Tutor Intelligence Deploys $16-an-Hour Robots to US Factories
Martin Casado, Raghu Raghuram, Josh Gruenstein
General-purpose robotic labor gives the US a greenfield manufacturing advantage over China, which faces the friction of tearing down billions of dollars in legacy, single-purpose automation.
Tutor Intelligence has deployed robots that have processed and shipped over 250 million items to U.S. consumers, bypassing academic R&D cycles by embedding directly into production supply chains. (0:00)
By offering standardized robotic labor under an hourly model starting at $16 an hour, Tutor compresses traditional multi-year industrial sales and deployment cycles down to 40–60 days to contract and single-day facility setup. (25:00)
To overcome the absence of ambient physical data on the internet, Tutor uses remote human 'tutors' to handle corner cases in real time, turning routine warehouse throughput into labeled training data for frontier models. (8:30)
Tutor operates 'Data Factory One' in Boston, a 100-robot semi-humanoid facility designed to run large-batch reinforcement learning experiments that generate targeted real-world experience traces. (40:00)
Detailed notes
## Background
Published on November 2, 2026, on RSS and web platforms by Martin Casado, Raghu Raghuram, and Josh Gruenstein, this episode examines what is required to deploy functional robots into American factories and warehouses. According to the show description, Tutor Intelligence CEO Josh Gruenstein joins a16z General Partners Casado and Raghuram to discuss Tutor's deployment-first methodology, their 100-robot data factory, robotic labor offered at approximately $16 an hour, and how pairing real-world work with human tutors generates training data to scale the physical economy.
## TL;DR
Josh Gruenstein argues that robotics is fundamentally a systems problem requiring vertical integration across hardware, software, supply chains, and go-to-market execution rather than isolated algorithmic research. To overcome the absence of pre-existing physical data on the internet, Tutor Intelligence deploys standardized robots like Cassie and Sunny directly into mid-market supply chain facilities, using human teleoperators to bridge corner cases while capturing high-value physical interaction data. By offering robots under an operational labor model starting at $16 an hour with zero capital expenditure, Tutor achieves rapid enterprise adoption and continuous two-week hardware-software iteration cycles. Anchored by domestic assembly in Massachusetts and a 100-robot testing fleet at Data Factory One, the company leverages frontier spatial reasoning models and large-batch reinforcement learning to build a scalable labor substrate for American manufacturing.
## Atmosphere & style
The discussion carries a grounded, fast-paced, and pragmatic tone, standing in sharp contrast to conventional venture and academic AI hype. Gruenstein sounds less like a theoretical roboticist and more like an operations executive fixated on line throughput, shift churn, and unit economics, creating an energetic and intellectually rigorous dynamic with the hosts. The mood blends the unvarnished realities of the industrial factory floor—such as cardboard handling, 55-year-old assembly line demographics, and tight manufacturing margins—with steady technical optimism about physical intelligence finally bridging the commercial gap.
## Bulletin
- `00:00` Tutor Intelligence has deployed robots that have processed and shipped over 250 million items to U.S. consumers, bypassing academic R&D cycles by embedding directly into production supply chains.
- `25:00` By offering standardized robotic labor under an hourly model starting at $16 an hour, Tutor compresses traditional multi-year industrial sales and deployment cycles down to 40–60 days to contract and single-day facility setup.
- `08:30` To overcome the absence of ambient physical data on the internet, Tutor uses remote human 'tutors' to handle corner cases in real time, turning routine warehouse throughput into labeled training data for frontier models.
- `40:00` Tutor operates 'Data Factory One' in Boston, a 100-robot semi-humanoid facility designed to run large-batch reinforcement learning experiments that generate targeted real-world experience traces.
## Key takeaways
- `05:00` Robotics is fundamentally a multi-layer systems problem across AI, hardware manufacturing, firmware, and go-to-market execution; excluding any layer eliminates most of the trade-off space required for commercial viability.
- `06:45` The trend toward vertical integration in robotics is a symptom of market immaturity, mirroring early computing before standardized horizontal abstractions like CPUs emerged.
- `08:30` Unlike digital AI models that benefited from decades of ambient web text and imagery, physical robotic intelligence produces no natural digital footprint and requires purpose-built networks to collect manipulation data.
- `10:00` Customer indifference to whether an arm moves autonomously or via remote teleoperation validates human-in-the-loop systems as an operational and economic success.
- `12:36` Paid production factory deployments ensure that training data remains strictly aligned with economically valuable distributions, avoiding the out-of-distribution pitfalls of passive video scraping.
- `12:36` Frontier AI lab breakthroughs in spatial reasoning and 3D foundation models for design software inadvertently bridged the semantic-spatial gap required for robotic manipulation.
- `16:44` Off-the-shelf industrial robots from legacy vendors were engineered for rigid, repeatable paths in shielded cages rather than co-design with dynamic, low-latency frontier models.
- `20:00` Tutor focuses on two core indoor morphologies—Cassie (single-arm suction) and Sunny (wheeled bi-manual semi-humanoid)—which together address roughly 80% to 90% of warehouse packaging and palletizing tasks.
- `20:00` Hyper-automated, bespoke machinery represents only 0.01% of American manufacturing, while the remaining 99.99% relies on manual labor suffering from three-month turnover and an aging workforce.
- `30:00` Tutor avoids the historical startup trap of building custom, multi-year bespoke machines for enterprise pilots by targeting mid-market customers as a forcing function for standardized, general hardware platforms.
- `30:00` Small cost reductions in automated manufacturing can trigger non-linear economic discontinuities, such as shifting modular home construction entirely from on-site building to automated factories.
- `33:38` In-house manufacturing in Watertown, Massachusetts, allows Tutor to push bi-weekly hardware updates and continuous firmware releases alongside daily software deployments.
- `33:38` Robotics unit economics have shifted: cost of goods sold is no longer dominated by multi-year hardware CapEx amortization, but by ongoing operational expenses like model inference, maintenance, and teleoperation.
- `37:10` Adaptable, greenfield robotic automation provides the United States an opportunity to bypass China's legacy investments in rigid, fixed-function automation lines.
- `40:00` Data Factory One houses a 100-robot fleet dedicated to capturing dense reinforcement learning traces at batch sizes inaccessible to standard academic research labs.
## Detailed walkthrough
### `00:00` The State of Robotics and Tutor Intelligence's Origin
Robotics pitches routinely flood venture capital desks, yet the vast majority of founders focus on narrow model architectures while giving little thought to what the robot will actually do in the real world. In contrast, Tutor Intelligence—a Boston-based company of roughly 160 people founded around 2020 by Josh Gruenstein—approached the challenge from the opposite direction. Gruenstein recognized that robotics has spent 60 years trapped as a cottage industry of custom-engineered, unscalable machines that remain a rounding error even within highly automated manufacturing and logistics environments. Rather than waiting for a fully autonomous artificial general intelligence (AGI) to materialize in a lab, Tutor set out to build a business that deploys real robots directly into supply chain enterprises today, helping manufacture and ship over a quarter of a billion products to American consumers.
This deployment-first model rests on a foundational thought experiment: there are eight billion people on Earth, all possessing the innate physical intelligence needed for factory and warehouse tasks. Even if one assumes a hypothetical software platform capable of connecting those people to any robot—functioning as an instant *deus ex machina* for unlimited physical intelligence—the broader economic problem remains unresolved. Deploying robots at scale still requires solving real-world manufacturing, assembling a resilient hardware supply chain, building entirely new operational software, and convincing customers who have never bought a robot before to adopt the technology.
```mermaid
flowchart LR
accTitle: The Deployment-to-Data Flywheel
accDescr: Flowchart illustrating how deployed robots use human tutors to complete tasks and generate training data for frontier models.
subgraph Operations["Real-World Deployment"]
A["Deploy robots to supply chain"] --> B["Perform factory work"]
end
subgraph Teleoperation["Physical Intelligence"]
B --> C["Human 'Tutors' bridge edge cases"]
end
subgraph Learning["Model Improvement"]
C --> D["Capture labeled physical data"]
D --> E["Train frontier robot AI"]
E --> A
end
```
- Tutor has manufactured and deployed robots that have processed and shipped more than 250 million items to U.S. consumers.
- The company targets a greenfield customer base, where almost every enterprise purchaser is buying a robot for the very first time.
- The company maintains a staff of roughly 160 employees based primarily in Boston.
- “Tutor robots have helped manufacture and ship over a quarter of a billion products to American consumers,” Josh Gruenstein points out.
`05:00` Robotics as the Ultimate Systems Problem and Vertical Integration
Martin Casado notes that despite Gruenstein’s background in MIT robotics, conversations with him sound less like academic technical discussions and more like deep dives with an operations and go-to-market executive obsessed with unit economics. Gruenstein explains that the shift is essential because robotics is fundamentally a systems problem rather than a pure software or hardware challenge. While an individual roboticist might specialize in FPGAs, ASICs, or neural network architectures, solving robotics commercially requires optimizing tradeoffs across the entire stack. Excluding any single layer—whether firmware, hardware manufacturing, supply chain logistics, or go-to-market execution—eliminates roughly 90% of the trade-off space required to make real progress.
Historically, robotics startups have struggled to achieve massive economic success due to two primary failures: the lack of a core intelligence primitive capable of driving down unit economics, and an excessively disaggregated market where fragmented vendors attempt to integrate disparate, narrow tools. Consequently, while most robotics startups set out to be horizontal software providers, almost all of them are forced to verticalize—often starting to 3D print their own parts when off-the-shelf components fail to meet their system requirements.
- The inability of prior robotics startups to attain massive economic scale stems directly from market fragmentation and the absence of generalized intelligence primitives.
- Restricting innovation to a single slice of the stack cripples progress by removing the cross-layer tradeoffs between hardware, software, and unit economics.
- “The only way you can make forward progress is by figuring out all of the clever tradeoffs across as many layers of the stack as you can,” argues Josh Gruenstein.
`06:45` The Evolution from Vertical Integration to General Platforms
Addressing the debate over whether physical robotics must stay vertically integrated forever or will eventually mirror the computing industry by horizontalizing, Gruenstein argues that the current trend toward verticalization is simply a symptom of market immaturity. Just as early computing required companies to build both the hardware chips and the operating environments, modern robotics requires full-stack platforms with viable internal business units to prove the system works. Over time, horizontal abstractions will emerge, analogous to how general-purpose CPUs abstracted computing workloads from underlying silicon design.
```mermaid
flowchart TD
accTitle: Market Evolution of Robotics Platforms
accDescr: Diagram showing the progression from custom specialized machinery to general-purpose platforms and eventual respecialization.
Stage1["Specialization (Custom Machines)"] -->|"Market maturation & scale"| Stage2["General Purpose Abstraction (e.g., Humanoids & CPUs)"]
Stage2 -->|"Scale expands massively"| Stage3["Respecialization (Targeted Compute & Form Factors)"]
```
Gruenstein views the humanoid form factor not as an arbitrary novelty, but as a clean, general-purpose physical abstraction comparable to a CPU. While a CPU is not perfectly optimized for every single computational task—which is why GPUs and ASICs exist—it provides a universally applicable baseline. Robotics is tracking this exact trajectory: starting with specialized custom machines, progressing toward general-purpose architectures (such as humanoids) as intelligence matures, and eventually splitting back into specialized form factors once the physical economy becomes large enough to support distinct hardware categories.
- “And in our view, the reality is you need to build a full stack platform,” Josh Gruenstein asserts regarding what it takes to succeed in modern robotics.
- “I think the beauty of something like a humanoid is it's the same. It's that clean abstraction,” Josh Gruenstein explains when comparing physical form factors to computing hardware.
`08:30` Bridging the Physical Data Gap: Why "Tutor" Intelligence?
The name "Tutor Intelligence" originates from the critical data bottleneck separating digital AI from physical robotics. The modern explosion of Large Language Models (LLMs) was made possible because the digital activities of billions of people—posting on social media, authoring articles, writing emails, and publishing code—are inherently networked. The internet captured that ambient exhaust of human intelligence as a massive, pre-labeled dataset that could train foundation models for the digital world.
In contrast, human physical intelligence produces no natural digital footprint. When a person performs an ordinary task like picking up a coffee cup, adjusting a misaligned box on a conveyor belt, or packing an irregularly shaped object, no network records the motion, the tactile feedback, or the spatial reasoning involved. Tutor Intelligence was founded to build that missing network: deploying robots into operational workplaces and pairing them with human operators ("tutors") who step in when the automation encounters an edge case. This teleoperation and assistance loop does not just keep warehouse lines moving; it systematically records and digitizes baseline physical intelligence, providing the raw data substrate needed to train general-purpose robotic models.
- Digital foundation models benefited from decades of ambient human data passively archived on the internet, whereas physical manipulation lacks an equivalent pre-existing dataset.
- Human tutors act as an operational failsafe while simultaneously generating the labeled, real-world physical interaction data required to train foundation models.
- “Tutor was founded kind of with this thesis that there is a whole internet of human intelligence that is networked, right? That is posting on social media, that's writing emails,” Josh Gruenstein explains when comparing digital data exhaust to the physical world.
`10:00` Building Tutor Intelligence: The Economics of Human Teleoperators
Josh Gruenstein explains the founding thesis of Tutor Intelligence: in text and vision, massive internet datasets enabled modern AI, but physical manipulation possesses no equivalent pre-existing digital archive of fingertip pressures, 6-DoF poses, and contact dynamics. Rather than trying to invent an algorithmic miracle that spawns physical intelligence out of thin air, Tutor was designed to generate its own training data directly from production environments. By deploying physical robots into factories and pairing them with human teleoperators—dubbed "tutors"—the company created a real-time feedback loop where humans step in to execute or guide tasks whenever autonomous policies fall short. Gruenstein emphasizes that this approach bypasses the need for massive, non-revenue-generating R&D capital projects: each deployed robot immediately produces economic value on day one, subsidizing the data capture that fuels the broader network.
The investors point out that when doing due diligence with Tutor’s factory customers prior to investing, factory operators rarely knew or cared whether an arm was moving via autonomous model weights or via a remote human tutor; all that mattered was throughput and reliability. Gruenstein views this indifference as the ultimate stamp of approval for an industrial product. When asked whether the end goal is to eliminate human tutors entirely, he frames the decision as a strict unit-economics calculation analogous to sizing cloud infrastructure. Just as one measures the duty cycle, utilization, and cost of an NVIDIA H200 cluster, an automation provider must balance the heavy engineering expense required to achieve "the last nine" of model autonomy against the marginal cost of routing tricky edge cases to human intelligence.
