Pickle Robot Targets $75B Supply Chain Bottleneck With Physical AI
Ai4
Fine-tuning pre-trained foundation models on real-world operational data turns autonomous truck unloading into physical AI's first scalable commercial beachhead.
## Background
Published by Ai4 on YouTube on August 28, 2025, this presentation features AJ Meyer, Founder and CEO of Pickle Robot Company. The session explores how physical and generative AI are transforming physical supply chain operations by turning perception into motion to autonomously unload trailers and shipping containers.
## TL;DR
Pickle Robot founder AJ Meyer highlights the unaddressed $75 billion market opportunity in automated truck loading and unloading within global supply chains. He traces the evolution of physical AI from 1950s Bellman formulations and historic MIT and Stanford milestones to modern advances like diffusion policies and hallucinated world models. Meyer details how Pickle Robot deploys an end-to-end vision-to-motion architecture using seven RGB cameras and fine-tuned foundation models to master high-frequency micro-motions in grueling cross-dock environments. Ultimately, he demonstrates how physical AI has transitioned from academic research into live, high-throughput commercial rollouts for major global enterprise accounts.
## Atmosphere & style
The presentation carries a sharp, authoritative tone that bridges deep MIT robotics lineage with the harsh, 130°F industrial reality of California shipping docks. Meyer speaks with the brisk, pragmatic enthusiasm of an experienced technical founder, effortlessly translating complex reinforcement learning paradigms into intuitive physical analogies. The narrative arc builds an engaging bridge from fragile historical prototypes to modern generative control systems, leaving the audience with an exciting, tangible sense of robotics finally crossing into real-world commercial viability.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=4s) Truck loading and unloading alone represents an unaddressed $75 billion annual market, with the potential to double the cumulative commercial footprint of the entire existing robotics and automation industry.
- [04:38](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=278s) Cross-dock logistics has historically resisted automation due to extreme unstructured environments, where human workers manually unpack shipping containers at temperatures reaching 130°F.
- [04:38](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=278s) Autonomous supply chain automation requires coordinating specialized interoperable robots across yard, dock, sorting, and transport operations rather than building a single general-purpose machine.
- [04:38](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=278s) Production robot fleets produce massive data scale—over 100 gigabytes per robot daily—driving a commercial flywheel where physical operational data continually improves learned models.
- [10:08](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=608s) Unlike traditional industrial automation that follows pre-programmed rigid trajectories ('trains on rails'), Physical AI implements a closed Bellman loop connecting perception, goal specification, world modeling, and action.
- [14:20](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=860s) Learning from physical reality marked a foundational shift from abstract game RL, demonstrated by early prototypes learning dynamic locomotion directly from real-world physics rather than hand-coded kinematics.
- [14:20](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=860s) Modern physical AI is powered by four complementary machine learning pillars: generative world models, discriminative perception models, reinforcement learning, and statistical inference.
- [16:22](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=982s) Imitation learning from human demonstrations circumvents the need to write complex analytical physics equations by allowing systems to directly infer underlying goals and control policies.
- [17:24](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1044s) Diffusion policy adapts generative denoising techniques from image and video synthesis to iteratively generate smooth, robust robot action trajectories directly from multimodal sensor inputs.
- [17:24](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1044s) Generative world models allow robots to synthesize internal simulations and 'hallucinate' future physical interactions before executing actions in unstructured environments.
- [18:14](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1094s) Pickle Robot's end-to-end vision-to-motion stack uses standard RGB cameras and neural networks to perform high-frequency micro-adjustments that prevent packages from slipping during high-speed unloading.
- [19:52](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1192s) The robotics engineering playbook has shifted from training vision and control models from scratch to fine-tuning pre-trained foundation models on customer-specific physical domain data.
- [19:52](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1192s) Physical AI has crossed into broad commercial adoption, with autonomous trailer unloading serving as the beachhead killer application scaling into major enterprise logistics pipelines.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=4s) The Rise of Physical AI and the $75B Dock Door Opportunity
AJ Meyer introduces himself as the founder of Pickle Robot Company, a 130-person robotics startup based in Boston and rooted in the MIT robotics ecosystem. He notes that the entire robotics field is currently having a breakthrough moment and sitting on the cusp of a major revolution. This surge is directly underwritten by rapid advancements in artificial intelligence over recent years, giving rise to the buzzword often championed by NVIDIA's Jensen Huang: "Physical AI." Meyer highlights the two questions he encounters most frequently: what physical AI actually means, and why heavyweights like Boston Dynamics, SoftBank, Honeywell, Pickle Robot, and a wave of startups are pouring hundreds of millions of dollars into the seemingly unglamorous problem of loading and unloading trucks.
The economic rationale centers on the staggering scale of the global supply chain, which incurs roughly $50 trillion in labor costs. In contrast, the entire historical commercial impact of all robotics and automation deployed to date totals about a $75 billion-per-year market—a mere drop in the bucket against the $50 trillion total. Truck loading and unloading alone constitutes an unaddressed $75 billion problem. Successfully solving this single operational chokepoint has the potential to double the cumulative impact of the entire robotics and automation industry combined.
```mermaid
flowchart LR
accTitle: Robotics Market Comparison
subgraph SCM["Global Supply Chain"]
L["Total Labor Bill<br/>$50 Trillion"]
end
subgraph Impact["Automation Value Pools"]
direction TB
A["Existing Automation Market<br/>$75 Billion/year"]
T["Truck Loading & Unloading<br/>$75 Billion/year (Pickle Target)"]
end
SCM -.-> Impact
```
- **Pickle Robot Company**: ~130-person MIT spinout startup based in Boston.
- **Global Supply Chain Labor Bill**: ~$50 trillion.
- **Total Existing Robotics/Automation Market**: ~$75 billion per year.
- **Truck Loading and Unloading Market**: ~$75 billion per year.
---
[04:38](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=278s) Orchestrating the Autonomous Supply Chain Network
Examining a real-world cross-dock facility in Los Angeles illustrates the brutal physical realities and lack of automation in modern logistics. In these facilities, international cargo arriving by sea is driven from ports to cross-docks featuring upwards of 150 dock doors on one side. Human crews work inside shipping containers and trailers where temperatures frequently hit 130°F during California summers, manually unloading boxes onto carts, sorting orders, wheeling them across the facility, and reloading outbound trailers headed for retail shelves at stores like Target. Existing industrial automation has historically offered nothing to address this unstructured, physically demanding environment.
Meyer argues against viewing truck unloading in isolation. True value creation requires connecting individual specialized robots into an integrated network that coordinates seamlessly across warehouse workflows, metropolitan fulfillment clusters, and national supply chains. Rather than building a single robot to do everything, the winning approach relies on combining commodity hardware with proprietary innovations, tackling extreme data challenges, and deploying commercial flywheels.
subgraph Facility["Intralogistics & Sorting"]
B --> C["AmbiSort Palletizer / Sorting"]
C --> D["Teradyne / Universal Robots AMR Jack"]
D --> E["Reflex Robotics Humanoid Sorter"]
end
subgraph Outbound["Outbound Logistics"]
E --> F["Pickle Gen 2 Loading Robot"]
F --> G["Outrider Yard Truck / Over-the-Road Semi"]
end
```
- **Hardware Ecosystem Convergence**: Exploits low-cost, highly reliable off-the-shelf sensors, compute, and motors, augmented by proprietary hardware such as Pickle's specialized robotic hand engineered to lift overhead 60 lb boxes.
- **Massive Data Scale**: Individual robots generate upwards of 100 gigabytes of operational data per day, creating data pipeline scales well beyond tens of petabytes across thousands of units.
- **Applied Foundation Models & Flywheel**: Frontier general-purpose models require fine-tuning on domain-specific customer experience; Pickle's flywheel uses initial reliable deployments to gather physical operational data that continuously improves performance.