> “If the feedback is like the robot is pretty, it's like, oh, it's probably a real bad robot.” — Josh Gruenstein
```mermaid
flowchart LR
accTitle: The Tutor Intelligence Data and Economic Flywheel
accDescr: A closed loop showing how customer deployments fund operations, generate training data via tutors, and improve autonomous models.
subgraph Site["Factory Floor"]
A["Robot on Production Line"] -->|"Real-time Work"| B["Factory Output & Revenue"]
end
subgraph HybridOps["Execution Layer"]
A -->|"Uncertainty / Corner Case"| C["Remote Human Tutor"]
C -->|"Guidance & Teleop"| A
end
subgraph DataEngine["Model Improvement"]
C -->|"Demonstrations & Sensory Data"| D["Model Retraining"]
D -->|"Expanded Autonomous Autonomy"| A
end
```
`12:36` Data Alignment, Frontier 3D Models, and Escaping the Out-of-Distribution Trap
The conversation shifts to the structural differences between generative AI and physical robotics. In pure software generative AI, the training distribution closely mirrors the inference distribution—models trained on internet text and web images produce text and images directly consumed by internet users. In physical environments, however, even a slight departure from the training distribution can cause physical manipulation to fail completely. Martin Casado and Gruenstein debate how to guarantee that robotic training data remains commercially viable. Gruenstein critiques brute-force academic data-collection efforts—such as strapping action cameras to workers across developing nations—because they lack market alignment. By contrast, collecting operational data as a byproduct of real, paid customer workflows guarantees that every frame and force-torque trajectory collected has proven commercial utility.
At the same time, Gruenstein acknowledges that Tutor cannot rely solely on physical teleoperation data to span high-dimensional physical state spaces. The team supplements production logs with physics simulations and internet-scale vision models to broaden distribution tails. Crucially, recent breakthroughs in spatial reasoning models have triggered a sea change for robotics. Because frontier AI labs face massive commercial incentives to generate CAD and 3D modeling tools for software and design markets, they have begun post-training foundation models on dense spatial data. Gruenstein explains that this commercial push inadvertently solved the missing link in robotics: connecting the abstract semantic reasoning of large language models to the fine-grained physical spatial awareness required by action diffusion transformers.
> “There's no invisible hand of the market that's guiding you to ensure that that data is actually economically valuable.” — Josh Gruenstein
- In contrast to open-ended video collection, paid factory deployments act as an automatic filter for economically relevant training distributions.
- Frontier AI post-training on CAD and 3D data provides the spatial grounding that previously prevented large language models from guiding robotic manipulators.
- Tutor has evolved its underlying machine learning stack continuously since its founding in 2021, transitioning from early convolutional neural networks to modern diffusion-based foundation models.
When asked what "building robots" actually means in practice, Gruenstein puts the state of American industrial automation into perspective. In 2025, the entire United States deployed roughly 38,000 industrial robot workcells—a microscopic figure compared to the roughly one million deployed internationally, primarily across Asian manufacturing hubs. Furthermore, the legacy supply chain dominated by traditional players like Fanuc and ABB is fundamentally mismatched with modern AI-driven systems. Traditional industrial arms were engineered for rigid systems integrators executing deterministic, high-speed, repeatable trajectory paths in shielded enclosures. They were not architected with the compliance, sensory feedback, or dynamic latency requirements needed to co-design with frontier foundation models.
Because neither legacy industrial giants nor consumer-oriented quadruped and humanoid vendors (such as Unitree) offered platforms suited for modern embodied AI, Tutor was forced to build its own hardware stack. While Gruenstein avoids reinventing commoditized subcomponents—sourcing raw actuators, motors, and camera sensors from large-scale original equipment manufacturers (OEMs)—Tutor designs, integrates, and manufactures its own complete robotic platforms in Watertown, Massachusetts. Owning the end-to-end integration of mobile bases, structural arms, and sensor suites gives the software stack direct, deterministic control over the full physical system.
> “They are not cut from the set of tradeoffs that you would want to co-design with frontier models.” — Josh Gruenstein
- Total US industrial robot installations in 2025 stood at approximately 38,000 units, demonstrating the vast domestic under-penetration of automated labor.
- Consumer-grade hardware vendors like Unitree had virtually zero verified enterprise production deployments in US industrial workcells over that period.
- Tutor centralizes its final system integration and custom manufacturing at its headquarters in Watertown, Massachusetts.
`18:25` Robot Morphologies: Cassie and the Physical Economy of Pallets
Gruenstein details the physical morphologies Tutor builds to solve real industrial bottlenecks, beginning with its workhorse machine, "Cassie." Cassie is designed deliberately as a heavy-duty industrial unit rather than a humanoid novelty. Equipped with a rugged single arm and an array of high-power suction grippers, Cassie runs vision-action foundation models operating in the 200-to-400-million parameter range. Rather than requiring open-ended, human-like dexterity for every motion, the robot co-designs hardware and software to predict physical affordances and grasp points directly from raw sensor feeds before executing rapid pick-and-place routines.
The specific morphology of Cassie reflects the mechanical realities of the American supply chain. Gruenstein notes that nearly every physical commodity manufactured in the United States eventually moves on a wooden pallet or inside a corrugated cardboard case. Because the supply chain relies universally on these standardized units, enormous volumes of manual human labor are spent simply shifting heavy cartons and stacking pallets. Cassie was purpose-built to automate this specific, grueling layer of logistics and manufacturing at scale across diverse facilities, serving as Tutor’s primary production workhorse alongside their second machine morphology, "Sunny."
> “Anything in the physical economy has to be transported on a pallet or in a cardboard box.” — Josh Gruenstein
- Cassie is built around a single manipulation arm paired with specialized industrial suction end-effectors to reliably move corrugated cardboard boxes and cases.
- The autonomy stack driving Cassie utilizes foundation models sized between 200 million and 400 million parameters to infer grasp affordances.
- Palletizing and depalletizing remain among the highest-volume, highest-churn manual tasks across the physical manufacturing and logistics ecosystem.
`20:00` Sonny, Cassie, and the Reality of American Manufacturing
Josh Gruenstein breaks down the engineering behind Tutor’s robot form factors, specifically highlighting Sunny. Sunny is an industrial semi-humanoid designed with different trade-offs than legged research platforms like Unitree. Because indoor industrial facilities have polished concrete floors, Sunny uses wheels rather than legs, allowing it to achieve a much higher payload capacity, operate at human speeds with industrial collaborative safety standards around people, and run continuously for 3 to 10 years. Tutor deploys Sunny internally to collect manipulation data and push forward their research agenda, while an expanding commercial fleet handles dexterous, bimanual manipulation in real customer facilities. When asked whether a company needs three or thirty robot form factors to capture the industrial market, Josh states that between Sunny and Cassie (their single-arm, sucker-equipped robot), Tutor can address roughly 80% to 90% of manufacturing and logistics tasks.
```mermaid
flowchart TD
accTitle: Robot Form Factors and Work Environments
accDescr: Mapping Tutor Intelligence robot form factors Sunny and Cassie to factory applications and identifying outdoor edge cases requiring future legged embodiments.
subgraph Limitations["Current Environmental Limits"]
direction TB
Outdoor["Outdoors & Uneven Terrain"] -->|"Gravel, construction sites, landfills"| FutureLegs["Future Legged Embodiments"]
Home["Residential Homes"] -->|"Too industrial & visually intimidating"| Consumer["Future Consumer Units"]
end
Factory -.->|"Cannot traverse gravel or outdoors"| Limitations
```
Josh contrasts the public imagination of manufacturing with what actually exists on the ground. Most people base their mental model on the television show *How It’s Made*, imagining hyper-automated, Rube Goldberg-style machines where no humans are visible. In reality, that level of bespoke hard automation represents only the top 0.01% of American industry. The remaining 99.99% consists of million-square-foot facilities where rows of people stand along conveyor belts or folding tables doing repetitive tasks entirely by hand. Labor in these positions suffers from constant three-month turnover, with the average remaining worker around 55 years old and nearing retirement. Tasks as ubiquitous as placing consumer goods into boxes or assembling final retail packaging—even on high-end electronics like iPhones—are still performed by hand at massive scale.
Addressing the physical boundaries of their current machines, Josh explains that while Cassie and Sunny cover indoor logistics, they cannot go outside due to hardware and locomotion constraints. Despite articulating hind legs, wheeled bases cannot navigate loose gravel, uneven earth, or construction environments. Expanding beyond polished concrete into construction, material recovery from landfills, or unstructured outdoor settings will eventually require legged platforms, while consumer homes represent another distinct environment where industrial machines like Cassie and Sunny are far too imposing.
- **Indoor vs. outdoor engineering:** Sunny relies on wheels over polished floors to maximize payload, speed, and 3-to-10-year durability.
- **Workforce demographics:** The manual packing jobs Cassie and Sunny automate suffer from 3-month turnover and an average worker age of 55.
- **Packaging reality:** Virtually every consumer good shipped in a cardboard box is placed there by a human worker rather than bespoke factory machinery.
---
`25:00` The $16/Hour Pitch and Frictionless Velocity
Customer acquisition in industrial automation is historically slow, but Tutor accelerates adoption by treating deployment as a greenfield opportunity. Out of the millions of commercial operations in the United States, roughly 9 out of 10 factories and warehouses have zero robots, and the entire country deploys only around 38,000 industrial robots annually. Traditional automation requires heavy capital expenditure, complex systems integration, and extensive custom engineering, leading most facility managers to conclude that robotics is simply out of reach. Tutor circumvents these historical hesitations entirely: rather than leading with pitches about artificial intelligence, foundation models, or remote human tutors, they cold-call operators with a simple economic framing: “Hey, have you ever considered a robot for $16 an hour?”
Because Tutor absorbs the capital and technical risk under an hourly labor model and ships standardized, generally capable hardware directly from the factory without site-specific customizations, their sales and installation cycles operate at speeds unheard of in industrial robotics:
- **Lead to contract:** 40 to 60 days on average from the initial cold call to a signed agreement.
- **Contract to delivery:** Approximately 40 days to have physical units on-site.
- **Deployment time:** Systems are stood up in production within a single day, often within the first two hours of arriving at the facility.
Once an operations team sees the robot perform, adoption spreads rapidly. Josh highlights the pragmatic nature of the operations community: once they verify that a machine works on one station, they immediately look across their footprint to deploy it on identical stations, allowing Tutor to scale from a single pilot machine to a dozen or more units across multiple sites.
> “A robot is a robot unless it works, in which case it’s a machine.”
> — Josh Gruenstein quoting an industry observation
---
`27:30` Human-Robot Interaction on the Factory Floor
Josh highlights a fundamental dynamic in industrial robotics: the economic buyer is not the actual user. While the COO or general manager signs the contract, the daily user is the line worker sharing a station with the machine. Because Tutor automates specific tasks rather than replacing entire facilities, managing Human-Robot Interaction (HRI) is essential for long-term operational success. Workers often feel initial apprehension, but as the robot relieves them from strenuous, repetitive strain, their relationship with the equipment warms significantly.
To build positive relationships between workers and their new mechanical partners, Tutor incorporates standard HRI practices, starting with naming every unit:
- **First deployment names:** Tutor’s internal manufacturing technicians name the units during assembly to reflect their personal care and craftsmanship.
- **Repeat order naming rights:** When a customer places a repeat order to expand their fleet, the customer’s floor workers are granted naming rights, creating an immediate sense of local ownership.
- **Naming themes:** Workers frequently name the machines after their own children, favorite football players, or pop stars.
- **Retirement of hurricane names:** Tutor initially used the meteorological alphabet (hurricane names) for production tracking, but abandoned the convention after cycling through the alphabet too many times to keep units organized.
The section closes with Raghu Raghuram noting that while most venture-backed robotics startups pursue Fortune 50 enterprise accounts with massive experimental budgets, the vast majority of Tutor's customer base consists of local, small-to-midsize industrial businesses. Josh confirms that this was an intentional strategic choice, pointing out that the overwhelming majority of manufacturing and logistics facilities across the United States are small and mid-market operations that have historically been completely ignored by the automation industry.
`30:00` Building a General Robot Platform vs. Academic Traps
Josh Gruenstein reflects on the founding thesis of Tutor Intelligence, noting that the startup ecosystem had been littered with academic, supply-chain-oriented robotics companies that raised money, launched, and swiftly collapsed. The root cause was that these startups failed to act like true product companies. Typically, a large enterprise would present a niche challenge accompanied by a non-binding letter of intent promising future expansion. The startup would spend two years designing a hyper-customized Rube Goldberg machine, deploy it, and watch it fail because hardware products inherently require ongoing iteration rather than a "one and done" release. Even if the bespoke machine worked, the customer's operational needs often shifted during those two years, immediately rendering the machine obsolete and driving the robotics company out of business. Gruenstein felt that watching this tragedy repeat five to ten times in a row served as an unmistakable warning.
To avoid that fate, Tutor focused on serving mid-market businesses. Gruenstein emphasizes that mid-market diversity acts as an essential forcing function to build a standardized, general-purpose platform rather than bespoke machinery. If the company can close a contract within two weeks with a mid-market customer and have a robot operating on the floor immediately afterward, it proves that both the robot body and the underlying model possess genuine generalization capabilities. When Raghu Raghuram and Martin Casado ask about the economic and productive convergence limit—whether things simply get cheaper or if quality and variety also explode—Gruenstein explains that robotics will unlock massive economic discontinuities.
```mermaid
flowchart TD
accTitle: Bespoke Robotics vs General Platform Approach
accDescr: Contrasting the failed custom engineering approach with Tutor's standardized platform model.
subgraph Legacy["Academic / Bespoke Approach (Fails)"]
L1["Big Customer with Specific Niche Problem"] --> L2["2 Years of Custom Rube Goldberg Engineering"]
L2 --> L3["Rigid Deployment with No Iteration Loop"]
L3 --> L4["Customer Needs Shift / Company Fails"]
end
subgraph Tutor["Tutor Intelligence Platform (Succeeds)"]
T1["Mid-Market Customers Across Diverse Verticals"] --> T2["Generally Capable Hardware & Models"]
T2 --> T3["2-Week Contract to Immediate Floor Deployment"]
T3 --> T4["Continuous Iteration & General Platform Moat"]
end
```
Gruenstein illustrates these discontinuities with an example from modular construction. In a conversation with an assembly line worker at a modular housing factory in Wisconsin, Gruenstein learned that constructing a house inside a factory is currently about 5% more expensive than erecting it on-site with traditional construction methods. However, if manufactured housing costs drop by even 1% below traditional on-site construction, real estate developers will divert the vast majority of housing demand directly into automated factories. Similar tipping points exist for recycling and material reclamation in landfills. When asked if Tutor will ever enter the home, Gruenstein remarks that they will go wherever the economics make sense, though home environments present extreme safety hurdles, high task complexity, and unfavorable unit economics compared to factories.
- “And so I think working with a diverse range of customers is this beautiful forcing function for building the right shape of product, which to us is a general robot, right?” — Josh Gruenstein
- “So I think our hope is like if Tutor is really successful, we can provide this robot substrate, this labor force that other businesses can use to transform the physical economy and make the world ultimately a cheaper and better place to live in.” — Josh Gruenstein
- “Take trade-offs across as many layers of the stack as possible to get a foothold, start building the platform, and as you build the platform, you have to take fewer and fewer trade-offs over time.” — Josh Gruenstein
---
`33:38` Rapid Hardware Iteration and the Shifting Cost Structure
Addressing how product innovation functions when dealing with physical machines, Gruenstein highlights that Tutor manufactures its own robots rather than outsourcing the core hardware stack. While most robotics companies stagnate on a single hardware generation for years, in-house manufacturing gives Tutor total control, enabling the engineering team to push hardware updates and manufacturing process improvements every two weeks. Software code is deployed daily to the fleet, accompanied by major releases every fortnight, backed by custom deployment infrastructure. At the same time, Tutor maintains an embedded research team that co-designs hardware and models simultaneously: when research produces a breakthrough in spatial reasoning or robotic manipulation, the engineering team can immediately alter the hardware and push the new capability straight into customer facilities.