- **Collaborative Partner Ecosystem**: Interoperates with Outrider (autonomous yard trucks), Ambi Robotics / AmbiSort (pallet sorting), Teradyne / Universal Robots (autonomous pallet jacks/forklifts), and Reflex Robotics (dexterous humanoid package sorters).
---
[10:08](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=608s) From Bellman's Equation to Modern Physical AI
To define "Physical AI," Meyer references the classic Bellman equation from 1950s reinforcement learning, deconstructing its core mathematical components into four essential functions of an autonomous agent:
Meyer contrasts this with conventional industrial automation—such as automotive welding arms or assembly line croissant straighteners—which he compares to a "train on rails." Traditional robots operate strictly along pre-programmed rigid trajectories, using vision sensors only to start, stop, or lightly adjust speeds, entirely lacking internal goals, physical world modeling, or autonomous planning. True physical AI began with pioneering research prototypes that integrated these four closed-loop pillars:
- **Shakey the Robot (SRI, 1966)**: The earliest physical AI system, equipped with onboard machine vision, camera sensors, mapping capabilities, and automated planning to independently navigate Stanford Research Institute’s AI lab.
- **Handey (MIT, 1986)**: Developed as the PhD thesis of MIT professor Tomás Lozano-Pérez, demonstrating model-based autonomous pick-and-place manipulation with goal formulation and motion planning.
- **MIT Leg Lab (1990s)**: Founded by Marc Raibert (who later founded Boston Dynamics) alongside Tomás Lozano-Pérez; early research lab where Pickle co-founder Dan Paluska built dynamic legged humanoids and backflipping machines decades before modern viral robotics videos.
---
[14:20](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=860s) The Evolution of Learning from Physical Data
Between 1954 and 2001, early physical AI research consumed millions of dollars in DARPA funding but yielded virtually zero commercially viable robots due to brittle heuristics and prohibitive hardware costs. The critical paradigm shift occurred around 2004 with Russ Tedrake’s PhD thesis at MIT on the "Toddler" robot, which marked the first time a physical robot learned to walk through reinforcement learning directly from real-world physical experience rather than pre-programmed kinematics or rigid simulation.
Learning from real-world physics differs fundamentally from reinforcement learning in abstract game environments like chess or backgammon. In modern physical AI, four complementary machine learning approaches bridge the gap between sensor data and the Bellman decision loop:
| Learning Paradigm | Primary Mechanism / Function in Physical AI |
|:--- |:--- |
| **Generative Models** | Simulating, predicting, and "hallucinating" plausible world states and trajectory futures. |
| **Discriminative Models** | Semantic classification and object detection (e.g., distinguishing package types and boundaries). |
| **Reinforcement Learning** | Optimizing action policies and motor primitives through trial-and-error reward signals. |
| **Inference & Regression** | Estimating continuous physical parameters, aerodynamics, and dynamic contact states. |
Meyer illustrates learning complex dynamics via the MIT "Perching Glider" project. When small foam glider aircraft are forced into high-angle aerodynamic stall maneuvers to land precisely on a wire, classical fluid dynamics equations become mathematically untenable. By applying statistical inference directly to physical flight data, the control system learns the nonlinear aerodynamic transition dynamics required for robust autonomous perching.
[16:22](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=982s) Historical Milestones: ASIMO Synthetic Vision and Stanford's Acrobatic Helicopter
Reflecting on the foundations of modern robotics, Meyer recounts his early work at Honda Research on the humanoid robot ASIMO. In 2008, his team was among the first to pioneer GPU-accelerated synthetic data pipelines for robotic vision. They constructed a discriminative model designed to evaluate whether ASIMO would successfully recognize a given synthetic image; whenever the discriminator predicted a failure, additional synthetic variations were generated and funneled back into the training pipeline to harden the vision system.
Around the same era, Pieter Abbeel completed his PhD thesis at Stanford, demonstrating human-level autonomous acrobatic helicopter flight—a benchmark Meyer still urges audiences to look up: "If you have to YouTube anything in the talk, YouTube this. It's unbelievable." Rather than manually programming intricate aerodynamic flight dynamics, Abbeel leveraged what is now recognized as imitation learning (or learning from demonstration). By training an inference model directly on telemetry captured from expert human radio-control pilots, the system learned to infer the underlying goal and control policies directly, overcoming the persistent challenge in physical AI where explicit mathematical descriptions of complex physical tasks are nearly impossible to write by hand.
- **Honda ASIMO Vision Pipeline (2008):** Early implementation of GPU-accelerated synthetic data generation paired with automated discriminators to identify perception edge cases.
- **Stanford Autonomous Helicopter:** Leveraged imitation learning and inverse reinforcement learning from human demonstrations to execute acrobatic maneuvers.
- **Goal Specification in Physical AI:** Highlighted how learning from human demonstration bypasses the friction of manually coding physical goals and dynamics.
---
[17:24](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1044s) Modern Physical AI Breakthroughs: Diffusion Policy and Hallucinated World Models
Transitioning from 2010 to recent developments, Meyer highlights how generative architectures have transitioned from digital media synthesis into robotics. A pivotal breakthrough emerged through diffusion policy, which adapts the noise-to-data generation process popularized by image and video models like Midjourney and Stable Diffusion. Instead of generating pixels, the model begins with noise over potential robot action trajectories, iteratively denoising the signal against real physical demonstration data to yield smooth, robust robot control sequences.
Complementing diffusion policies are synthesized world models, such as the generative simulation research from Yilun Du at Harvard. In these systems, the robot actively "hallucinates" future visual and physical states, synthesizing full hypothetical rollouts to imagine how an interaction will play out. By evaluating these internal simulations before executing an action in the real world, the robot can determine the optimal sequence for manipulating unstructured physical objects, such as opening a drawer and stowing items inside.
```mermaid
flowchart LR
accTitle: Modern Physical AI Generative Mechanisms
subgraph GenerativeControl["Diffusion Policy"]
A["Action Noise Space"] --> B["Iterative Denoising"]
B --> C["Physical Actuation"]
end
subgraph WorldModeling["Synthesized Imagination"]
D["Sensor Observations"] --> E["Hallucinated World Model"]
E --> F["Predicted Future States"]
F --> C
end
```
- **Diffusion Policy for Robotics:** Converts the iterative denoising formulation of generative image/video models into robot trajectory generation conditioned on multimodal physical inputs.
- **Hallucinated World Models:** Synthesizes internal forward simulations, allowing robots to visually "imagine" manipulation outcomes prior to physical contact.
---
[18:14](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1094s) End-to-End Robotics and Pickle Robot's Vision-to-Motion Stack
By 2025, the convergence of four core machine learning paradigms—generative models, discriminative models, reinforcement learning, and statistical inference—has replaced the traditional, brittle paradigm of hand-crafted perception and motion planning pipelines. When these four learning systems are unified from raw sensory perception directly to motor actuation, it forms what the industry terms "end-to-end robotics."
```mermaid
flowchart TD
accTitle: Pickle Robot End-to-End Physical AI Architecture
subgraph Perception["Perception Stack"]
A["7 RGB Cameras"] --> B["3D Scene Reconstruction & Spatial Parsing"]
end
subgraph LearningStack["Physical AI Models"]
B --> C["Generative & Hallucination Models"]
B --> D["Discriminative & Inference Models"]
C & D --> E["Diffusion Policy & Reinforcement Learning"]
end
subgraph Control["Production Execution"]
E --> F["Subtle Motion Optimization"]
F --> G["High-Speed Trailer Unloading"]
end
```
At Pickle Robot, this architecture runs in live production environments. Equipped with an array of seven standard RGB cameras, the robot reconstructs real-time 3D scenes, identifies floor planes, reads package labels, and synthesizes motion paths. Beyond macroscopic path planning, neural models learn subtle, high-frequency kinetic adjustments during package transfers. Minor physical variations that are almost imperceptible to the human eye dictate whether an irregularly packed carton slips or holds; by learning these nuances directly from data, the system significantly reduces dropped packages, accelerates cycle times, and boosts overall operational throughput.
- **Pickle Sensor Suite:** Operates on seven standard RGB cameras to build real-time 3D spatial reconstructions and parse semantic scene details.
- **Micro-Motion Optimization:** Utilizes learned models to correct subtle trajectory variations during high-speed package handling, preventing grip failures.
- **Development Investment:** Reaching production-grade reliability with this learned physical AI stack required "tens of millions of dollars" in targeted R&D.
---
[19:52](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1192s) Crossing the Adoption Curve: Fine-Tuning Foundation Models for Commercial Scale
Meyer concludes by noting that physical AI has transitioned past the early adopter phase into broad commercial deployment—climbing the steep part of the adoption curve. Autonomous truck and trailer unloading serves as the initial "killer application" for physical AI in supply chain operations, paving the way for a broader ecosystem of industrial automation startups.
Addressing the practical engineering playbook, Meyer updates the company's "secret sauce": teams no longer need to train foundational vision and robotics models from scratch. Instead, the modern standard is to take pre-trained foundation models and fine-tune them directly on customer-specific physical domain data. Operating on this paradigm in late 2025, Pickle Robot has deployed across its first 10 enterprise customer accounts out of hundreds waiting in the commercial pipeline, spanning logistics giants like Amazon to global retail apparel brands such as Zara, Puma, and Helly Hansen.
- **Market Inflection:** Robotic container and trailer unloading is operating as the beachhead commercial application driving physical AI adoption across global logistics.
- **Updated Deployment Playbook:** Fine-tuning pre-trained foundation models on target customer edge cases replaces bespoke model training from scratch.
- **Enterprise Customer Pipeline:** Live deployments across early enterprise adopters, scaling into an active sales pipeline comprising hundreds of distribution hubs.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=4s) Physical AI — AJ Meyer
> [10:08](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=608s) train on rails — AJ Meyer
> [14:20](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=860s) Toddler — AJ Meyer
> [14:20](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=860s) Perching Glider — AJ Meyer
> [16:22](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=982s) If you have to YouTube anything in the talk, YouTube this. It's unbelievable. — AJ Meyer
> [17:24](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1044s) Diffusion policy — AJ Meyer
> [18:14](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1094s) end-to-end robotics — AJ Meyer
> [18:14](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1094s) tens of millions of dollars — AJ Meyer
> [19:52](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1192s) killer application — AJ Meyer
> [19:52](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1192s) secret sauce — AJ Meyer
## Mentioned
AJ Meyer, Pickle Robot Company, MIT, Jensen Huang, NVIDIA, Boston Dynamics, SoftBank, Honeywell, Target, Outrider, Ambi Robotics, AmbiSort, Teradyne, Universal Robots, Reflex Robotics, Richard Bellman, Shakey the Robot, Stanford Research Institute (SRI), Handey, Tomás Lozano-Pérez, Marc Raibert, MIT Leg Lab, Dan Paluska, DARPA, Russ Tedrake, Honda Research, ASIMO, Pieter Abbeel, Stanford University, Midjourney, Stable Diffusion, Yilun Du, Harvard University, Amazon, Zara, Puma, Helly Hansen
## Video description
AJ Meyer, Founder & CEO, Pickle Robot Company
Abstracts
While foundation models are revolutionizing digital workflows, AJ explores how generative AI is now being applied to physical supply chain operations - turning perception into movement and predictions into action to drive autonomous robots that unload trailers and
import containers. See how robots with physical AI move goods, not just data, to deliver real ROI for real companies.
AJ Meyer, founder of Pickle Robot Company, argues that physical artificial intelligence has reached a commercial inflection point, identifying trailer loading and unloading as a $75 billion market that could double the historical impact of the robotics industry.
The $75 Billion Logistics Automation Chokepoint (0:04)
Meyer contrasts the global supply chain's estimated $50 trillion labor expense with the existing commercial robotics market, which generates approximately $75 billion annually. According to Meyer, truck loading and unloading represents an unaddressed $75 billion annual labor pool. Automating this single operational stage could effectively match the cumulative financial footprint of all existing industrial robotics deployments.
The target environment reflects severe physical and ergonomic challenges. In high-volume distribution hubs, such as Los Angeles cross-dock facilities with more than 150 dock doors, human workers manually unload loose cartons from shipping containers where summer temperatures regularly reach 130°F. Conventional industrial automation has historically failed to address these unstructured, sweltering, and physically hazardous environments.
Evolution From Scripted Kinematics to Generative Control (10:08)
Traditional factory robots operate like trains on rails, executing pre-programmed kinematic trajectories with minimal environmental responsiveness. Meyer frames physical AI around Bellman's classical decision loop—coupling perception, goal specification, world transition modeling, and action execution. While early research prototypes such as SRI's Shakey (1966) and MIT's Handey (1986) formulated these concepts, brittle heuristics prevented commercialization for decades.
The modern robotics control stack has transitioned from hand-coded heuristics to data-driven paradigms. Foundational stepping stones—such as Russ Tedrake's 2004 reinforcement learning walking robot, Pieter Abbeel's imitation learning with acrobatic helicopters, and Honda's 2008 synthetic vision pipelines for ASIMO—laid the groundwork for modern generative techniques. Today, diffusion policies adapt generative image denoising to construct smooth robot motor trajectories, while synthesized world models project hypothetical physical interactions to plan complex manipulations before physical contact.
End-to-End Vision-to-Motion in Production (18:14)
Pickle Robot replaces disjointed perception and motion-planning pipelines with an end-to-end physical AI architecture. Running in active customer production, the company's system utilizes a sensor array of seven standard RGB cameras to reconstruct 3D scenes, identify boundaries, parse package labels, and synthesize motor actuation commands.
Meyer states that tens of millions of dollars in research and development were required to achieve production reliability. In addition to broad path planning, the learned models execute high-frequency micro-adjustments during high-speed package handling. Learning these subtle physical nuances directly from operational data prevents package slippage and dropped cartons, directly improving trailer unloading cycle times and operational throughput.
Fleet Interoperability and Foundation Model Fine-Tuning (4:38)
Achieving end-to-end automation across cross-dock facilities requires coordinating specialized machines rather than deploying a single general-purpose robot. Meyer illustrates an interconnected workflow incorporating Outrider autonomous yard trucks, Pickle unloading and loading arms, AmbiSort parcel sorters, Teradyne autonomous mobile pallet jacks, and Reflex Robotics humanoid sorters. Deploying fleets at scale introduces significant data infrastructure demands, as individual robots generate more than 100 gigabytes of operational data daily, accumulating into multi-petabyte datasets across broad networks.
Meyer outlines a critical shift in the deployment playbook: robotics companies no longer need to train core vision and robotic architectures from scratch. Instead, engineering teams fine-tune pre-trained foundation models using domain-specific physical interaction data. Under this strategy, Pickle has deployed systems across its first 10 enterprise customer accounts, including retail and logistics brands such as Amazon, Zara, Puma, and Helly Hansen.
Meyer presents trailer unloading as the initial beachhead proving the commercial viability of physical AI across supply chains. However, scaling beyond initial enterprise deployments requires sustaining high-volume physical data flywheels and ensuring seamless interoperability across heterogeneous autonomous machinery.