Prospective warehouse and factory operators do not need multi-year integration roadmaps; they can visit Tutor's website, input their operational throughput into an interactive pricing calculator, and evaluate cost reductions. A customer can see projected annual savings ranging from $100,000 to upwards of $100 million and deploy units within two months with zero capital expenditure (CapEx). When Casado and Raghuram probe into whether the dominant cost driver is upfront amortization or ongoing operations, Gruenstein explains a profound shift in robotics unit economics: cost of goods sold (COGS) is no longer dominated by heavy CapEx depreciated over a decade, but by ongoing operational costs like inference, teleoperation support, and break-fix maintenance.
- “We are shipping code to robots either every day or major releases every two weeks.” — Josh Gruenstein
- “And so if we're able to compete in these highly liquid labor markets that exist across American physical economy, and we can offer something that delivers greater equivalent capability to lower cost, the market for that is the whole market, right?” — Josh Gruenstein
- When Casado asks whether the upfront hardware amortization or ongoing operational cost dominates COGS, Gruenstein responds: “Yeah, it is more ongoing at this point.” — Josh Gruenstein
- Tutor's foundational economic heuristic: value equals the total work a robot can perform divided by its overall cost.
---
`37:10` Strategic Automation and Reshaping Global Manufacturing Geography
Gruenstein rejects the notion that customers view Tutor robots as temporary surge staffing for peak seasons like Black Friday. Instead, adopting robots is a fundamental, strategic survival imperative. Industrial manufacturers—whether producing deodorant, crayons, or consumer packaged goods—face an aging blue-collar workforce, escalating domestic labor costs, and surging baseline demand. Because physical manufacturing operates on thin gross margins that lack the buffer of software businesses, these facilities either solve their structural labor deficit or go out of business. Factory operators prefer to remain masters of their specific manufacturing domain and partner with Tutor to manage the operational complexity of automation.
Looking at the broader macroeconomic and geopolitical horizon, Gruenstein predicts that general-purpose robotics will fundamentally upend global manufacturing geography and the traditional concept of single-industry "factory towns." In the prior industrial automation cycle, countries like China poured immense capital into rigid, fixed-function machinery tailored to specific workflows. Gruenstein suggests this historic asset base may become a liability: while China must eventually tear down rigid automation lines, the United States has the opportunity to deploy greenfield automation powered by adaptable, generalized robot workers. In parallel, lowering the capital barrier to physical automation will mirror the software revolution catalyzed by cloud computing: anyone will be capable of founding, launching, and scaling physical manufacturing enterprises without massive upfront industrial footprints.
- “Because, you know, these are not software businesses that have these, you know, huge gross margins that you can you can play with and can compress a little bit. And it's like, it's fine. It's like, no, the business either works or it doesn't.” — Josh Gruenstein
- “I think one big thing is this is going to be a great equalizer between the United States and other countries like China invested in the previous regime of industrial automation.” — Josh Gruenstein
- “Anybody will be able to start a business in the physical world. Anybody will be able to scale a business in the physical world if we kind [of succeed].” — Josh Gruenstein
`40:00` Global Supply Chains, In-House Manufacturing, and Data Factory One
The discussion shifts to the hardware supply chain and the heavy role that China plays in robotics manufacturing. When asked about Tutor Intelligence’s reliance on overseas suppliers and how that risk is handled over time, Josh Gruenstein emphasizes that the manufacturing reality cannot simply be decoupled overnight: “I think ultimately, no matter what anybody says, today in 2026, we have a highly interconnected, interdependent global supply chain.” Both American and Chinese firms are navigating long-term industrial strategies through cross-border investment and domestic capacity building. Martin Casado points out that because robotics starts as a cottage industry, rapid scaling fundamentally changes supplier dynamics. As Tutor's deployment volume climbs to make it one of the largest robotics customers in the country, external component suppliers begin functioning more like contract manufacturers, forcing Tutor to decide whether to specify custom designs or vertically integrate production directly.
Josh confirms that Tutor views building domestic capacity as an active responsibility: “I think ultimately, for Tutor to be successful, we will need a mature robotic supply chain in the United States.” To that end, the company builds its systems in Watertown, Massachusetts, and is progressively verticalizing component production to reduce foreign dependency.
Turning to AI and machine learning architectures, Josh draws a direct analogy to the progression of large language models. While LLMs leverage diverse, heterogeneous web data, high-performance models ultimately depend on the very peak of the data pyramid: high-quality reinforcement learning traces of the exact behaviors being targeted. To replicate this paradigm physically, Tutor established a massive facility in Boston known as Data Factory One. The operation houses an active fleet of 100 robots, representing what Josh describes as the largest semi-humanoid data collection farm in the United States, specifically engineered to run large-batch reinforcement learning experiments that neither academic labs nor standard startups can execute.
```mermaid
flowchart TD
accTitle: The Robotics Learning Data Pyramid
accDescr: Diagram showing how base datasets feed into real-world RL traces gathered at Data Factory One.
subgraph DataInputs["Data Sources"]
A["Heterogeneous Base Datasets"]
B["Synthetic & Simulation Data"]
end
subgraph PyramidApex["Top of Data Pyramid"]
C["Real-World Experience Traces"]
D["100-Robot Fleet: Data Factory One"]
end
subgraph RLTraining["Model Optimization"]
E["Large-Batch Real-World Reinforcement Learning"]
F["Deployed Production Capabilities"]
end
A --> C
B --> C
C --> D
D --> E
E --> F
```
- Tutor Intelligence conducts its primary assembly and manufacturing operations out of Watertown, Massachusetts.
- Data Factory One, based in Boston, deploys a 100-robot fleet dedicated to generating continuous reinforcement learning traces.
- Reinforcement learning requires large batch sizes that cannot function effectively on small research fleets, making an industrial-scale test bed essential for rapid algorithmic progress.
`43:08` Full-Stack Robotics Hiring and Interdisciplinary Culture
Operating at this scale requires a team structure completely divergent from a traditional software research laboratory. When asked what profiles Tutor Intelligence is seeking, Josh responds that they are hiring across the entire spectrum, from machine learning scientists down to assembly floor technicians. He clarifies that vertical integration leaves no room to bypass physical realities: “It's robotics. You're going to engage with it whether you want to or not. The question is whether you understand it.” Engineering challenges span software models, physical materials, custom electronics, and industrial hardware, requiring continuous handoffs between disciplines.
Josh highlights the rare work environment that emerges from this setup, describing Tutor as one of the few places where an engineer might debate transformer or model architectures on a whiteboard in the morning and spend the afternoon inspecting real robot operations directly on a factory floor. While core engineering and production operations are anchored in Massachusetts, field technicians and go-to-market teams are deployed across the United States to service customer factory installations nationwide.
- Hiring covers all technical layers: foundational AI research, software engineering, materials science, electronics, and assembly technicians.
- Field technicians and commercial operations are distributed across the country to maintain on-site deployments on active factory lines.
`45:00` Outro and Closing Remarks
The conversation concludes with final thanks from the hosts, followed by standard network sign-offs directing listeners to subscribe, leave reviews, and follow the a16z podcast across YouTube, Apple Podcasts, Spotify, Substack, and X.
## Notable quotes
> `25:00` A robot is a robot unless it works, in which case it’s a machine. — Josh Gruenstein
> `18:25` Anything in the physical economy has to be transported on a pallet or in a cardboard box. — Josh Gruenstein
> `05:00` The only way you can make forward progress is by figuring out all of the clever tradeoffs across as many layers of the stack as you can — Josh Gruenstein
> `43:08` It's robotics. You're going to engage with it whether you want to or not. The question is whether you understand it. — Josh Gruenstein
## Mentioned
Tutor Intelligence, Josh Gruenstein, Martin Casado, Raghu Raghuram, a16z, MIT, NVIDIA, Fanuc, ABB, Unitree, Cassie, Sunny, Data Factory One, How It's Made, iPhone
## Video description
What does it take to build robots that can actually work in factories and warehouses? Tutor Intelligence CEO Josh Gruenstein argues robotics is a systems problem spanning AI, hardware, manufacturing, and the work itself.
Josh joins a16z General Partners Martin Casado and Raghu Raghuram to explain Tutor’s approach: deploy robots into real businesses, use human “tutors” when needed, and turn that work into training data that makes the robots more capable.
They discuss Tutor’s 100-robot data factory, robots starting at roughly $16 an hour, advances in spatial reasoning, and what flexible robotic labor could mean for American manufacturing and the physical economy.
Tutor Intelligence founder Josh Gruenstein argues that commercial robotics must be approached as a full-stack systems and unit-economics problem rather than an academic AI challenge, relying on production deployments and human teleoperators to generate the training data needed for scalable automation.
Solving the Physical Data Bottleneck Through Tutors (8:30)
Unlike generative language models that leverage decades of passively archived text on the internet, physical robotics lacks a pre-existing digital archive of contact dynamics, fingertip pressures, and 6-DoF poses. To bridge this gap, Tutor pairs production robots on factory floors with remote human teleoperators ("tutors") who intervene when automation confronts an edge case. Gruenstein explains that this setup produces immediate economic value for customers while systematically capturing the labeled real-world interaction traces necessary to train foundation models.
Standardized Platforms and Form Factors Versus Custom Engineering (20:00)
Gruenstein observes that earlier robotics startups repeatedly failed by building bespoke, Rube Goldberg-style machinery for individual enterprise accounts, only to become obsolete when customer requirements changed. Tutor instead addresses roughly 80% to 90% of industrial tasks using two standardized form factors: Cassie, a single-arm system equipped with suction grippers tailored to palletizing cardboard cartons, and Sunny, a wheeled, semi-humanoid bi-manual robot designed for polished concrete factory floors. To continuously advance its learning models, Tutor complements customer deployments with Data Factory One in Boston, a facility operating 100 robots to collect large-batch reinforcement learning traces.
Targeting the Mid-Market via Hourly Labor Economics (25:00)
While traditional industrial robotics involves steep capital expenditures, custom engineering, and integration cycles that exclude the vast majority of facilities, Tutor pitches mid-market operators with a labor model framed around $16 an hour. With approximately 90% of American manufacturing and warehouse sites lacking robots and only about 38,000 industrial robot workcells installed domestically in 2025, Gruenstein views the market as predominantly greenfield. This business model allows Tutor to move from cold call to contract within 40 to 60 days and achieve live floor operations within hours of delivery.
In-House Manufacturing and Full-Stack Tradeoffs (16:44)
Gruenstein contends that commercially viable robotics requires controlling cross-layer tradeoffs between firmware, hardware manufacturing, and software, noting that legacy arms were designed for deterministic, shielded environments rather than co-design with foundation models. Tutor centralizes its final system assembly in Watertown, Massachusetts, sourcing raw commoditized components like actuators from OEMs while maintaining internal production. This setup enables bi-weekly hardware and manufacturing updates, daily fleet software deployments, and a shifting cost structure where ongoing operational expenses—such as inference, maintenance, and teleoperation—dominate over upfront equipment depreciation.
Gruenstein anticipates that lowering the barrier to adaptable physical automation will allow domestic greenfield operations to leapfrog rigid legacy automation lines, though significant hurdles remain—including a reliance on an interconnected global hardware supply chain and locomotion limits that restrict current machines from unstructured outdoor environments.
## Background
该内容由 Martin Casado、Raghu Raghuram 与 Josh Gruenstein 于 2026 年 11 月 2 日在播客及网络平台发布。根据节目简介，Tutor Intelligence 联合创始人兼 CEO Josh Gruenstein 与 a16z 普通合伙人深入探讨了让机器人真正走进工厂和仓库所需应对的系统级工程挑战，以及 Tutor 结合现场部署、人类远程教学与数据工厂来重塑美国制造业的实战路径。
## TL;DR
Josh Gruenstein 详细阐述了 Tutor Intelligence 避开纯算法学术幻想、从真实工业现场逆向切入机器人的全栈哲学。公司通过部署轮式重载机器人 Cassie 与半人形协作机器人 Sunny，以极其务实的“时薪 16 美元”、零前期资本支出的服务模式打入美国中小型制造与仓储企业的空白市场，目前已协助运送超 2.5 亿件商品。在技术闭环上，Tutor 借助人类远程导师（Tutor）处理长尾极端工况，并利用包含 100 台真机的数据中心 Data Factory One 沉淀高质量强化学习轨迹，实现自给自足的数据飞轮与双周级软硬件迭代，旨在构建推动实体经济变革的通用机器人劳动力底座。
## Atmosphere & style
对谈洋溢着硅谷顶尖投资人与一线硬核创业者之间既严谨深刻又极为务实的交流氛围，完全剥离了机器人行业常见的空泛炒作与科幻滤镜。Josh Gruenstein 的分享充满了直击一线车间现实的敏锐洞察与黑色幽默，将复杂的模型架构推导与蹲在工厂流水线旁排查螺栓与纸箱卡死问题的粗糙现实巧妙交织。整场对话既有对美国制造业劳动力断层和地缘供应链博弈的清醒研判，又流淌着一种脚踏实地重构物理世界基础设施的昂扬雄心。
## Bulletin
- `00:00` Tutor Intelligence 避开过度追求前沿算法的脱节路线，通过直接在真实工厂落地部署并结合人类远程“导师”（Tutors）接管干预，累计已协助制造和交付超过 2.5 亿件商品。
- `23:40` 团队打破昂贵的传统自动化采购模式，以“时薪 16 美元”和零前期资本支出（Zero-CapEx）的极简商业模式切入，将签约到上线的周期压缩至两个月内。
- `20:00` Tutor 推出重载拆码垛机器人 Cassie 与轮式工业半人形态机器人 Sunny，并断言这两种形态结合即可覆盖工业制造与仓储物流领域 80% 至 90% 的工种需求。
- `42:47` 为突破真机强化学习的数据瓶颈，公司在波士顿自建拥有 100 台机器人的数据农场“Data Factory One”，旨在通过大规模真实交互轨迹训练下一代具身基础模型。
## Key takeaways
- `05:00` 机器人的本质不是单点突破的技术，而是涵盖模型、硬件、制造、供应链及客户业务全链路的“终极系统工程”。
- `10:00` 物理世界不存在互联网天然沉淀的语料副产物，具身智能所需的高价值物理交互数据唯有通过真实商业部署的生产过程才能低成本捕获。
- `12:00` 人类远程接管（Teleoperation）应被视为高阶异构的计算与劳动输入，用以平衡消灭长尾错误的高额工程研发成本与实际生产毛利。
- `15:00` 顶尖 AI 实验室针对下游 3D/CAD 软件需求进行的后训练，意外补齐了基础模型在三维空间推理（Spatial Reasoning）上的短板，直接赋能机器人的动作规划。
- `17:30` 传统工业机械臂是为固定轨迹与系统集成商量身打造的，无法适配端到端基础模型，倒逼 Tutor 必须自研硬件整机与控制架构。
- `20:00` 工业厂房地面多为平整硬质环境，轮式半人形态机器人（如 Sunny）在满足安全协作、高速作业及长期耐用性上远比双足机器人更符合工业现实。
- `20:00` 绝大多数美国普通工厂仍依赖纯手工完成繁重、枯燥的流水线工序，一线劳动力老龄化与高流动性使工厂对机器人产生了迫切的战略级用工需求。
- `30:00` 许多学术型机器人初创公司失败的根源在于沦为了定制工程外包商，陷入长周期开发并错失敏捷迭代机制；机器人公司必须转型为通用产品型平台。
- `34:00` 机器人的价值公式极其纯粹：机器人完成的总工作量除以其综合成本，而当前综合成本的主体已从硬件折旧转变为云/边推理与持续运营支持。
- `34:00` 硬件制造自主化使 Tutor 能以两周为周期将产线改进迅速推向生产，构建起软硬件协同迭代的敏捷开发壁垒。
- `36:30` 通用机器人劳动力底座将重塑全球实体经济，美国有望在工业自动化的空白地带跳过旧有固定资产包袱，直接构建灵活制造的“云计算时刻”。
- `40:00` 当采购规模上升到足以影响供应商业务时，OEM 便会退化为代工厂，促使 Tutor 进一步走向美国本土化制造与关键组件的深度垂直整合。
- `42:47` 借鉴大语言模型的演进逻辑，具身智能模型能力的最终壁垒位于数据金字塔尖顶——即通过真机集群在真实任务中探索出的大规模强化学习轨迹。
## Detailed walkthrough
### `00:00` 真实世界部署的断层与“人类导师”思想实验
在当前的全球产业链中，美国与中国企业都在加紧构建自身的工业制造能力。然而在当下的机器人领域，大多数创投项目（Pitches）都陷入了脱离实际的误区：99% 的项目都在谈论前沿 AI 模型或某种特定的数据结构，却鲜有人真正思考机器人最终在工厂终端究竟要做什么。与此相对，Tutor Intelligence 选择了一条反向路径：直接切入从未购买过机器人的广阔空白市场（Greenfield Market），并已制造并在真实生产环境中部署了大量机器人。Tutor Intelligence 的机器人已协助制造并向美国消费者交付了超过 2.5 亿件产品，几乎所有美国消费者的生活底层都在直接或间接地被其服务。
公司联合创始人兼 CEO Josh Gruenstein 解释了公司名称“Tutor”的灵感来源：它始于一个反向的思想实验。面对机器人行业长达 60 年来依赖定制工程化机器、无法规模化扩张的软件与智能瓶颈，他们思考：地球上有 80 亿人，每个人都天生具备我们希望机器人拥有的基础物理智能（Physical Intelligence）。如果拥有某种无限强大的软件，能够把地球上的任意人类连接到任何一台机器人上，赋予它人类智慧（就像戏剧中的“天降神兵”机械降神 Deus Ex Machina 一样，瞬间拥有无限的机器人智能），那接下来会发生什么？结论是：即便拥有了这种智能，问题依然远未解决。美国的机器人制造业依旧处于手工作坊式（Cottage Industry）阶段，甚至在自动化程度最高的制造和物流行业里也只是微不足道的尾数误差（Rounding Error）。硬件大规模制造能力不存在、适配软件不存在，能真正完成客户期望工作的落地场景与客户群体也同样不存在。
- **核心数据**：Tutor Intelligence 创立约五年半，团队规模约 160 人（主要位于波士顿），其协作制造与运送的商品已超 2.5 亿件。
- **行业痛点**：机器人售卖面临极大的空白市场，绝大多数客户此前从未采购过任何工业自动化机器人。
```mermaid
flowchart TD
accTitle: Tutor Intelligence Closed-Loop Model
accDescr: 真实工业部署与人类远程导师闭环驱动机器人进化的流程图
subgraph Factory["工厂与仓库终端"]
Deploy["规模化部署机器人"]
Task["执行真实物流/制造任务"]
end
subgraph Intervention["远程介入系统"]
Human["人类操作员 (Tutor)"]
Teleop["接管与示范操作"]
end
subgraph DataEngine["数据与模型引擎"]
Data["收集真实物理世界标注数据"]
Train["训练前沿具身 AI 模型"]
end
Deploy --> Task
Task -->|"遇到长尾复杂问题"| Teleop
Human --> Teleop
Teleop --> Data
Data --> Train
Train -->|"能力迭代升级"| Deploy
```
---
### `05:00` 机器人是终极系统工程与全栈纵向化必然
针对 a16z 合伙人 Martin Casado 关于其从 MIT 机器人技术背景转向极度关注单位经济模型（Unit Economics）、供应链与 GTM（Go-to-market）商业化的提问，Josh Gruenstein 强调，这并不意味着底层技术不再重要，而是因为机器人绝非单点突破的技术，而是一个“终极系统工程”（Ultimate Systems Problem）。如果在设计时人为切除模型、硬件、制造、供应链或客户需求中的任何一环，就会主动放弃全局权衡（Trade-offs）的探索空间。许多机器人初创公司最终在经济效益上未获实质成功，一方面是因为缺乏核心智能图元以拉平单位经济，另一方面则是由于市场极度碎片化，各方相互割裂。
谈及机器人企业为何大多从“横向通用定位”起步最终却纷纷走向“垂直全栈整合”（购买 3D 打印机自研硬件制造），Josh 认为这并非物理世界无法抽象的本质缺陷，而是当前市场成熟度不足所带来的阶段性现象，正如早期计算机工业的发展轨迹。未来，像人形机器人（Humanoids）这样的通用形态在物理世界中就扮演着类似通用 CPU 的干净抽象层；随着行业规模不断扩大，机器人也将重走计算领域“专用化机器 -> 通用计算平台（CPU） -> 再次分化（GPU/ASIC）”的演进路径。在此之前，要构建一个能让未来软件创业者无需关心底层硬件即可运营的通用平台，唯一行之有效的方式是亲自下场搭建垂直全栈平台，并用真实的业务闭环来验证。正如 Josh Gruenstein 所总结的：
> “And then starting a business, it became really clear to me that like, oh, okay, robotics is not actually full stack, like model hardware, software. It's full stack, like model hardware, software, supply chain, customers, product, right?”
此外，Josh 深刻对比了数字智能与具身智能在数据采集上的根本鸿沟：互联网天然汇集了人类发布在社交平台、电子邮件中的数字交互，大语言模型（LLM）得以利用这笔丰沛的副产物数据完成训练；但物理世界并不存在这种天然联网机制，正如他在对话尾声所生动比喻的——自己在现实中随手端起一个咖啡杯，互联网上没有任何系统会感知到这一动作，物理世界的数据必须依靠主动部署的终端才能捕获。
### `10:00` Tutor Intelligence 的核心理念：构建真实世界的数据飞轮
谈及公司创立的初衷，Josh Gruenstein 指出，物理世界机器人面临的最根本短板在于缺乏高质量的交互动作数据——世界上并没有任何记录人类指尖压力、三维姿态等细致物理特征的现成资产。与大语言模型能够轻易爬取整个互联网的语料不同，试图单凭算法“无中生有”地训练出具身智能（Embodied Intelligence），在统计学和工程逻辑上都无异于一场胜算极低的赌博。计算机视觉和自然语言处理的发展历史已经证明，海量数据才是破局前提。
因此，Tutor Intelligence 的破局思路是构建一个部署在现实工厂中的机器人网络，并通过人类远程操作员（即“导师”/ Tutors）在机器人遇到困难或学习新任务时进行介入教学。正如互联网用户在日常活动中“不经意间”排放并积累了足以训练前沿大模型的海量文本一样，Tutor 意在通过商业部署，自发沉淀出一套“机器人训练数据的互联网”。这种模式不仅无需承担凭空烧钱的庞大资本开支，还在铺设数据飞轮的同时创造了真切的商业价值，直接降低了实体工厂的制造成本。Josh Gruenstein 强调：
> “And so when we thought about, okay, what are we going to name this business? You know, it's not strictly artificial intelligence. There's some artificial intelligence, but it's ultimately like this tutor intelligence that we're building that was connecting and controlling these robots.”
Raghu Raghuram 补充道，在他们投资前与客户访谈时发现一个有趣的现象：大多数工厂客户根本不知道、也完全不在乎眼前的机器臂究竟是由自动模型驱动，还是后台有人类 Tutor 在远程操作，这种无缝衔接正证明了产品的成熟度。Josh 笑着回应，评判机器人在产业中是否成功的金标准绝不是外观有多抢眼，而是务实的生产结果：
> “You want the feedback to be like, my shampoo is that I'm manufacturing at volume. It's now five cents cheaper than it was. If the feedback is like the robot is pretty, it's like, oh, it's probably a real bad robot.”
```mermaid
flowchart LR
accTitle: Tutor Intelligence 数据与商业飞轮
accDescr: 展示现场部署、人类远程教学、模型能力进化与商业收益的闭环关系
subgraph Operations ["现场与操作层"]
A["工厂真实生产任务"] --> B["现场机器人作业"]
B -->|"遇到长尾/复杂动作"| C["人类远程介入 (Tutors)"]
end
subgraph DataEngine ["数据与智能层"]
C -->|"采集高价值交互数据"| D["沉淀真实环境训练集"]
D -->|"训练升级"| E["具身基础模型"]
E -->|"提高自主执行率"| B
end
subgraph ValueLayer ["商业价值层"]
B --> F["单件生产成本下降"]
F --> G["客户付费支撑商业闭环"]
G -.->|"支撑更多设备部署"| A
end
```
---
### `12:00` 经济副产物与数据有效性：为何不能靠“人肉采集”？
关于未来是否会彻底淘汰人类 Tutor，Josh 认为这本质上是一个算力与人力的单元经济学（Unit Economics）平衡问题。这就好比评估一块英伟达 H200 GPU 的利用率和占空比一样，企业应当将人类智慧视为一种高阶、异构的智能资源输入。为了消灭模型最后“几个 9”的长尾极端情况，究竟是无休止地堆砌工程研发成本更划算，还是直接接入兼职人类 Tutor 更具成本效益，完全取决于经济账。
Martin Casado 提出质疑：生成式 AI 的奇迹在于“输入什么分布，产出即满足该分布的商业价值”（训练文本便能输出文本工具）。然而机器人的严峻之处在于其解空间维度极高，哪怕模型能学会 90% 的已知操作，剩下 10% 稍微脱离分布（Out-of-Distribution）的微小变动，就可能让机器人彻底卡壳，从而需要源源不断的人力长期填坑。
Josh 分析道，核心命题在于“数据到底从何而来”。如果一家企业是作为应用层工具深深嵌入在软件工程师的工作流中，它自然而然就身处极具经济价值的数据洪流中心。机器人亦然，许多团队尝试给工作人员头戴摄像头、派往印尼等发展中国家满大街盲目采集物理交互，这无异于碰运气，因为没有市场的“无形之手”能够检验这些数据的实际价值。唯有让机器人真正运行在客户愿意自掏腰包买单的工业产线上，作为生产副产物沉淀下来的动作与决策，才是板上钉钉的高商业价值数据：
> “There's no invisible hand of the market that's guiding you to ensure that that data is actually economically valuable. But if it's actually this kind of emission byproduct, if you have robots in real customer sites that customers are paying for, for those robots to do that task... you know for sure that that data is economically valuable.”
- **多模态补充机制**：团队并非纯粹的“实机数据至上主义者”，而是通过仿真（Simulation）与互联网规模的公开数据配合，主动拓宽长尾场景的数据分布。
- **动态经济权衡**：如果让人类介入能直接维持极佳的毛利与生产平稳度，团队完全乐于接受“人机协同”，不盲目追求绝对意义上的完全无人化。
---
### `15:00` 空间推理的质变与数据飞轮的代际壁垒
讨论转向大模型对机器人领域的技术溢出效应。Josh 表示，近几年学术界和前沿实验室（如 Google Astra 等多模态工作）的突破为机器人带来了翻天覆地的“剧变”（Sea change）。长期以来，行业都知道大模型内部蕴藏着极其广博的物理世界通识，但苦于缺乏最后一步的“空间推理（Spatial Reasoning）”桥梁，无法将 GPT 类通用知识有效输入给动作扩散 Transformer（Action Diffusion Transformer）这类具身执行端。
如今，顶尖 AI 实验室因下游 3D 建模、CAD 设计等软件市场的强劲变现需求，开始大规模对基础模型进行 3D 空间数据后训练（Post-training）。令人欣喜的是，这一为了软件变现而训练出的三维空间理解能力，恰好直接填补了工业机器人动作规划所需的核心拼图。
面对大模型通用能力快速进化的外部环境，Josh 直言这不仅不会侵蚀 Tutor 的生存空间，反而极大地助推了业务迭代。Tutor 自 2021 年成立以来，经历了从卷积神经网络（CNN）到现代扩散架构的数次技术换代，深知底层的算法小技巧（Architectural tricks）会不断过时，但穿透周期的核心资产始终未变：
> “We started this business in 2021. We were doing this initially with convolutional neural networks, right? So like we've, we've gone through every sea change and, and the thing that is always valuable is the flow of data, right? And the flow of data is, is always the thing that uniquely you can do.”
> “Ultimately, we think the thing that has kind of the longstanding generational value in a business is... is the flywheel and loops themselves.”
---
### `17:30` 重构硬件供应链：从零自研与主力机型 Cassie
探讨到硬件路线时，Martin Casado 追问 Tutor 的造机模式到底处于光谱的哪一端——是直接购买 Unitree（宇树）等现成四足/双足本体写 Prompt，还是自己从最底层的机加工做起。
Josh 给出了一组直观的背景数据：在 2025 年全美仅部署了约 3.8 万套工业机器人工作单元（Industrial robot work cells），这与全国数百万的潜在工位需求相比微不足道；而且其中几乎全都是 Fanuc（发那科）、ABB 等传统老牌工业臂，没有任何像 Unitree 这类新型机器人的规模部署案例。更关键的是，传统工业机器人的硬件设计理念主要是为了服务系统集成商（Systems Integrators），其机械刚性与高加减速是针对固定、重复的绝对轨迹优化的，完全不适配与前沿基础模型进行软硬件协同设计（Co-design）。
因此，Tutor 不得不在硬件层面白手起家。虽然团队不自研底层电机执行器（Actuators）和摄像头等由成熟供应链规模供应的标准件，但整台能够自主移动、带有机械臂、能执行多样任务的整机系统，全部由 Tutor 在马萨诸塞州沃特敦（Watertown, Massachusetts）的工厂自主设计并规模化装配。
当谈及机器人的形态学（Morphology）时，Josh 介绍了公司目前规模化量产部署的主力重载机型 Cassie（另一款为 Sunny）：
| 机型名称 | 定位与核心形态 | 模型架构与驱动机制 | 核心落地场景 |
|:--- |:--- |:--- |:--- |
| **Cassie** | 重载型“粗重活机器（A brute）”，配备单机械臂与高负载真空吸盘阵列 | 2 亿至 4 亿参数量级（200M–400M）的基础模型，负责预测物体的抓取可操作性（Affordances/Grasps） | 针对美国仓储物流中“一切货物都基于托盘（Pallet）或纸箱运输”的国情，专职替代高负荷的托盘拆码垛与人工搬箱工作 |
谈及将特定硬件与市场需求深度结合的设计初衷，Josh 概括道：
> “So Cassie is kind of a brute, right? Cassie is designed to do heavy duty work. It has a single arm and a lot of suction cups, and it's powered by foundation models in like the 200 to 400 million perimeter range and kind of came out of this world that Tutor was started out of in 2021 of like, what can foundation models do for robots?”