## Background
本视频由 Ai4 频道于 2025 年 8 月 28 日发布在 YouTube 上，演讲嘉宾为 Pickle Robot Company 创始人兼 CEO AJ Meyer。内容主要探讨生成式 AI 与“实体 AI（Physical AI）”如何落地于物理供应链场景，通过将环境感知与预测转化为自主机器人的实际行动，实现卡车与集装箱的自动化卸货并为企业带来切实的商业投资回报。
## TL;DR
AJ Meyer 深入解析了实体 AI 的技术本质及其在物流供应链中的突破性应用，指出单是攻克卡车装卸货这单一痛点即可撬动 750 亿美元的潜在市场。演讲系统回顾了实体 AI 从贝尔曼方程、MIT 早期物理实验到当代扩散策略与世界模型（生成式预演）的演进历程，并详细介绍了 Pickle Robot 运用 7 组多模态相机矩阵与亚毫米级微动作学习实现的端到端控制系统。目前该细分赛道已跨越早期采纳鸿沟，Pickle 依靠“基座大模型 + 垂直场景数据微调”的策略，已在亚马逊、Zara、Puma 等顶级客户的实际仓储场景中实现首批规模化交付。
## Atmosphere & style
整场演讲充满硬核科技与产业落地交融的振奋气息，展现出一种严谨务实又极具开拓精神的叙事风格。AJ 将 MIT 实验室数十年的学术积淀与炎热货箱内的严酷工况紧密相连，在回顾特技直升机与 ASIMO 早期探索时充满敬畏与热情，而在展示当下实机作业与商业订单突破时则格外自信沉稳，让观众强烈共鸣到物理 AI 跨越技术奇点、迈向黄金时代的行业脉搏。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=4s) 实体 AI（Physical AI）正迎来爆发式临界点，将感知、预测与运动控制深度结合，推动机器人进入物理供应链的核心场景。
- [03:34](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=214s) 全球供应链的人力成本规模高达 50 万亿美元，而传统机器人自动化产业年产值仅 750 亿美元，单是卡车装卸货这一单一细分市场的潜在规模就达 750 亿美元。
- [03:34](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=214s) 实体 AI 的真正商业价值在于多设备协同构建的网络效应，打通从卸货、分拣、码垛、搬运到干线自动驾驶卡车的全流程闭环。
- [10:08](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=608s) 实体 AI 系统的核心是贝尔曼方程（Bellman Equation）四大要素构成的闭环：感知、目标、状态转移推演及行动，区别于传统沿固定轨道运行的工业自动化。
- [14:19](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=859s) 过去半个世纪中基于手写规则的自主机器人项目在商业上均以失败告终，直至 2004 年从真实物理经验数据（Experience Data）中学习的强化学习方法才带来历史性转折。
- [17:20](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1040s) 现代物理 AI 建立在生成模型、判别模型、强化学习与推理模型四大支柱上，扩散策略（Diffusion Policy）与世界模型彻底取代了手写复杂规则的旧范式。
- [18:50](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1130s) Pickle Robot 利用 7 组 RGB 相机实现实时三维重建与轨迹预演，并通过机器学习捕捉亚毫米级的动态偏差，确保高速作业下的零跌落。
- [19:50](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1190s) 现代硬科技创业的最佳方法论是以成熟通用 Foundation Models 为基座，集中资源利用客户现场真实业务数据进行针对性微调（Fine-tune）。
- [19:50](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1190s) 卡车自主卸货已跨越早期采纳鸿沟，成为物理 AI 的首个杀手级应用，Pickle Robot 已交付包括亚马逊、Zara、Puma 等在内的标杆客户。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=4s) 机器人时代的临界点与“实体 AI”的崛起
在波士顿这片孕育了深厚学术与工程底蕴的 MIT 创新生态圈中，Pickle Robot 作为一家约 130 人的机器人初创企业，正站在行业剧变的最前沿。正如创始人兼 CEO AJ Meyer 在 AI 峰会开场时所言，整个机器人领域当前正处于某种“迎来高光时刻的临界点（having a bit of a moment）”，一场巨大的行业变革呼之欲出。而这一切的底层支撑，正是过去数年间人工智能技术的突飞猛进。如今，以英伟达黄仁勋等行业领袖频繁提及的新词——"Physical AI"（实体 AI / 物理 AI）为代表，技术演进正在重塑大众与投资人的认知。
面对行业内最常被问及的两个核心疑问——“究竟什么是实体 AI？”以及“为什么像 Pickle Robot、软银（SoftBank）、霍尼韦尔（Honeywell）、波士顿动力（Boston Dynamics）乃至众多明星初创公司，要向看似平平无奇的‘卡车装卸货’场景砸入数亿美元？”AJ 在现场展示了一段从未公开发布过的 90 秒实机作业独家视频，直观呈现了具身智能机器人在真实仓储环境中的自动化搬运表现。
```mermaid
flowchart LR
accTitle: 实体 AI 供应链协同全景架构
subgraph S1["月台与堆场调度"]
A["自主牵引车<br/>Outrider"] --> B["卸货机器人<br/>Pickle Gen 2"]
end
subgraph S2["库内分拣与转运"]
B --> C["精细分拣机器人<br/>Reflex"]
C --> D["码垛机器人<br/>AmbiSort"]
D --> E["自主叉车/托盘车<br/>Teradyne"]
end
subgraph S3["出库装载与干线"]
E --> F["装车机器人<br/>Pickle"]
F --> G["干线自动驾驶卡车"]
end
```
### [03:34](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=214s) 750 亿美元卡车装卸难题与供应链网络效应
从商业逻辑来看，机器人与自动化的核心使命是撬动全球供应链中高达 50 万亿美元的人力成本账单。然而，整个传统机器人与自动化行业自诞生以来的商业化总规模每年仅约 750 亿美元，在 50 万亿美元的盘子面前不过是“沧海一粟（a drop in the bucket）”。极为戏剧性的是，仅“卡车装卸货（loading and unloading trucks）”这单一场景的潜在市场规模就达到了 750 亿美元。这意味着，只要攻克这一个应用环节，就能让过去几十年所有机器人与自动化创造的市场价值直接翻倍。
以洛杉矶一处典型的越库配送中心（Cross-dock）为例，远洋货轮靠港后，卡车将集装箱运至仓库一侧排开的 150 多个月台门之一。在加州炎炎夏日高达 130°F（约 54.4°C）的密闭货箱内，工人们需要顶着酷暑人工将单件重达数十磅的货物搬卸到推车上，推至仓库对侧重新装车，最终送往 Target 等零售卖场的货架。这种恶劣且高度依赖人工的场景长期缺乏技术渗透，不仅因为环境复杂多变，更因为传统的自动化方案根本无法应对非结构化的货箱。
AJ 强调，孤立地看待单台机器人是狭隘的（myopic），真正的价值爆发点在于设备之间相互通信、协同作业所构建的“网络效应（network effect）”。通过将卸货、库内分拣、搬运、重新装车以及未来 4 到 5 年内落地的干线自动驾驶卡车串联成一张协作网络，实体 AI 将实现从单座仓库、都市物流圈到全球供应链的整体效率跃迁。
- 全球供应链人力成本规模约 50 万亿美元（$50T）。
- 传统机器人与自动化整体年产值仅 750 亿美元（$75B）。
- 单一卡车装卸货市场的潜在商业价值达 750 亿美元（$75B）。
- 典型越库仓库夏季集装箱内作业温度可达 130°F（约 54.4°C）。
### [06:32](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=392s) 软硬件创新、海量物理数据与商业化飞轮
构建能够胜任极端物流场景的实体 AI 机器人，需要一套融合了尖端硬件、分布式软件架构与垂直场景数据的“独门配方（secret sauce）”：
1. **成熟廉价的硬件基础与专用创新**：2025 年的硬件供应链具备了前所未有的高算力、高可靠性传感器与高扭矩电机，结合自主创新的专用末端执行器（例如 Pickle 自研的吸附机械手，能够轻松在头顶高度抓取 60 磅重的重型纸箱），构成了稳定运行的物理躯体。