### `20:00` 半人形态机器人 Sunny 与真实工厂中的劳动现状
谈及 Tutor Intelligence 的另一款机器人形态，Josh Gruenstein 介绍了工业级半人形态机器人 Sunny。如果说普通人形机器人（例如宇树科技的产品）追求的是机械姿态，那么 Sunny 则是完全围绕“在工厂中以人类的作业速度和高可靠性连续运行 3 到 10 年”这一工业诉求做出的工程权衡。工业室内环境基本都是平整打磨过的硬质地面，因此 Sunny 采用了轮式底盘而非双足；同时它具备极高的负载能力、更快的运行速度，并且通过了工业协作安全标准（collaborative safety standards），能够与人类并肩工作。Tutor 不仅在内部通过大量 Sunny 收集数据以推进研发，更有不断扩大的 Sunny 机队部署在真实客户现场，负责高灵活度的双手协同操作任务。
```mermaid
flowchart TD
accTitle: Sunny 与传统足式机器人的工业设计权衡
accDescr: 对比工业半人形态机器人 Sunny 针对室内工厂所做的硬件和形态选型优势。
subgraph Design["工业化设计权衡"]
A["作业环境: 平整硬质地面"] --> B["移动底盘: 轮式替代双足"]
C["工业标准要求"] --> D["高负载能力 & 高速移动"]
C --> E["协作安全标准 (可与人同台作业)"]
end
subgraph Deployment["实际场景应用"]
B & D & E --> F["Tutor 内部数据工厂: 推进模型研发"]
B & D & E --> G["真实客户车间: 双手高灵活性分拣与组装"]
end
```
当被问及 Tutor 是否需要开发 30 种不同形态才能覆盖市场时，Josh Gruenstein 表示：“如果你只想深耕制造与物流领域，今天的 Sunny 和 Cassie 大概就能承担 80% 到 90% 的工作。”大众对美国工业和制造业物流的认知多半来源于科普电视节目《制造的秘密》（*How It's Made*），以为工厂全是不见人影、自动化机械臂高速飞转的复杂鲁布·戈德堡式机器（Rube Goldberg machines）。但那种全自动化产线仅代表全美制造与物流业最顶尖的 0.01%。当你真正走进全美占地百万平方英尺的各类普通厂房，映入眼帘的往往是成群的工人们站在传送带或折叠桌旁，全凭纯手工在完成枯燥甚至折磨人的繁重工作。
- 劳动力短缺极其严峻：工厂中需要弯腰搬运的重活往往招不到人，工人平均流动周期仅为 3 个月，留下的核心工人平均年龄高达 55 岁，即将面临退休潮。
- 绝大部分组装与包装仍依赖人工：从最普通的盒装消费品到高度精密如 iPhone 的外包装，最终都离不开人工装配环节。
- 轮式形态的局限性：当前的 Sunny 和 Cassie 无法胜任室外工况，并非动力不足，而是底盘硬件无法在砾石等复杂地形上行进；未来针对建筑工地、垃圾填埋场资源回收等场景，具有足式行进能力的机器人将是形态图谱中的下一块拼图。
- 家用场景尚不成熟：当前的工业机器人外形和体积过于硬核，普通人不会希望把如此庞大且“吓人”的工业设备直接放入家中。
### `23:40` 时薪 16 美元带来的无摩擦商业模式与飞速落地
在谈及客户采纳（Customer Adoption）时，Josh Gruenstein 指出目前美国工厂在自动化上面临巨大的空白——全美每年部署的机器人仅有 3.8 万台左右，十分之九的工厂和仓库完全没有任何机器人，几乎整个市场都属于“荒原”（greenfield）。许多工厂老板早已见惯了老旧昂贵的传统自动化方案，先入为主地认为机器人根本不适合自己的厂子。因此，Tutor 在接触客户时完全避开 AI 模型、空间智能或人类导师系统等宏大技术概念，而是直接给工厂主打电话提出最直白的方案：“嘿，你考虑过雇一个时薪 16 美元的机器人吗？”这直接打破了传统认知。
```mermaid
flowchart LR
accTitle: Tutor Intelligence 极简销售与交付周期
accDescr: 展示 Tutor 从触达工厂客户到实现机器人量产上岗的完整时间节点。
A["首次销售电话: 时薪 16 美元方案"] -->|"40 至 60 天"| B["签署部署合同"]
B -->|"约 40 天"| C["机器人运抵车间现场"]
C -->|"数小时至 1 天内"| D["正式并网投入生产"]
```
这种将机器人供应商利益与客户生产成本紧密绑定的商业模式，使分发效率较传统工业自动化提升了数个数量级。传统自动化方案需要漫长的定制开发周期并让客户承担高昂的资本支出风险；而 Tutor 生产的是通用智能机器人，直接从工厂整机出货，客户几乎无需承担经济风险。正如 Josh Gruenstein 解释运营团队为何迅速接受这一方案时所言：“运营管理圈的人特别务实，因为他们看到一个东西有效，那就是有效。就是这么直接。”一线管理层一旦看到某道工序被机器人解决，就会迅速要求将方案推广到全厂所有相同的工位上，使得 Tutor 能够以极快速度从小微企业扩张到供应链头部企业，从单点单台试用裂变为跨厂区的数十台复购。
### `25:40` 工厂交互中的人机关系与“给机器人起名字”
谈到工厂现场人与机器人的社会化关系，Josh Gruenstein 引用了一句耐人寻味的行业俏皮话：“机器人只要还没正常工作，它就是个机器人；一旦它工作正常了，它就成了一台机器。”但在实际场景中，机器人买家（COO 或厂长）与终端用户（每天与机器人搭班的产线工人）是割裂的。由于目前的技术并非要实现整座工厂 100% 的无人化，而是接管特定工种，机器人不可避免地需要和人类工人长时间协作，因此建立融洽的人机交互（Human-Robot Interaction, HRI）变得至关重要。
为了拉近工人与机器人的心理距离并消除抵触情绪，Tutor 引入了命名机制：
- 消除工人敌意：产线工人一旦发现机器人接管了那些让人腰酸背痛的重活，自己的日常劳动强度大幅下降，往往会迅速对机器人产生依赖与好感。
- 专属命名权激发归属感：Tutor 规定客户在追加复购订单时，有权为新到店的机器人命名。这极大激发了客户的主人翁意识，许多现场工人甚至会以自己孩子、最喜欢的橄榄球星或流行歌星的名字来称呼这些机器人。
- 命名规则的迭代：Tutor 最初按照飓风命名法（按 26 个英文字母轮转），但由于机器出货量增长过快，字母表很快就不够用了；如今，制造这批机器人的美国本地装配技术人员会根据自己的喜好和带来快乐的事物自由命名。
### `28:40` 避开 Fortune 50 执念：深入美国本土中小型工厂的蓝海
Raghu Raghuram 提出，许多初创机器人公司习惯于追逐《财富》50 强（Fortune 50）巨头，因为大企业资金充裕、愿意斥资试错，但 Tutor 的绝大多数客户反而是分布在全美各地的中小型本土企业。Josh Gruenstein 指出，这绝非偶然，而是因为这类中小型制造与物流企业构成了全美工业经济中最庞大、最真实、也最迫切需要生产力补充的主体市场。
### `30:00` 机器人公司为何屡屡失败：定制化陷阱与通用产品哲学
Josh Gruenstein 深入剖析了机器人行业的历史教训。他指出，Tutor Intelligence 创立时，正值大量源自高校实验室、专注于供应链物流的学术型机器人初创公司纷纷折戟沉舟。追溯其根源，许多机器人团队压根没有建立真正的“产品公司”思维，而是陷入了致命的定制化工程陷阱：某家巨头企业带着特定痛点找上门，签下一纸承诺未来会大规模采购的合作意向书；创业团队随后闭门造车，耗费整整两年为其研发高度定制化的专用系统；等到实际部署时，系统往往根本无法正常运转——因为任何做过产品的人都知道，产品绝不可能一蹴而就，必须依赖高频迭代。更糟糕的是，即便历经千辛万苦跑通了流程，客户两年前的需求可能早已不复存在，导致初创公司瞬间走向终局。
为了打破这种反复上演的悲剧，Tutor Intelligence 将服务多样化的腰部及中端市场（Mid-market）客户作为核心原则。Josh Gruenstein 将其视为倒逼团队做正确产品的极佳外在约束机制。他们拒绝为个别工业大客户搭建复杂的“鲁布·戈德堡机械”（Rube Goldberg machine），而是专注于打造通用平台——即通用的机械本体搭载通用的大模型，进驻现场就能立刻上岗。如果在两周内就能与中端客户签约并迅速上线作业，这就证明商业模式与产品架构走对了路。
当被问及若这一路径完全走通、行业的终局会是什么时，Josh Gruenstein 认为将带来全方位的质变与物理经济的“断层式跃迁”（Discontinuities）。以预制房屋制造为例，他曾与威斯康星州一家装配式建筑工厂的产线员工交流，对方提到目前在工厂内建房成本比传统工地现场施工高出约 5%；但只要工厂建房的成本比现场施工哪怕便宜 1%，全美开发商的所有订单就会在一夜之间全部涌向工厂化制造。垃圾填埋场资源回收等诸多物理行业同样存在这种阈值拐点。
关于机器人是否会进入家庭场景，Josh Gruenstein 明确表示团队不设限制，但短期内并不可行。这不仅是因为家庭环境在任务复杂度和安全性上的技术门槛极高，更根本的在于经济账算不过来：机器人在工厂中能够产生明确的商业价值，而普通消费者愿意为家庭机器人支付的溢价完全无法支撑其成本。因此，团队的核心战略是在全栈各层级主动做取舍以抢占立足点，随着底层平台不断成熟，需要妥协的环节才会越来越少。
> “And one of the big things I think that robotics companies get wrong is they're not actually really product companies.” —— Josh Gruenstein
> “We don't want to build a Rube Goldberg machine that works really well for big industrial customer A or B. We want to build a platform that is generally capable, robot bodies running generally capable robot models, that can immediately get to work.” —— Josh Gruenstein
> “And so I think our hope is like if Tutor is really successful, we can provide this robot substrate, this labor force that other businesses can use to transform the physical economy and make the world ultimately a cheaper and better place to live in.” —— Josh Gruenstein
---
### `34:00` 硬件敏捷迭代与极简商业模式：零资本支出的即战力
在产品研发机制上，大部分机器人公司至今仍停留在初代原型（V1），而 Tutor 已经迭代出两代成熟机型。为了让硬件迭代拥有接近数字软件的敏捷度，Tutor 选择自主制造机器人。通过完全掌控生产制造流程，团队能以两周为周期将产线工艺改进直接推向生产环节；在软件层面，代码每日推送到端侧机器人，主版本每两周发布一次。与此同时，其研究团队既紧贴客户一线获取高质量实景数据，又承担突破性模型研发的任务。一旦端到端大模型取得突破，软硬件协同设计团队便能将其迅速封装进产品并部署至一线。
对于仓库或制造企业而言，引入 Tutor 机器人的门槛被压缩到极致。客户无需等待长达数年的定制工程，只需在官网上输入邮箱，利用内嵌的定价计算器输入当前运营数据，就能直观测算出采用 Tutor 机器人能省下多少钱——无论是每年节省 10 万美元、100 万美元还是上亿美元。更为关键的是，客户无需预付任何资本支出（CapEx），签约两个月内即可落地并产生真实效益。
谈及市场需求，Josh Gruenstein 强调需求极为旺盛，但客户买的根本不是“机器人”这一概念，而是纯粹的劳动产出。团队内部将机器人的价值公式精简为一句话：**机器人完成的工作总量除以其综合成本**，此外的一切都是多余的解释。在被问及成本结构时，Josh Gruenstein 指出，与传统认知中“前期硬件资本投入巨大、需要十年才能收回投资”的模式截然不同，Tutor 机器人的综合成本（COGS）中，占主导地位的反而是运行过程中的持续成本（Ongoing costs），例如云端/边缘推理算力和人机协同（Teleoperation/远程接管），这一反直觉的事实让很多行业局外人感到惊讶。
```mermaid
flowchart LR
accTitle: Tutor Intelligence Rapid Iteration and Delivery Architecture
accDescr: Diagram showing the tight feedback loop between research, two-week manufacturing iterations, daily code updates, and zero-CapEx customer deployment.