2. **PB 级物理世界数据规模**：在软件与数据层面，每台作业机器人每天可产生约 100GB 的多模态传感器数据，当机队规模扩展至成千上万台时，数据处理吞吐量将迅速超越 80PB 级别，带来全新的工程挑战。
3. **垂直场景数据飞轮**：学术界前沿的通用基座模型（Foundation Models）虽具潜力，但距离工业落地仍有很长的路要走，必须针对具体业务场景微调。Pickle 打造的商业闭环是：先交付能满足当前工况的机器人系统，让其在真实经验中收集物理交互数据，通过持续训练自我迭代，进而以飞轮效应不断拓展服务边界。
在合作伙伴生态中，Pickle 与制造自主堆场牵引车的 Outrider、研制灵巧人形分拣机器人的 Reflex、专注于码垛包裹的 Ambi Robotics（AmbiSort）以及投资方 Teradyne（Universal Robots 母公司）的自主移动托盘车紧密协同，共同组成覆盖整个物流链条的实体 AI 舰队。
```mermaid
flowchart TD
accTitle: 实体 AI 决策闭环流程
subgraph Math["贝尔曼决策模型 (Bellman Equation)"]
P["感知 Perception (X)<br/>机器视觉/多模态传感器"]
G["目标 Goal (f)<br/>作业任务定义"]
T["状态转移与规划 Transition (T)<br/>预测物理世界响应并制定计划"]
A["行动 Action (A)<br/>机械臂抓取与底盘移动"]
P --> T
G --> T
T --> A
A --> P
end
```
### [10:08](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=608s) 实体 AI 的数学本质与半个世纪的演进史
究竟何为实体 AI？AJ 将其核心溯源至 1950 年代提出的强化学习基石——贝尔曼方程（Bellman Equation）。实体 AI 的核心由四个相互关联的要素构成：感知（$X$）、目标（$f$）、状态转移与推演规划（$T$）以及行动（$A$）。系统不仅要看清周围环境，更要能够在大脑中“预演”自身动作对物理世界产生的反应，据此规划最佳路径来实现既定目标。
相比之下，目前耗资 750 亿美元的传统工业自动化体系（如流水线汽车车门焊接、传送带上的可颂面包整理等）在本质上更像“按既定轨道行驶的火车（akin to a train）”。这些机械臂完全沿着预设轨迹运行，传感器仅充当“加速、减速或急停”的触发开关，缺乏主动设定目标、自主规划路径和认知物理规律的智能。
回顾实体 AI 的学术探索史，其开端可以追溯到数十年前：
- **1966 年（斯坦福研究院 SRI）**：诞生了名为 **Shakey** 的移动机器人，首次配备了摄像头与机器视觉，能自主建图并在实验室内规划导航路线。
- **1986 年（MIT）**：AJ 的导师 Tomás Lozano-Pérez 教授在其博士课题中推出了 **Handy** 机器人，打破传统固定轨迹限制，基于环境感知与自主规划实现 Pick-and-Place 抓取作业。
- **1990 年代（MIT 腿足实验室 Leg Lab）**：波士顿动力创始人 Marc Raibert 与 Pickle 联合创始人 Dan Paluska 等人在此深耕，早于大众熟知的 Atlas 数十年便已实现了机器人的后空翻与双足平衡控制，并在 2001 年打造出全功能人形双足机器人。
然而，从 1954 年到 2001 年近半个世纪里，倾注了大量 DARPA 研发资金的自主思考机器人项目在商业上均以失败告终——由于技术过于脆弱、成本奇高且缺乏泛化能力，未能交付任何具备商业实用价值的成熟产品。
### [14:19](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=859s) 物理经验学习的突破与空气动力学推断
这一僵局在 2004 年迎来历史性转折。MIT 教授 Russ Tedrake 在其名为 "Toddler" 的双足行走机器人博士课题中，首次尝试让实体 AI 摆脱纯仿真环境或棋类博弈的限制，直接利用真实物理交互数据（Experience Data）进行强化学习。AJ Meyer 于 2005 年加入该实验室，亲历了这一从物理数据中学习控制规律的技术突破。
在 2025 年的技术语境下，利用物理世界数据进行实体 AI 训练的路径主要划分为四种本体论范式：
- **生成式模型（Generative Models）**：对物理环境变化与未来状态进行预判和“想象”；
- **判别式模型（Discriminative Models）**：进行精准的物体检测、分类与位姿识别；
- **强化学习（Reinforcement Learning）**：基于真实物理奖励反馈进行策略微调；
- **传统推断与回归（Traditional Inference & Regression）**：拟合复杂动力学规律以实现极限制导。
AJ 结合其早年在 MIT 参与的“栖息滑翔机（Perching Glider）”前沿项目解释了物理推断的重要性：当小型泡沫滑翔机在尝试如鸟类般精准降落在电线上时，飞机会主动进入大迎角“失速（stall）”状态。在这一瞬间，传统经典空气动力学的解析数学公式完全失效（completely untenable）。研发团队必须依靠从物理飞行数据中推断和学习非线性气动模型，进而设计出能够自主控制姿态并精准停靠在细电线上的实体 AI 控制系统。这种从未知物理规律中学习并反哺控制决策的能力，构成了当代实体 AI 突破工业落地边界的核心驱动力。
### [16:22](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=982s) 从 ASIMO 视觉到特技直升机：物理 AI 的早期探索
在回顾物理 AI 的演进历程时，AJ 分享了自己在早期技术探索中的亲身经历。早在 2008 年，他在本田研究院（Honda Research）短暂工作，参与了知名人形机器人 ASIMO 的视觉系统开发。当时他们团队成为了业内最早利用 GPU 加速合成数据训练管线的先驱之一：通过构建一个判别器（Discriminator），系统会主动评估 ASIMO 是否能够成功识别当前图像；一旦判定可能无法识别，管线就会自动注入更多数据进行强化训练。这套机制在当时取得了难以置信的成功。
随后，AJ 重点提及了斯坦福大学 Pieter Abbeel 在 2000 年代后期具有里程碑意义的博士论文。他极力向现场观众推荐去搜索相关视频：“如果你听完这场演讲只想去 YouTube 搜一个视频，一定要看这个，它简直令人难以置信。” 该研究展示了一架能够完成媲美人类顶级水准特技飞行动作的自主直升机。这项突破的核心正是如今被称为模仿学习（Imitation Learning）或演示学习（Learning from Demonstration）的技术——通过利用在人类专家操纵数据上训练出的推理模型来反向学习目标函数。因为在物理 AI 领域，很多时候单凭手工编写代码来精确定义机器人的目标是极其困难的。
- **ASIMO 判别器数据管线**：2008 年利用 GPU 加速，通过判别模型动态检测视觉盲区并补充合成数据。
- **Pieter Abbeel 特技直升机**：利用模仿学习让算法自主掌握人类极限级别的控制策略，解决了物理动作目标难以数学化描述的痛点。
---
### [17:20](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1040s) 扩散策略与端到端架构：现代物理 AI 的四大模型支柱
时间跨入最近几年，物理 AI 迎来了爆发式的范式转移。AJ 重点展示了两项前沿突破：其一是大约两年前诞生的“扩散策略”（Diffusion Policy）。它借鉴了 Midjourney、Stable Diffusion 以及前沿视频生成模型中为人熟知的加噪-去噪（Noise-to-Data）机制，初始阶段将机器人动作视为噪声分布，随后通过真实物理数据训练逐步去噪，生成精准的连续动作指令；其二是来自哈佛大学杜逸伦（Yilun Du）团队的完整“世界模型/幻觉模型”（Hallucinated Model）。画面中呈现的一切动作场景完全由模型合成，机器人借由这种“富有想象力的幻觉”在脑海中预演动作过程，从而做出更优的决策，比如自主判断如何拉开抽屉并将香蕉准确放进去。
```mermaid
flowchart TD
accTitle: 物理 AI 端到端技术架构
subgraph Models["四大核心模型支柱"]
G["生成模型 Generative<br/>(扩散策略 / 想象预演)"]
D["判别模型 Discriminative<br/>(视觉识别 / 质量检验)"]
R["强化学习 RL<br/>(策略优化 / 闭环反馈)"]
I["推理模型 Inference<br/>(意图理解 / 目标推导)"]
end
subgraph Pipeline["端到端控制流程 (End-to-End Robotics)"]
S1["环境感知 Perception"] --> S2["路径规划 Planning"]
S2 --> S3["物理执行 Action"]
end
Models --> Pipeline
```
到了 2025 年，曾经困扰物理 AI 领域数十年的难题——依赖博士们熬夜手写复杂的感知、规划、目标设定与运动控制规则——彻底被数据驱动的学习工具所取代。生成模型（Generative）、判别模型（Discriminative）、强化学习（RL）与推理模型（Inference）这四大支柱如今贯穿了整个技术栈，构建起了新兴的“端到端机器人学”（End-to-End Robotics）。
- **扩散策略在物理控制的应用**：从纯动作噪声起步，经由真实世界物理交互数据迭代去噪，生成稳定动作轨迹。
- **合成预演世界模型**：利用生成式想象评估运动分支，显著提升了机械臂处理复杂交互任务时的决策容错率。
---
### [18:50](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1130s) Pickle Robot 的技术实践：7 组相机与亚毫米级的动作学习
在 Pickle Robot 的卸货机器人上，这套全新的 AI 范式得到了全面落地。曾经极其棘手的机器视觉挑战，如今被布置在机身上的 7 台 RGB 相机轻松化解。系统不仅能瞬间精准识别车厢地面、箱体表面标签以及货品边界，还能实时构建出三维环境重建，并“幻觉”预演多种不同的机械臂运动轨迹。
在运动控制层面，Pickle 正在利用机器学习捕捉极其细微的动态差异。AJ 指出，在机械臂高速移动的过程中，肉眼甚至很难看出两次抓取路径之间有何不同，但正是那些极为微小的姿态与受力偏差，决定了包裹究竟是被稳稳放置还是中途滑落。通过针对这些亚毫米级的运动细节进行模型学习与纠偏，机器人得以在保障零跌落的前提下实现极致的作业速度，为客户带来切实的生产力价值。
- **多视角感知矩阵**：全机搭载 7 组 RGB 相机，完成实时 3D 点云重建与标签/地面语义分割。
- **高动态微动作优化**：解决包裹在高速搬运过程中的脱落风险，依靠海量微调消除人工示教无法捕捉的细微物理干扰。
- **研发投入规模**：为了将这一软硬件系统打磨至目前的量产级性能，公司已累计投入数千万美元。
---
### [19:50](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1190s) 跨越采纳鸿沟与商业爆发：卡车卸载成为首个杀手级应用
在演讲的收尾阶段，AJ 兴奋地宣布，至少在卡车自主卸载这一物理 AI 细分赛道上，行业已经成功跨越了“早期采纳者”（Early Adopters）阶段，正式迈入陡峭的规模化高速普及期。他坚信货箱卸载只是物理 AI 的第一个杀手级应用（Killer Application），后续必将涌现出更多的垂直场景与创业团队。AJ 感慨道：“在我从事这一行业的 25 年里，眼下绝对是最令人兴奋的黄金时代。”
对于打造硬科技产品的商业方法论，AJ 也更新了他的“独门秘诀”（Secret Sauce）：如今开发者不再需要从零开始耗巨资预训练大模型，而是应该以通用 Foundation Models（基础模型）为基座，把全部精力集中在利用客户现场的真实业务问题数据进行微调（Fine-tune）。
依靠这套精准的产品与技术打法，Pickle Robot 在 2025 年下半年取得了强劲的商业化突破。公司已经正式交付了储备管线中数百家意向客户里的首批 10 家标杆企业，客户群体不仅包括像亚马逊（Amazon）这样的物流与电商巨头，还广泛渗透到了 Zara、Puma 以及 Helly Hansen 等全球知名快消与服装零售品牌的供应链腹地。
- **技术方法论升级**：从底层自研模型转向“基础大模型 + 客户现场问题数据微调（Fine-tuning）”。
- **商业化落地进展**：管线中储备了数百家企业订单，首批 10 家正式部署客户横跨物流巨头与顶级零售服饰品牌。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=4s) 整个机器人领域当前正处于某种“迎来高光时刻的临界点（having a bit of a moment）”，一场巨大的行业变革呼之欲出。 — AJ Meyer
> [03:34](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=214s) 整个传统机器人与自动化行业自诞生以来的商业化总规模每年仅约 750 亿美元，在 50 万亿美元的盘子面前不过是“沧海一粟（a drop in the bucket）”。 — AJ Meyer
> [03:34](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=214s) 孤立地看待单台机器人是狭隘的（myopic），真正的价值爆发点在于设备之间相互通信、协同作业所构建的“网络效应（network effect）”。 — AJ Meyer
> [10:08](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=608s) 这些机械臂完全沿着预设轨迹运行，传感器仅充当“加速、减速或急停”的触发开关，缺乏主动设定目标、自主规划路径和认知物理规律的智能。 — AJ Meyer
> [14:19](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=859s) 当小型泡沫滑翔机在尝试如鸟类般精准降落在电线上时，飞机会主动进入大迎角“失速（stall）”状态。在这一瞬间，传统经典空气动力学的解析数学公式完全失效（completely untenable）。 — AJ Meyer
> [16:22](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=982s) 如果你听完这场演讲只想去 YouTube 搜一个视频，一定要看这个，它简直令人难以置信。 — AJ Meyer
> [18:50](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1130s) 在机械臂高速移动的过程中，肉眼甚至很难看出两次抓取路径之间有何不同，但正是那些极为微小的姿态与受力偏差，决定了包裹究竟是被稳稳放置还是中途滑落。 — AJ Meyer
> [19:50](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1190s) 在我从事这一行业的 25 年里，眼下绝对是最令人兴奋的黄金时代。 — AJ Meyer
> [19:50](https://www.youtube.com/watch?v=kn1DQqyWCGI&t=1190s) 至少在卡车自主卸载这一物理 AI 细分赛道上，行业已经成功跨越了“早期采纳者”（Early Adopters）阶段，正式迈入陡峭的规模化高速普及期。 — AJ Meyer
## Mentioned
Pickle Robot Company, AJ Meyer, 黄仁勋 (Jensen Huang), 软银 (SoftBank), 霍尼韦尔 (Honeywell), 波士顿动力 (Boston Dynamics), Outrider, Reflex, Ambi Robotics, AmbiSort, Teradyne, Target, 贝尔曼方程 (Bellman Equation), Shakey, 斯坦福研究院 (SRI), Handy, Tomás Lozano-Pérez, Marc Raibert, Dan Paluska, Russ Tedrake, Toddler (机器人), ASIMO, 本田研究院 (Honda Research), Pieter Abbeel, 杜逸伦 (Yilun Du), Diffusion Policy (扩散策略), Midjourney, Stable Diffusion, Amazon, Zara, Puma, Helly Hansen
## Video description
AJ Meyer, Founder & CEO, Pickle Robot Company
Abstracts
While foundation models are revolutionizing digital workflows, AJ explores how generative AI is now being applied to physical supply chain operations - turning perception into movement and predictions into action to drive autonomous robots that unload trailers and
import containers. See how robots with physical AI move goods, not just data, to deliver real ROI for real companies.