subgraph RND["研发与制造闭环"]
Research["研究团队<br/>(算法突破 & 数据挖掘)"]
Hardware["自主制造产线<br/>(每 2 周产线工艺迭代)"]
Software["软件交付基建<br/>(每日更新 / 双周大版本)"]
end
subgraph Field["客户作业现场"]
Robot["通用机器人本体"]
CloudOps["持续运营与推理<br/>(模型推理 + 远程接管)"]
end
Research --> Hardware
Research --> Software
Hardware --> Robot
Software --> Robot
Robot --> CloudOps
CloudOps -.->|"实景作业数据回传"| Research
```
> “And that's not like, okay, this is a roadmap over five years of custom engineering. It's like, no, you could save that a million dollars a year, two months from today, without putting CapEx down.” —— Josh Gruenstein
> “And so it's ultimately like, you know, we have this framework of the value of a robot is you take how much work a robot can do and you divide it by its cost. And that's it, right? Everything else is commentary.” —— Josh Gruenstein
---
### `36:30` 战略级劳动规划与全球制造业重构：从工业重镇到分布式工厂
客户引入机器人并非为了应对“黑五”等短期促销爆仓，而是一项关乎生存的长期战略决策。当前美国制造业企业面临着严峻的劳动力老龄化问题，同时在各种宏观趋势推动下市场需求持续攀升，用工成本水涨船高。制造业不是拥有丰厚毛利率缓冲空间的纯软件公司，成本账本一旦算不过来，企业就会面临停产倒闭。因此，无论是生产除臭剂还是制造蜡笔的工厂，企业管理者希望专注于自身产品的工艺运营，而迫切需要一个长期战略伙伴来全面托底自动化劳动力。
这进而引出了更深远的全球产业格局变迁。Raghu Raghuram 提问：如果通用机器人可以随时随地部署，传统的“单一工厂小镇”（Factory Town）以及全球工厂的地理分布会发生怎样的改变？
Josh Gruenstein 认为，通用实体智能将成为重塑地缘政治和产业竞争的“伟大平衡器”（Great equalizer）：
- **中美自动化路径的范式代差**：中国在上一轮固定形态的工业自动化浪潮中投入了海量固定资产，这虽然奠定了过去的成功，但从长远来看可能成为历史包袱；因为下一代灵活通用机器人普及后，上一代高度固化的流水线设施可能面临高昂的拆除与改造成本（Tear down）。相反，美国由于制造业出现断层，可以直接在白地（Greenfield）上依托通用大模型和灵活机器人建立全新形态的工业产能。
- **物理世界的“云计算时刻”**：正如云计算基础设施和现代软件工具的大爆发催生了海量轻量级数字企业一样，通用的机器人劳动力底座将彻底打破实体创业的重资产门槛。未来，全球工厂将不再高度集中于极少数超大型产业集群，任何个人或团队都能低门槛地开启并快速扩展一家物理实体的制造企业。
> “These businesses, I think, are faced with a really hard choice, which is they know the age of their workforce and, they know that demand is ramping up due to all these secular trends and the cost basis of fulfilling that work is going up.” —— Josh Gruenstein
> “I think one big thing is this is going to be a great equalizer between the United States and other countries like China invested in the previous regime of industrial automation.” —— Josh Gruenstein
> “Anybody will be able to start a business in the physical world. Anybody will be able to scale a business in the physical world if we kind [of provide this substrate]...” —— Josh Gruenstein
### `40:00` 供应链依赖、本土化制造与垂直整合
谈及机器人领域的全球供应链格局时，主持团队提到他们曾就机器人供应链发布过一份研究报告，显示中国在其中占据着极为核心的地位。针对 Tutor Intelligence 对海外供应链的依赖度及其长期演变趋势，Josh Gruenstein 直言不讳地指出，无论外界怎么说，在当前节点全球供应链都处于高度互联和相互依赖的状态，机器人硬件只是其中的一个典型缩影。中美两国的企业都在极为严肃地思考这一格局的长期演变方向，这涉及中国企业在美投资、美国企业在华投资，以及双方竭尽全力增强各自本土工业制造能力的博弈。
```mermaid
flowchart TD
accTitle: 机器人硬件垂直整合演进路径
accDescr: 展示从采购通用组件到驱动OEM代工，再到建立美国本土制造能力与自研组件的演化逻辑。
subgraph Phase1["初创阶段：家庭作坊式采购"]
A["采购现成标准件<br/>（外部专业 OEM 厂商）"]
end
subgraph Phase2["规模拐点：需求反客为主"]
B["出货规模急剧攀升<br/>采购量占 OEM 绝大份额"]
C["OEM 沦为事实上的合同制造商<br/>（由 Tutor 决定指令与技术规范）"]
end
subgraph Phase3["深层自研：本土垂直整合"]
D["在马萨诸塞州沃特敦组装整机"]
E["自研核心硬件组件并推动美国本土化制造"]
end
A --> B
B --> C
C --> D
C --> E
```
随着 Tutor 的部署规模迅速扩大，主持人注意到该公司正在进行深度的垂直整合，将更多环节收归内部。Josh 解释道，机器人产业目前仍带有很强的“作坊工业”（cottage industry）特征；初创期你可能依赖 OEM 提供各种子部件，但一旦采购规模占据了对方总业务量的实质性比例，这层关系的性质就彻底改变了——这些供应商不再是传统意义上的 OEM，而是退化成了拿订单办事的代工厂（glorified contract manufacturers）。此时企业就必须算账：是继续把生产外包给代工厂并指导他们生产，还是干脆自主制造？鉴于 Tutor 已成为全美部署量最大的机器人企业之一，也自然成了众多硬件厂商的最大买家，未来必将持续加深对关键组件的掌控。
- “I think ultimately, no matter what anybody says, today in 2026, we have a highly interconnected, interdependent global supply chain.” —— Josh Gruenstein 解释当前全球供应链紧密交织的客观现实。
- “I think ultimately, for Tutor to be successful, we will need a mature robotic supply chain in the United States.” —— Josh Gruenstein 阐明 Tutor 致力于重塑美国本土机器人工业基础的核心原因。
- Tutor 目前在马萨诸塞州沃特敦（Watertown, Massachusetts）设立了整机制造基地，并计划把越来越多零部件的生产留在美国本土。
---
### `42:47` 借鉴大语言模型：强化学习与 Data Factory One
在探讨具体的研究与学习架构时，主持人回忆起双方接触之初的情形——虽然 Tutor 的商业落地进展抢眼，但底层核心始终离不开强有力的前沿研究。关于机器人具身智能该押注何种学习范式，Josh 分享了他们从大语言模型（LLM）发展历程中总结的得失：LLM 固然需要吞吐海量异构数据，但处在整个数据金字塔尖顶、最终拉开能力差距的，永远是真实世界中的强化学习轨迹（real-world reinforcement learning traces）。
为了将这一范式照进现实，Tutor 搭建了一整套面向实战经验捕获的研究体系，并在波士顿建立了一个由 100 台机器人构成的自动化数据收集中心——“Data Factory One”。Josh 强调，这是全美已知规模最大的半人形（semi-humanoid）数据农场，它充当了大规模探索真机强化学习的研究试验场。强化学习天然受制于数据吞吐，在小批量样本（small batch sizes）下几乎无法发挥威力；学界和以往的工业界之所以难以施展这一方法，根源就在于缺乏大规模真机集群，而拥有庞大机队的 Tutor 恰恰填补了这块拼图。
- “And I think one thing that clearly has been very, very important in LLMs as there are lots of heterogeneous sources of data, but ultimately there's a top of the data pyramid, which is real-world reinforcement learning traces doing the thing you want to be able to do.” —— Josh Gruenstein 剖析机器人学习为何必须以真实任务的强化学习轨迹为核心。
- “Data Factory One in Boston”是 Tutor 设立在波士顿的核心科研枢纽，由 100 台机器人同时运转，为大规模在线强化学习提供源源不断的真机经验。
---
### `44:03` 跨全技术栈的人才吸纳与软硬件穿梭日常
如此不同寻常的研发路径对团队架构提出了截然不同的要求。当被问及当前的人才招募需求时，Josh 脱口而出：“我们正在招揽所有人（We're looking to hire everybody）。”这不仅涵盖从顶层软件、底层系统到硬件工程的各领域研究员与工程师，还囊括了负责在车间组装机器人的现场技术员。Josh 强调，机器人系统是一个庞大而严密的整体，团队绝不能在任何一层推卸责任（abdicate responsibility）——无论是软件、硬件、材料科学还是电子工程，只要踏入机器人领域，客观现实都会强迫你面对这些挑战，唯一的区别仅在于你究竟懂不懂它们。
这种软硬一体的全栈环境造就了极其独特的工作氛围与文化。Josh 笑称，这大概是全宇宙极少数能让你在同一天经历两种极端现实的地方：早上还在白板前热烈推导最新的模型架构，下午就得蹲在工厂的流水线旁边，屏息凝神地观察机器人在实际生产工况下的运转情况。除了沃特敦与波士顿的研发中心，Tutor 的业务与服务网络已覆盖全美，遍布各地的销售、GTM 团队以及技术支持人员正与全国的客户工厂保持着紧密联动。
- “It's robotics. You're going to engage with it whether you want to or not. The question is whether you understand it.” —— Josh Gruenstein 总结机器人研发对全栈工程能力不可回避的苛刻要求。
- 团队的日常工作场景呈现高度反差：上午研讨最前沿的模型数学结构，下午深入工业车间排查机器人执行瑕疵。
## Notable quotes
> `34:00` And so it's ultimately like, you know, we have this framework of the value of a robot is you take how much work a robot can do and you divide it by its cost. And that's it, right? Everything else is commentary. — Josh Gruenstein
> `10:00` You want the feedback to be like, my shampoo is that I'm manufacturing at volume. It's now five cents cheaper than it was. If the feedback is like the robot is pretty, it's like, oh, it's probably a real bad robot. — Josh Gruenstein
> `30:00` We don't want to build a Rube Goldberg machine that works really well for big industrial customer A or B. We want to build a platform that is generally capable, robot bodies running generally capable robot models, that can immediately get to work. — Josh Gruenstein
> `30:00` And one of the big things I think that robotics companies get wrong is they're not actually really product companies. — Josh Gruenstein
> `42:47` And I think one thing that clearly has been very, very important in LLMs as there are lots of heterogeneous sources of data, but ultimately there's a top of the data pyramid, which is real-world reinforcement learning traces doing the thing you want to be able to do. — Josh Gruenstein
> `15:00` Ultimately, we think the thing that has kind of the longstanding generational value in a business is... is the flywheel and loops themselves. — Josh Gruenstein
> `05:00` And then starting a business, it became really clear to me that like, oh, okay, robotics is not actually full stack, like model hardware, software. It's full stack, like model hardware, software, supply chain, customers, product, right? — Josh Gruenstein
## Mentioned
Tutor Intelligence, Josh Gruenstein, Martin Casado, Raghu Raghuram, a16z, MIT, Google Astra, Cassie, Sunny, Data Factory One, Unitree, Fanuc, ABB, Nvidia H200
## Video description
What does it take to build robots that can actually work in factories and warehouses? Tutor Intelligence CEO Josh Gruenstein argues robotics is a systems problem spanning AI, hardware, manufacturing, and the work itself.
Josh joins a16z General Partners Martin Casado and Raghu Raghuram to explain Tutor’s approach: deploy robots into real businesses, use human “tutors” when needed, and turn that work into training data that makes the robots more capable.
They discuss Tutor’s 100-robot data factory, robots starting at roughly $16 an hour, advances in spatial reasoning, and what flexible robotic labor could mean for American manufacturing and the physical economy.
[Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/9130a518-123f-457b-8d4e-90650dff71d0/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=9130a518-123f-457b-8d4e-90650dff71d0&feed=JGE3yC0V)
Source quotes
[34:00] Josh Gruenstein: And so it's ultimately like, you know, we have this framework of the value of a robot is you take how much work a robot can do and you divide it by its cost. And that's it, right? Everything else is commentary.
[10:00] Josh Gruenstein: You want the feedback to be like, my shampoo is that I'm manufacturing at volume. It's now five cents cheaper than it was. If the feedback is like the robot is pretty, it's like, oh, it's probably a real bad robot.
[30:00] Josh Gruenstein: We don't want to build a Rube Goldberg machine that works really well for big industrial customer A or B. We want to build a platform that is generally capable, robot bodies running generally capable robot models, that can immediately get to work.
[30:00] Josh Gruenstein: And one of the big things I think that robotics companies get wrong is they're not actually really product companies.
[42:47] Josh Gruenstein: And I think one thing that clearly has been very, very important in LLMs as there are lots of heterogeneous sources of data, but ultimately there's a top of the data pyramid, which is real-world reinforcement learning traces doing the thing you want to be able to do.
[15:00] Josh Gruenstein: Ultimately, we think the thing that has kind of the longstanding generational value in a business is... is the flywheel and loops themselves.
[5:00] Josh Gruenstein: And then starting a business, it became really clear to me that like, oh, okay, robotics is not actually full stack, like model hardware, software. It's full stack, like model hardware, software, supply chain, customers, product, right?
Source quotes
[25:00] Josh Gruenstein: A robot is a robot unless it works, in which case it’s a machine.
[18:25] Josh Gruenstein: Anything in the physical economy has to be transported on a pallet or in a cardboard box.
[5:00] Josh Gruenstein: The only way you can make forward progress is by figuring out all of the clever tradeoffs across as many layers of the stack as you can
[43:08] Josh Gruenstein: It's robotics. You're going to engage with it whether you want to or not. The question is whether you understand it.
Original transcript
[0:00] No matter what anybody says, today in 2026, we have a highly interconnected, interdependent global supply chain. Both U.S. companies and Chinese companies are thinking very hard about what is the long-term strategy of this look like? And ultimately, everybody trying to build their industrial capacity. We see a lot of robot pitches. Very few people are thinking about what the robot is supposed to do at the end of all of this. Almost everybody we're selling a robot to has never purchased a robot before. Almost all of the market is greenfield. The reality is you need to build a full stack platform. Tutor robots have helped manufacture and ship over a quarter of a billion products to American consumers. We're training frontier robot AI models, manufacturing robots at scale, and directly deploying into the largest businesses in the supply chain. How does product innovation work in robotics? Most companies are just on V1, whereas you've got two versions of the robot and you continue to iterate really fast. on people for physical work. What does it take to build robots that can actually do those jobs? In this episode, Martine Cassado and Raghu Raguram sit down with Tudor co-founder and CEO Josh Gruenstein to unpack a different approach to building general purpose robots. Tudor deploys robots into factories and warehouses, connects them to human operators when needed, and uses that real-world work to generate training data that makes the robots increasingly capable. manufacturing and go-to-market at the same time, and why Josh believes general-purpose robots could ultimately become a new labor substrate for the physical economy. Josh, thanks for joining us. So we're with Josh Gruenstein, who is the CEO and co-founder of Tutor Intelligence. And we're going to be talking about robots today. And so maybe before we jump in, just give a quick background, maybe of yourself and then the company. Yeah, so I'm Josh. I started Tudor about five and a half years ago with the core thesis that If you could just start deploying robots out into the world and all the places in the physical economy where robots are needed the most, you could start to build the data sets, the models and the supply chain that's needed to proliferate robots throughout the physical economy. So we're about 160 people based primarily in Boston, and we're training frontier robot AI models, manufacturing robots at scale and directly deploying into the largest businesses in the supply chain. Maybe talk a little bit about, we see a lot of pitches. There's a lot of companies that talk about doing robotics, but don't actually deploy robots or have not yet. And you're one of the few companies that's actually deployed robotics at scale. So maybe just talk about the state of the business. most American consumers are a Tudor customer a couple layers down. I think this comes from kind of The origin story of the business, which is us looking at the state of robot learning, it was obvious to us that robot intelligence and robot software was this bottleneck, right? Robots for 60 years has been this custom engineer machines that can't scale. Yeah, so we started the company with this thesis that if we could just start deploying robots, we could build the label data sets that help us train these models and deliver this value to the American consumer. We see a lot of robot pitches. 99% of them start with some insights about the model and how you're going to build a model or some insight about the data, shape of the data that is. And very few people are thinking about it. what the robot is supposed to do at the end of all of this. You started from the reverse. You had the insight that you could actually Tutor to the robot, so to speak. That's the name of your company. Where did that insight come from? It came from kind of this thought experiment where we had this insight that, hey, there are 8 billion people on Earth who all have this baseline physical intelligence that we want our robots to have. And so imagine you had infinitely powerful software that could connect any human being on Earth to any robot and kind of embody that robot with intelligence. That's kind of like this deus ex machina of like suddenly you have unlimited infinite robot intelligence. Let's say you take that for granted. Now what? The problem isn't actually solved. You still need to actually manufacture robots at scale. Robots are this cottage industry today in the United States. Robots are rounding error, even in the most automated industries like manufacturing logistics. So that doesn't exist. Certainly none of the software exists and none of the customers exist because, again, the robots can't do any of the jobs that customers want robots to do. It's about content.