[Source](https://www.youtube.com/watch?v=kn1DQqyWCGI)
Source quotes
[0:04] AJ Meyer: 整个机器人领域当前正处于某种“迎来高光时刻的临界点（having a bit of a moment）”，一场巨大的行业变革呼之欲出。
[3:34] AJ Meyer: 整个传统机器人与自动化行业自诞生以来的商业化总规模每年仅约 750 亿美元，在 50 万亿美元的盘子面前不过是“沧海一粟（a drop in the bucket）”。
[3:34] AJ Meyer: 孤立地看待单台机器人是狭隘的（myopic），真正的价值爆发点在于设备之间相互通信、协同作业所构建的“网络效应（network effect）”。
[10:08] AJ Meyer: 这些机械臂完全沿着预设轨迹运行，传感器仅充当“加速、减速或急停”的触发开关，缺乏主动设定目标、自主规划路径和认知物理规律的智能。
[14:19] AJ Meyer: 当小型泡沫滑翔机在尝试如鸟类般精准降落在电线上时，飞机会主动进入大迎角“失速（stall）”状态。在这一瞬间，传统经典空气动力学的解析数学公式完全失效（completely untenable）。
[16:22] AJ Meyer: 如果你听完这场演讲只想去 YouTube 搜一个视频，一定要看这个，它简直令人难以置信。
[18:50] AJ Meyer: 在机械臂高速移动的过程中，肉眼甚至很难看出两次抓取路径之间有何不同，但正是那些极为微小的姿态与受力偏差，决定了包裹究竟是被稳稳放置还是中途滑落。
[19:50] AJ Meyer: 在我从事这一行业的 25 年里，眼下绝对是最令人兴奋的黄金时代。
[19:50] AJ Meyer: 至少在卡车自主卸载这一物理 AI 细分赛道上，行业已经成功跨越了“早期采纳者”（Early Adopters）阶段，正式迈入陡峭的规模化高速普及期。
Source quotes
[0:04] AJ Meyer: Physical AI
[10:08] AJ Meyer: train on rails
[14:20] AJ Meyer: Toddler
[14:20] AJ Meyer: Perching Glider
[16:22] AJ Meyer: If you have to YouTube anything in the talk, YouTube this. It's unbelievable.
[17:24] AJ Meyer: Diffusion policy
[18:14] AJ Meyer: end-to-end robotics
[18:14] AJ Meyer: tens of millions of dollars
[19:52] AJ Meyer: killer application
[19:52] AJ Meyer: secret sauce
Original transcript
[0:04] I'm AJ. I'm the founder of Pickle Robot
[0:06] about a 130 person robotic startup uh
[0:08] out of Boston, part of the MIT
[0:10] community. And some of you may be aware
[0:13] that robotics is having a bit of a
[0:15] moment right now. Seems like we're whole
[0:17] field is on the cusp of something really
[0:19] big. Uh and so I think it's a very
[0:22] appropriate venue here at this AI
[0:24] focused conference to kind of talk about
[0:26] why that might be the case. Um because
[0:30] fundamentally the current moment that
[0:33] robotics is having is completely
[0:35] underwritten by the tremendous progress
[0:37] that's being made in AI and has been for
[0:39] the last several years. Uh so much so
[0:42] that we now have a new emerging term
[0:44] that you might see in the headlines from
[0:45] time to time. I know Jensen likes to
[0:47] talk about it all the time, physical AI.
[0:50] Uh and you know I get two questions all
[0:53] the time. The first one is what the heck
[0:55] is physical AI? So I hope to answer that
[0:57] question for you today. And then the
[0:59] other question that I get all the time
[1:01] is why is it the case that companies
[1:04] like mine, Pickle Robot, the most famous
[1:07] name maybe in robotics, Boston Dynamics,
[1:09] Soft Bank, Honeywell, a slew of other
[1:11] startups, are literally investing
[1:14] hundreds of millions of dollars into the
[1:17] problem of unloading and loading trucks,
[1:19] which seems a little unobvious. So
[1:21] hopefully we'll get a chance to talk
[1:22] about those things. um you know the
[1:24] basic uh stuff I want to cover maybe
[1:26] I'll start with our vision for you know
[1:28] not just what our company can be but
[1:30] what's going on in the industry and how
[1:31] we see it developing uh over the next
[1:33] several years uh and I'd like to talk a
[1:35] bit about our approach on how to get
[1:36] there uh but most importantly I really
[1:39] want to help contextualize for you if
[1:41] the claim is that robotics is having a
[1:42] moment and is at the cusp of a
[1:44] revolution you know what's the
[1:46] historical context for that uh and where
[1:49] did this all come from uh and then
[1:51] finally I want to catch you up on maybe
[1:52] kind of the latest going on in the field
[1:54] and at our at our company.
[1:57] So, to start, I'm just going to uh play
[2:00] about a 90-second clip here that
[2:02] introduces you to the uh pickle.
[2:06] That doesn't seem to want to uh play
[2:09] here. There we go.
[2:14] [Music]
[2:26] You are the
[2:31] robots.
[2:37] [Music]
[2:47] Hey,
[2:52] [Music]
[2:58] [Music]
[2:59] hey,
[3:03] hey.
[3:08] Cool.
[3:11] [Music]
[3:14] [Applause]
[3:15] [Music]
[3:27] So, everybody with their cell phones
[3:28] out, we've never actually published that
[3:30] clip before, so you guys are the only
[3:31] ones that have it.
[3:34] Anyway, so why the doctor door? Um,
[3:37] great question.
[3:38] Something to know about robotics and
[3:40] automation is that fundamentally what
[3:42] it's for is trying to make a dent in the
[3:45] labor bill on the supply chain, which is
[3:47] about $50 trillion. And so the entirety
[3:51] of what robotics and automation have had
[3:53] to deliver so far in terms of commercial
[3:55] impact since inception is a $75 billion
[3:59] a year market. So a drop in the bucket
[4:01] against 50 trillion. Meanwhile, loading
[4:04] and unloading trucks is another $75
[4:07] billion problem. So suddenly there's a
[4:10] task out there, a particular application
[4:12] that has the potential to double the
[4:14] impact of all of robotics and automation
[4:17] that came before. So it's worth asking
[4:19] the question, why has nobody tried this
[4:21] before? Uh and the answer is many people
[4:23] have. And I think uh it's also important
[4:26] to understand that unloading and loading
[4:28] trucks may be just the first of what
[4:29] will be a whole slew of new products
[4:31] that are taking multi-billion dollar
[4:33] slices uh by bites out of the uh supply
[4:36] chain automation problem.
[4:38] So here's uh actually the very same
[4:40] building where that last clip was was
[4:42] coming from. Uh this is a building in
[4:43] Los Angeles. Uh what happens in this
[4:46] building, it's called a cross dock, is
[4:47] that boats arrive from overseas. A truck
[4:50] drives the freight up to one of those
[4:52] doors on the lefth hand side. There's
[4:53] about 150. Uh, a team of people are
[4:56] going to go in the truck, which can be
[4:58] 130° in the California summer. U, move
[5:00] all those boxes onto those carts, sort
[5:02] out a bunch of different orders. Then
[5:04] they're going to push the carts to the
[5:05] right hand side of the building, and
[5:06] then they're going to load the truck
[5:08] again where ultimately the material in
[5:10] the truck is going to end up on the
[5:11] shelf at Target. Uh, and this is exactly
[5:13] what it looks like today. So, not a lot
[5:15] of technology. If the $75 billion
[5:18] robotics and automation market had
[5:19] something to offer to this customer in
[5:21] this problem, they would have. Uh so
[5:23] something has changed that that kind of
[5:25] uh lets us address this for the first
[5:26] time.