[5:00] kind of, figuring out how to put all these three pieces together And it's the combination of all three that actually lets you build, in my view, a successful platform at any one of those layers. Yeah, internally we talk about robotics being the ultimate systems company as opposed to being one of the other. You know what's funny about this though, and the question I wanted to ask you is, I mean, you've got a pretty deep technical background. I mean, MIT robotics, is that fair? But whenever I talk to you, I feel like I'm talking to an operations guy that talks about unit economics and operations and supply chain. So, I mean. I mean, I go to market guy. I go to market guy. So it's kind of you've done this like the transformation is complete. And so maybe this is kind of an open ended question. But like, I've always wondered, is the harder problem here actually that stuff? No. But I think you have to do all of it is kind of the thing. And like when I was just a roboticist, right, only working on the technology side of this problem, I think my strength as a roboticist was that it was very full stack. And I was willing to think about the manufacturing, the hardware. I was willing to think about FPGAs and ASICs. And if you view robotics in the same way that I do, that it's this like ultimate systems problem. The only way you can make forward progress is by figuring out all of the clever tradeoffs across as many layers of the stack as you can. of the layers of the stack and you're like, well, we're not going to touch that. you've eliminated 90% of your trade-off space to make progress. And then starting a business, it became really clear to me that like, oh, okay, robotics is not actually full stack, like model hardware, software. It's full stack, like model hardware, software, supply chain, customers, product, right? Exactly. And similarly, you asked the question, like, why hasn't robotics worked, right? There's been a lot of robot startups. I would argue no robot startup has been like really massively economically successful. Why is that? And I think it's because, one, you don't have this core primitive, which is intelligence, and that lets you kind of bring... the unit economics of the system to a certain level, but two, it's that it's the super sort of disaggregated market. And you have infinite players, and they're all trying to work with each other. And the reality of what we need in robots is to be able to make tradeoffs across the full stack. So this is an observation that we have had. It's interesting to articulate it this way, which is, Most robot companies start horizontal. and they all go vertical. Everyone, they literally start like, we have a model or whatever. And then they start buying 3D printers. And do you think that that is kind of because the market is so immature, like you had to do in the early days of cars, or like you think it's actually inherent in this domain? Like even in 20 years, you will have to verticalize just because the economy of movement or something requires that. No, I think it's a maturity in the market, and you can see similar things in the early computing market. And so to us, the question then becomes, okay, how do you build the general platforms... that let people build businesses that aren't robotics businesses, right? We have all these incredible software businesses today, and the people who run them don't have to think about chip architectures, right? How do we make that possible in the physical world with robotics, and how do you go about building that platform? And I agree with you, people have tried to solve that problem by thinking about very narrow layers of the stack. And in our view, the reality is you need to build a full stack platform. And the only way to know that full stack platform really works is to actually have successful business units operating on top of it. I mean, are you open to, I think that's a good thing. it being the case that actually the physical world is different than the digital world. And the reason we horizontalized compute is because you have clean abstractions that are generally applied. And that's just not the case for like navigating physical space or. I think the beauty of something like a humoid is it's the same. It's that clean abstraction. You think about what is a CPU, and it's this general computing machine. And it's not perfectly optimized for every use cases, but it's general enough that you can run most computing worlds on. So that's your horizontal component. Exactly. For example, would you use the same... brain or like subcomponents to make a human as you would to make a drone or to make some other form factor? No, but you can make that same analogy and compute, right? We have CPUs, we have GPUs, we have ASICs, right? So I think you track the maturity of the market and you start in this world of specialization, custom machines, because that's the only thing the market can support. You move towards general purpose. And in robotics, this is going to look like humanoids and similar form factors. And then eventually you kind of break past to that respecialization as the market is big enough. Why is this thing called Tutor intelligence? So Tutor was founded kind of with this thesis that there is a whole internet of human intelligence that is networked, right? That is posting on social media, that's writing emails. And that network of human intelligence for LLMs and other kind of digital world models, by natural emission, has created this beautiful data set that we can use to train general purpose models. similarly networked, right? Like I pick up my coffee cup and nobody knows, right? Like nobody
[10:00] recorded all of the pressures of my fingertips and where they were in the poses. And so we don't have this asset. And... The core idea behind Tutor is, what if instead of trying to find kind of this algorithm where from no data, we could build urban intelligence. And for us, we're like, that seems like a problem that's rigged against us, right? That's not how this worked in computer vision. It's not how this worked in NLP. What if we could build this network of robots out in the real world? We could connect them with human teleoperators, which we call tutors, that teach the robots how to do new tasks. And in the same way, the Internet kind of by a mission created this data set to train these frontier models. build this internet of robot training data. And there are all these beautiful kind of emergent properties where like, okay, suddenly it's not this massive capital project. It is to go build this network, but also you're creating incredible economic value as you're doing it, right? You're making everybody's goods cheaper as you deploy these robots and collect this data. And so it's very positive sum, I think, in this very unique way. And so when we thought about, okay, what are we going to name this business? you know, it's not, strictly artificial intelligence. There's some artificial intelligence, but it's ultimately like this tutor intelligence that we're building that was connecting and controlling these robots. Yeah, that's an amazing insight. And the funny thing is when we talked to your customers before we invested, most of them didn't seem to know or didn't even seem to care that there was a tutor or when the tutor was operating and when your learning model was operating. And that credit to you guys that you made those very, very seamless too. Yeah, this is to me a signal that you know the robots are good, right? You want the feedback to be like, my shampoo is that I'm manufacturing at volume. It's now five cents cheaper than it was. If the feedback is like the robot is pretty, it's like, oh, it's probably a real bad robot. Is the goal to not have tutors in the future or does it matter? H200, right? And you can math that out. And what's the duty cycle and utilization of that H200? And you can think very similarly about human intelligence, right? What is sort of the heterogeneous profile of intelligence that you really want from humans that you can't get yet from your models? And then it becomes a unit economics question, right? Like, how much engineering do you have to do to get that last nine out of your models versus using tutors? And really, you know, going back to what we're talking about, you want to be able to make as and it's like this work. I mean, you know, it's tempting to say that, like, the economic marvel that happened with generative AI can also happen here. And the kind of the marvel that happened with generative AI is, like, you train these, you know, these models with, you know, some distribution of data, and it turns out basically the same distribution is economically viable, right? It's like you train it on images and people on images, you train it on text and people, you know, it's like kind of distribution in, distribution out. with these robots, right? You could train it on a distribution of data, but then actually what you need is just slightly out of that distribution. So how, I don't know, you probably have the best visibility into how this scales and what the convergence is and whether you can actually remove tutors. What is your intuition on that? I think this is one of the big questions, and I think really it ultimately comes down to what is the data and where are you getting it from? And the magic of the internet is if you are a company and you provide the application layer, tools for software engineers to accelerate their software development, you are naturally in flow. Of. all of this really, really, really economically valuable work, right? If you can make money as a then you're probably also making data as a business, right? Those like economic incentives are aligned. And similarly for robots, if you're like, okay, we're just going to strap cameras and everybody's heads and send them throughout Indonesia to collect data doing X, Y, and Z, you have no guarantee, right? You can guess, but There's no invisible hand of the market that's guiding you to ensure that that data is actually economically valuable. But if it's actually this kind of emission byproduct, if you have robots in real customer sites that customers are paying for, For those robots to do that task... you know for sure that that data is economically valuable. Doesn't it kind of depend on also the dimensionality of the solution set, though? It could be the case that the economically valuable stuff is so high-dimensional that you could solve 90% of it, but the last 10% – you know, you'll need a human being to do for a while just because getting that data set is just going to require too many tutors over a period of time. Yeah, and that might be fun, right? You might be willing to accept that in the economics. That also might be a place where other forms of data matter. I think we're not purists. Again, we're willing to take any tradeoff, right? We love using simulation to kind of get it.
[15:00] of widen these data distributions. We love using, like, internet-scale data to widen these distributions. So we have a lot of tools in our toolbox other than that real-world robot data. And... Talk a little bit more about the simulations of these other tools and the impact models like Astra are having on that. Yeah. That seems to be another cottage industry of excitement for the robotics world, right? How are you… using that. It's a sea change, right? I think we lived in this world for a while where everybody knew that all of this general knowledge about the world was baked into these models, but they couldn't really express it, right? They didn't have that last mile of spatial reasoning to connect the dots from the stuff that, like, you know, a GBT model knows to something that an action diffusion transformer uses. head can know. And that's changed, right? The labs are now post-training their models on 3D spatial data like CAD and... like three modeling, which by the way, they're doing because there's this direct economic incentive because there's a real market for it. So they can immediately monetize it. And it turns out that's the exact spatial reasoning that our models need. So I think we've been catapulted many years ahead in the future because we're starting to be able to use all of that general intelligence that lives in these big frontier models for robotics. And because it's a systems problem, the net effect for you is accelerated as opposed to slowing you down or saying somebody's going to come and eat your lunch. Yeah, no, we, and we've seen a couple of these happen, right? We started this business in 2021. We were doing this initially with convolutional neural networks, right? So like we've, we've gone through every sea change and, and the thing that is always valuable is the flow of data, right? And the flow of data is, is always the thing that uniquely you can do. And it's like, yeah, those like the million architectural tricks, like do them if it helps you accelerate the fly robots, collect more data, manufacture at a bigger scale, do more tasks like they're worth doing. Um, But ultimately, we think the thing that has kind of the longstanding generational value in a business is... is the flywheel and loops themselves. Can you talk a little bit about what you mean when you say you build robots? I think on one end of the continuum, it's like, you buy a unit tree. Yeah. And you prompt it. And the other end, whatever, you're machining parts. Where are you in this? Yeah. So... For context, for everybody, in 2025, the United States deployed 38,000 industrial robot worksups, which is tiny compared to roughly a million. You're not counting the Fanucs and the old ones. The Fanucs, right. But that's it, right? Nobody deployed any unitries in the United States into any. They would have shown up in those numbers, but they were probably zero. And so... The supply chain that exists for robots is both too small, And as you called out, wrong, right? Like the robots that Fanuc and ABB are building... They're mostly not building them in the United States, but even when they are building them in the United States, they're building these objects that are designed for systems integrators, that are designed for fast, repetitive motions. They are not cut from the set of tradeoffs that you would want to co-design with frontier models. And so we're kind of starting from scratch. on the hardware side. So Tudor, in order to deploy robots at scale, has actually been forced to design and manufacture our own hardware. For us today, like we're not building actuators, right? So we're not building cameras, there's a whole layer of stuff that we're sourcing from OEMs where they're manufacturing at a much larger scale. But the actual robot itself of, you know, the thing that drives around and has the arm on it and can do all the tasks, we actually are building at scale in Watertown, Massachusetts. Can you talk about the morphology of the robot? Is that called morphology? Anatomy? You can use whatever word you want. What do they look like? What do they look like? We make two. So one robot we make... Look at his website, Monjean. So Cassie. Cassie is... kind of, uh, a brute, right? Cassie is designed to do heavy duty work. It has a single arm and a lot of suction cups, and it's powered by foundation models in like the 200 to 400 million perimeter range and kind of came out of this world that Tudor was started out of in 2021 of like, what can foundation models do for robots? They can predict kind of affordances, grasps that you can go then execute. So it's this co-design of the hardware that you can manufacture at scale, a big market need, where in the United States, Anything in the physical economy has to be transported on a pallet or in a cardboard box. So there ends up being an enormous amount of physical human handwork of transporting pallets cases. Thank you. by hand. And so Cassie is kind of the result of that co-design, and we manufacture it at a very large scale, and we deploy it to a big diversity of businesses across manufacturing logistics. Our second robot is called Sunny.
[20:00] And Sunny is kind of an industrial semi-humanoid, right? It's as if you... You were looking at a unitree and you were like, I need this robot to run for three to 10 years in a factory at human rates and reliability. And so you end up with a slightly different set of design decisions. Sunny is wheeled, not legged, because most indoor industrial settings have polished ground floors. It has a much higher payload capacity and can move much faster and has industrial collaborative safety standards that enable it to operate around humans. collecting data for internal use and advancing our research agenda. And then we also now have a growing fleet of Sunnies out in the real world, doing more dexterous by manual manipulation. and you see How broad do you see this product line of different robot form factors? Is it Can you capture A big chunk of the useful market with three of them, or do you need 30 of them? Depends how you define the useful market. Where do you want to go? If you want to just live inside of manufacturing logistics, Sonny and Cassie today can probably do 80% to 90% of the work, right? These are designed for those cuts. Maybe just describe what that work, is because Yeah, so most people have this mental picture of what sort of the industrial base or manufacturing logistics in the United States, which comes from the TV show How It's Made. I don't know if either of you have seen this. Of course, of course. Yeah. And you see these like Rube Goldbergian machines of levers. And it's like, yeah, you don't see any humans in those pictures. Right. It's somebody built a machine that is the factory. Yeah. And that does exist. Yeah. And if it exists, they probably filmed it and put it on how it's made, right? Because that's the top 0.01% of American manufacturing logistics. And so what's below that? And you walk into these million square foot facilities, and you see a bunch of people. standing around on conveyor belts or folding tables, doing things by hand. And it's kind of shocking for anybody who doesn't come from this industry and community. And so you look at a robot like Cassie and it's like, you know... a huge amount of work, which is backbreaking work that's really hard to hire for. There's a three-month turnover where your average age of the worker doing that work is 55 years old and about to retire. And it's like, okay, we could use a Cassie to do that backbreaking work. And similarly, Sunny is like, basically anything you ever bought that came in a box, a human had to put that in a box, which is a crazy revelation for most people. Anything that requires any degree of assembly, right? Like even, you know, iPhones, you know, you would assume it's a robot, you know, putting that packaging together. No, it's a person even at the highest level of scale. Then to his question, which is so like you have like the single arm with the sucker, you've got the two armed humanoid. Yeah. What can't they do? Yeah. What can't they do? Where do you go from there? Yeah. So they can't go outside. And outside is really, really interesting. Oh, is that just like a locomotion issue? It's not locomotion. It's the hardware, right? These robots have... wheels and they have like a little bit of traction, you know, uh, they have like these articulating hind legs, but, uh, they can't drive over gravel, right? There's huge opportunities of robots in construction, robots reclaiming materials from landfills, um, uh, We need robots with legs for that. So I think that'll eventually be part of our pantheon of embodiments. And then similarly, robots in the home, you probably don't want a Cassie or Sunny in your home just now. It's a little bit too industrial of a device. Yeah, it looks scary. Yeah. Let's talk about the... Customer adoption side, how has that been obviously not – strictly from a, I don't have people, so I'll take whatever I can get. Yeah. to once they have it in their assembly line or whatever it is, how does adoption go in terms of their mindset? It's an interesting challenge because almost everybody we're selling a robot to has never purchased a robot before. And that kind of comes out of these stats, right? If the United States is only deploying 38,000 robots a year, if nine out of 10 US factories and warehouses don't have any robots, then almost all of the market is greenfield. And these people, you know, understand what this old guard of automation looks like, and they assume a robot doesn't work for them. And so the way we break through that actually isn't by talking about AI or models or tutors or anything like that. We call them up and we're saying, hey, have you ever considered a robot for 16 bucks an hour? And they haven't, right? That's a new proposition for them, right? And then you can kind of explain to them this thing that we're doing and get people interested and involved. And kind of by... Creating that business model where there is that true incentive alignment between us as the robot vendor and the customer, we're able to run distribution of robots at velocities that, you know, are many orders of magnitude faster than traditional industrial automation industry. We have an average time... from the, that first call to a robot signing contract of 40 to 60 days, we have an average time
[25:00] from that signed contract to robots actually in the customer site of about 40 days doing work. We stand up robots in production within a day, often within the first couple of hours that a robot is on site. So we're able to move very, very, very quickly because not only is the robot this generally intelligent, generally capable machine that we're not customizing, we're just, you know, shipping directly from the factory to the customer, but we're not putting any sort of economic risk on the customer. So we're able to get them on board. And once they deployed... How do their attitudes about labor versus machines change? the operations community is really, really fantastic because they see a thing work and it's like, it works. Like, that's it, right? It's like... I can see obviously it does the job. I know all the places where this job needs to be done. Let's go put it in all those places. So we've been able to scale up very, very quickly, both with small businesses and to some of the largest companies in the supply chain from one robot in one site to a dozen plus robots across many sites. It is interesting, though, because a lot of Gen AI was like... There's a social element where people get very attached to their models, and it's hard to swap models. Listen, you probably have more experience deploying robots than most people alive at this point. Are you saying the same thing? Are they viewed like equipment, or are they starting to be viewed like beings and coworkers? Yeah, so I saw a very funny tweet, which was, a robot is a robot unless it works, in which case it's a machine. I got to figure out the attribution. But is it real? It's pretty real. like they just view it as like a piece of equipment once it works certainly at like the factory manager level yes i think there's still the layer of like interestingly in robotics your buyer is not your user Your buyer is the COO or the general manager, and the user is a factory worker who's working with the robot all day long. And there you actually do need to cultivate the relationship between those workers and the robots because we don't fully automate 100% a factory. We automate some jobs in a factory or a warehouse. Do you have a philosophy on how you even approach this? Do you try to make them look more like co-workers or less like co-workers? Do you make them talk? How do you view this philosophically? of academic research called HRI, Human-Robot Interaction. A lot of people have spent a lot of time thinking about this, including a bunch of people from my academic background, and there are a whole bunch of tricks. So one trick is every robot has a name, and so that lets us actually build relationships with the robots, and also mostly our manufacturing technicians. So you name it, not the... It's not like a dog's when you get a new dog, you can rename it? when they do a repeat order, then they're allowed to name the robots. Oh, you're not supposed to rename a boat because that's bad luck. So you've got to keep the boat named. You can rename a dog. So I'm going to... The new robots get to be named. That solves it. And we get customers. They get a high sense of ownership now they've named it. Exactly. And we get really strong signals. We had people name the robots after their kids. People name the robots after their favorite football player, their favorite pop star. People get very attached because the robot ultimately helps. It makes their jobs easier. If you're a line worker in the facility, you're not doing backbreaking work as much. position and none of these places can hire. And then if you're a factory manager, your job is just a thousand times easier. What naming scheme do you use from when you... So we started off doing hurricane naming. So that's not... You didn't read the HRI book, clearly. Well, it's... I don't know. The customer doesn't know that Brittany is, you know, Nmod26 or whatever equals 2. It's... But we actually stopped because we kind of went through the hurricane. We went through the alphabet too many times to be able to keep track of it. Okay. So what are you using now? Uh, I think the technicians just pick whatever brings them joy. We have a really wonderful and privilege to employ a big staff of people in the United States who are helping to build these robots. And they are also naming them. Awesome. The other thing that surprised me when we first did the work before investing was, Most of the time when we talk to them, robotics companies, they're going after these humongous companies, right? Yeah. The big Fortune 50 type companies that seem to have – the budget for lots of experiments, etc. Yeah. Whereas, I think a significant majority of your customers And all these local businesses almost. Yeah. And... Is there a message in that or was it a conscious strategy or – Yeah, it's a lot of things. One is that's most of the market, right? The vast majority of U.S. businesses in manufacturing logistics are small businesses or mid-month.
[30:00] market businesses. So if you want to serve the market, that's what it takes to serve them. I think we also started this company at a time where we're seeing kind of all of these very academic supply chain oriented robotic startups kind of get started and then not really succeed. Um, And we had the opportunity of being like, okay, what went wrong here? And one of the big things I think that robotics companies get wrong is they're not actually really product companies. They're like, okay, some really big company has a very specific problem. They're willing to write us a shiny piece of paper that says they're going to expand with it. We're going to spend two years building a very custom system to solve their problem. We'll deploy it. Chances are it won't work because like anybody who's built products knows that you can't just one and done a product, right? Doing product requires iteration. And then even if it does work, who's to say that customer still has that problem? And so boom, the company's done. And I feel like I saw this happen like five to 10 times in a row to the point of like, oh my God, if this happened to us, like that would just be embarrassing. And so I think working with a diverse range of customers is this beautiful forcing function for building the right shape of product, which to us is a general robot, right? We don't want to build a Rube Goldberg machine that works really well for big industrial customer A or B. We want to build a platform that is generally capable, robot bodies running generally capable robot models, that can immediately get to work. And if you can sign a contract in two weeks with a mid-market customer and have a robot running immediately after that, and that's a totally different type of product and business than your other customers, you know you're being successful. How do you think can the limit what the industry looks like, assuming you continue to be successful? Is it just that... this work gets cheaper faster, or does somehow quality go up, or can they do more things? What is the economic conversions or the productive conversions? Yeah, I think everything gets better. So there's a whole set of goods that... If they are cheaper, we will buy more of them. And if you believe those goods are valuable, that's good for everybody. There are also a bunch of discontinuities in the world. I was talking with somebody the other day who worked on the line in a housing factory in, I think, Wisconsin. And they were describing how the houses that they're building are about 5% more expensive. to build them in a factory than to do traditional construction on site. But if these houses were even just 1% cheaper... than kind of the equivalent quality of house manufactured or constructed on site in the traditional way, all of the demand from developers would immediately go to that other shape of building these homes in a factory. And the same thing could be said about like reclaiming materials and landfills and a whole bunch of other markets. So I think our hope is like if Tudor is really successful, we can provide this robot substrate, this labor force that other businesses can use to transform the physical economy and make the world ultimately a cheaper and better place to live in. Will you ever go in the home? Um... Yeah, we're not opinionated. We'll go anywhere. But I would have mentioned that would be kind of a much tougher technical problem, just from like... I don't know, complexity of task and safety and... And the economics of it, right? Like how many, what is the economic value a tutor robot delivers to... in a factory and how many consumers are willing to to pay for that in their home. And it doesn't quite work out. But again, that comes back to our overall philosophy is, take trade-offs across as many layers of the stack as possible to get a foothold, start building the platform, and as you build the platform, you have to take fewer and fewer trade-offs over time. How does product innovation work? in I mean, most companies are just on V1, whereas you've got two versions of the robot and you continue to iterate really fast. How is it different than like, digital product innovation or other categories. We're trying to make it as similar as possible. So we're manufacturing our own robots. This gives us the ability to own that manufacturing process and figure out how do we push changes to the manufacturing process every two weeks. We are shipping code to robots either every day or major releases every two weeks. And it built a huge amount of infrastructure to support that level of iteration. So they're as close to the customer as humanly possible, which I think has been really successful for us. And then we also have a research team, which is able to be both, on the one hand, deeply embedded in kind of the customer problem space. And we're able to pull all sorts of levers to get specific types of data from customers, but also able to take these moonshots. And, you know, if we have a breakthrough on our modeling side that lets robots do a whole new shape of thing, we can kind of co-design
[35:00] a product around that in a way that I think most companies couldn't and get that out into the field much quicker. If I'm listening to this and I have a warehouse and I want to get a robot, do I just go to the website and put in my email? That's it. How does this work? You can go to our website, you can put an email. We actually have cool pricing calculators where you can actually put in some operational data about how your warehouse or factory runs today. And you can get an estimate of, hey, with an initial deployment of Tutor Robots, we could be saving $100,000 a year, $1 million a year, $100 million a year. And that's not like, okay, this is a roadmap over five years of custom engineering. It's like, no, you could save that a million dollars a year. two months from today. without putting CapEx down. And have you found that there is demand for, I mean, it's just such a sci-fi thing. Is there demand out there? There is so much demand. And it's not even demand for robots, because again, people don't, know that the shape of a robot exists or that it works or that it can do this job. So it's ultimately like, you know, we have this framework of the value of a robot is you take how much work a robot can do and you divide it by its cost. And that's it, right? Everything else is commentary. And so if we're able to compete in these highly liquid labor markets that exist across American physical economy, and we can offer something that delivers greater equivalent capability to lower cost, the market for that is the whole market, right? It's all physical word of dut. And just one quick wonky question on this, which is how much when you say cost, do you mean amortizing the upfront cost versus like the ongoing cost to operate it? It's both, which is very interesting and different, right? we've thought about robotics as ROI, right? Big HAPEX expense and then 10 years to make it back. And now it's actually like, no, more and more of the cost is going into ongoing inference teleoperation. But does that also include like maintenance and break-fix and all of that stuff? And then what is it like? What is the dominant cost or are they roughly equal? We drive down all of them over time. No, no, I'm just saying like from your standpoint, when you like pencil out your cogs, how much of is it the upfront cost versus the ongoing cost? Yeah, it is more ongoing at this point. And that's surprising people. Oh, wow. Yeah. Not surprising to me. I mean, you're really competing with something that you feed rice to, and it's very, very efficient, right? The human body. Yeah. So... When you get colored to these customers, is it because... Black Friday or something is coming up and they need a bunch of units done in a really short time? Or is this actually now part of their Actual labor planning cycle? No, it's strategic. Right? It's... These businesses, I think, are faced with a really hard choice, which is they know the age of their workforce and, They know that demand is ramping up due to all these secular trends and the cost basis of fulfilling that work is going up. And they're like, how does our business pencil out? Because, you know, these are not software businesses that have these, you know, huge gross margins that you can you can play with and can compress a little bit. And it's like, it's fine. It's like, no, the business either works or it doesn't. Um... And so I think our customers are ultimately looking for a long-term partner that can do the robots, right? If one of our customers is manufacturing deodorant, they're doing the deodorant, right? And they're going to be operational experts in deodorant or operational experts in crayons or whatever the thing they manufacture is, and they need a partner who's, you know, the expert on robots. I have a question on the nature of factories and buildings. I mean, you can ship a robot to anybody, anytime, not today because of your shortages, but at some point in the future, right? And if your robots become more and more general purpose over time, what does it mean for where factories are located and how people organize around the world? These factories and And you know these one-town factories and one-town factory town, whatever the expression is. Yeah. All of that fundamentally changes now, right? Yeah. Uh... I think a lot changes, right? There's a lot of geopolitics in this. I think one big thing is this is going to be a great equalizer between the United States and other countries like China invested in the previous regime of industrial automation. And in some ways, that made them very successful. And in other ways, I think in hindsight, that might end up being an enormous mistake. Because in the United States, we just discussed, we're doing this greenfield builds of automating our industrial capacity using generally capable robots. In China, they're going to have to tear down. Right. We're not going to have to do that in the United States in the same way. I think another thing that's going to shift, which is going to be really interesting, is we're just going to have way more factories in the same way that we have way more businesses because kind of the cloud and all of the software tooling let people build new businesses. Anybody will be able to start a business in the physical world. Anybody will be able to scale a business in the physical world if we kind
[40:00] have this infinite flexible labor substrate. You brought up China. We actually published a report on the supply chain and robotics, and it seemed to be that that, China was so central to it. So can you talk a little bit about your dependency on it and then how that evolves over time and how much is there risk in China to the supply chain? Yeah. I think ultimately, no matter what anybody says, today in 2026, we have a highly interconnected, interdependent global supply chain. And this is one of many examples of this. And I think both U.S. companies and Chinese companies are thinking, well, very hard about kind of what is the long-term strategy of this look like? And it's Chinese companies investing in the United States, it's United States companies investing in China, and ultimately everybody trying to build their industrial capacities. Yeah. I've noticed that you've gotten, as far as I can tell, you've verticalized more. Is that right? Like you're taking more in-house? And do you plan to continue to do so? Yeah. Well, there's this tough thing in supply chain where, and again, because robotics is such a cottage industry, is you might start in a world where subcomponents are coming from OEMs, and then you grow your volume to the point of they're not OEMs anymore. They're glorified contract manufacturers, right? Once you're doing a material percentage of their volume. dynamics change, right? It's like, is that the right command for that thing? Do I want to be using a command or do I want to be manufacturing it in-house? I mean, Tudor must be one of the largest deployers of robots in the United States. As a result, he must be one of the largest customers for some of these OVMs. As a result, I would imagine over time, you just continue to drive these components yourself. I think that's right. Does that mean taking it in-house or just telling them what to do? I mean, is it just, are they still suppliers? Some combination. Some combination. I think ultimately, for Tudor to be successful, we will need a mature robotic supply chain in the United States. And we're in this position where, Because we are so much of the demand for that supply chain, I think it's both our obligation and our opportunity to invest very, very, very heavily to building this capacity in the United States. And that's why we build robots in Watertown, Massachusetts. It's why we expect to continue to verticalize and build more and more of our robot componentry in the United States. Again, going back to when we first started talking, I remember it was probably the second meeting when we actually started about the research, right? Most of the meeting because we have such a great business. But that's – a big strength of yours, right? And Talk about how you approach it, and there's all sorts of theories about which robotic learning models will win and so on and so forth. How are you approaching the learning bet? Yeah. in research? I think we're thinking about how this played out in LLMs and what's similar and what's different. And I think one thing that clearly has been very, very important in LLMs as there are lots of heterogeneous sources of data, but ultimately there's a top of the data period, which is real-world reinforcement learning traces doing the thing you want to be able to do. And so we've built our research program in such a way where, We are thinking about how do we get the most out of robot experience out in the real world? And so we built this 100-robot fleet called Data Factory One in Boston. This is, to our knowledge, the largest sort of semi-humanoid data collection farm in the United States. And we're able to use this as kind of this research test bed of how do we do real-world reinforcement learning at scale with robots? You know, reinforcement learning is a technique family where... It doesn't really work at small batch sizes. And so if you have a very large fleet out in the real world, you can start to play with these methods, which I think people haven't been able to play with either in academia or industry. So I presume like building a team for this is quite different, right? Than like a lab. So maybe talk about kind of like the construction of the team and maybe if there's anybody you're looking to... To hire? We're looking to hire everybody. Everybody. So we're looking to hire researchers and engineers, and that's across every single layer of the stack. Wow. We are also looking to hire technicians to build robots. By the way, every layer of the stack, you mean like software? Hardware, materials, electronics. Like, yeah, there's nowhere that you... can abdicate responsibility, right? It's robotic. You're going to engage with it whether you want to or not. The question is whether you understand it. Um, So we're building that team, and I think that's a really unique shape of team, not only because it's so interdisciplinary and collaborative across all those layers of the technical stack, but because I think it's one of the only places in the universe where you can be on a whiteboard talking about model architecture in the morning, in the afternoon, like... looking at a factory line and seeing how a robot is doing. - Is this primarily in Massachusetts? - So we have a team across the United States. We have especially folks on the go-to-market and technicians who are kind of engaging. Our customers are everywhere across the United States. Awesome, thank you.
[45:00] This episode of the A16Z podcast. If you liked this episode, be sure to like, comment, subscribe, leave us a rating or review and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts and Spotify. Follow us on X and A16Z and subscribe to our substack at a16z.substack.com. Thanks again for listening and I'll see you in the next episode. As a reminder, the content here is for informational purposes only. It should not be Thank you. you