[5:28] Uh and so what I want to maybe convince
[5:30] you of uh in the first section of the
[5:33] talk here is that we're not really
[5:35] talking about just loading and unloading
[5:36] trucks. We're talking about everything I
[5:38] just walked you through because it's a
[5:40] multi-step process the supply chain with
[5:42] a lot of buildings. So, one robot might
[5:44] be unloading a truck, but another one is
[5:46] going to be starting the boxes onto
[5:48] carts, and then another one is going to
[5:50] drive the cart to the other side of the
[5:51] building, and then another one's going
[5:52] to load the truck, and hopefully pretty
[5:54] soon, in the next four, or five years
[5:55] here, we'll have over the road trailers,
[5:57] uh, tractor trailers that'll be able to
[5:58] drive the truck to the next building.
[6:00] So, it's myopic to be thinking about one
[6:02] robot doing a job. What we ought to be
[6:04] thinking about is the network effect
[6:06] that's created by robots talking to each
[6:09] other and coordinating the behaviors to
[6:11] act as a team. Uh and that's the way
[6:13] we're going to create the most value for
[6:14] customers because such a network doesn't
[6:16] just scale a single building, but it
[6:19] actually scales across buildings in a
[6:21] particular metro and ultimately through
[6:23] the national and international supply
[6:25] chains. So we're talking about building
[6:26] one network uh out of an emerging class
[6:29] of robots that I've called physical AI.
[6:32] Let's take another look at that
[6:33] building. So this is a uh overhead view
[6:35] of that same customer. Uh and there's
[6:38] literally thousands of buildings like
[6:40] this in the United States. I encourage
[6:41] you to uh next time you're on an
[6:43] airplane, look out the window as you're
[6:44] landing or taking off and just count how
[6:46] many dock doors we have uh with trucks
[6:48] on them. And so it starts in the yard
[6:51] where these uh uh trailers get um get uh
[6:55] a staged there and then we need to bring
[6:56] them onto the lefth hand doors. Then we
[6:58] need to process the material and then we
[6:59] need to bring them off of the right-hand
[7:01] doors and drop them back in the yard so
[7:02] a tractor trailer can pick them up. uh
[7:05] and there's an emerging group of
[7:07] companies that are all converging around
[7:09] a secret sauce for how to build a
[7:12] product that might have a role to play
[7:14] uh in this transformation. Uh and the
[7:16] secret services secret sauce is uh made
[7:19] up of hardware software uh but some
[7:22] other elements as well. So on the
[7:23] hardware side, I don't have too much I
[7:25] want to say about it other than 2025 is
[7:27] remarkable in that we have tremendously
[7:30] powerful, tremendously reliable and very
[7:32] inexpensive access to compute and to
[7:35] sensors and to motors that we did not
[7:37] have before. Uh and so if you take those
[7:40] things from the catalog and the various
[7:41] suppliers that produce them and combine
[7:43] them with some proprietary hardware of
[7:44] your own, like in our case the hand
[7:46] itself is a very important innovation
[7:48] that helped us pick up 60 lb boxes from
[7:50] over our head. uh you've basically got
[7:52] the the the guts of of the robot, the
[7:54] visible part. On the software side,
[7:56] well, we're going to talk all about
[7:57] that. That's be the physical AI
[7:59] component. Although, I will hook on to
[8:01] what Joel was talking about about
[8:02] consolidating your data. One of the
[8:04] interesting new software challenges here
[8:05] is that we produce in some cases 100
[8:07] gigabytes per day per robot. So, if
[8:09] you're talking about thousands or
[8:10] millions of robots, you have an
[8:12] incredibly new scale of data challenge
[8:14] uh well beyond 80 pabytes. Um finally uh
[8:19] I think all of the successful companies
[8:22] uh or the ones that are starting to be
[8:23] emerges as leaders have focused on an
[8:25] application. Uh I know there's a lot of
[8:27] enthusiasm for foundation models,
[8:29] general purpose things in AI that are
[8:31] supposed to do everything for everyone.
[8:32] But the reality is is the most frontier
[8:34] general purpose foundation models from
[8:36] academia. We're still a long way from
[8:38] commercialization absolutely need to be
[8:40] fine-tuned on a customer problem before
[8:42] they are commercially useful. Uh, and so
[8:45] given that you're going to fine-tune it
[8:46] on a customer problem, you need some
[8:47] recipe for getting data out of your
[8:49] customer and training on it. And what
[8:51] better data to get than the data you get
[8:53] from experience. And so what we've been
[8:56] working on at Pickle is putting these
[8:57] pieces together to get a flywheel going
[8:59] where we ship a robot that's good enough
[9:00] to serve you today. It gets better as it
[9:03] does that and that lets us expand
[9:05] further into your business or to other
[9:06] customers like it. So here's some other
[9:09] products that we're all friendly with
[9:10] and these are all Pickle partners. uh
[9:12] we're starting to think of pushing these
[9:13] things together. So on the far left you
[9:15] see out rider which makes autonomous
[9:17] yard goats. Then in the middle you have
[9:19] our gen two uh pickle unload and load
[9:21] robot. You have reflex a humanoid robot
[9:23] on the right hand side which is quite
[9:24] dextrous and good at picking up pieces
[9:26] of paper and sorting packages. You have
[9:28] ambi sort on the bottom left which
[9:29] stacks boxes onto pallets or other
[9:31] robots. Uh you have a a uh autonomous
[9:34] forklift or it's actually a pallet jack
[9:36] from Pterodine which is one of our
[9:37] investors and a key supplier for us.
[9:39] Some people might know them as Universal
[9:40] Robotics. And then ultimately those
[9:43] things all go in a chain and deliver
[9:45] some boxes to another pickle robot that
[9:47] loads the truck and on it goes back to
[9:48] the out rider.
[9:50] And so here's that building again to
[9:52] just kind of remind yourself of how all
[9:53] those products are going to fit in. This
[9:55] is what the building looks like today.
[9:56] Modulo, the small handful of robots we
[9:58] have in the corner. But we're really uh
[10:01] very close to completely transforming
[10:03] what this and many thousands of other
[10:04] operations across the supply chain look
[10:06] like.
[10:08] So I want to tell you about why what is
[10:10] physical AI? Um and it goes back to an
[10:13] equation that probably few of you
[10:15] recognize but maybe a couple people do.
[10:17] Uh this is called a Bellman equation and
[10:18] it's very very common in the field of
[10:20] reinforcement learning. First introduced
[10:22] in the 1950s and while it may look
[10:25] complicated uh its mystique goes away
[10:27] once I give you names for what all those
[10:29] little symbols are. So anywhere you see
[10:31] X we're talking about perception. Some
[10:33] you know sometimes that's machine vision
[10:34] but there's other sensors to process
[10:36] besides cameras. Uh when we're looking
[10:38] at that function f, we're talking about
[10:39] giving our system a goal. Uh when we
[10:42] talk about this transition function t,
[10:44] we're talking about the ability to
[10:45] imagine what the world is going to do in
[10:48] reaction to the things that we do to it.
[10:50] Imagining that allows us to make plans.
[10:52] Uh and making plans allows us to meet
[10:54] our goals. And then finally, anywhere
[10:56] you see an A, we're talking about
[10:57] action. And what's interesting about
[10:58] this equation is that of the current $75
[11:01] billion or so being spent on robotics
[11:03] and automation uh of the current ilk,
[11:05] things that uh weld car doors in in
[11:08] factories or maybe straighten up
[11:10] croissants on an assembly line uh or do
[11:12] some um like pick and place tasks, they
[11:15] do not use this. Uh fundamentally what
[11:18] you see when you see uh any one of those
[11:20] uh established kind of mature robotic
[11:22] and automation technology is something
[11:24] akin to a train. You have to build a set
[11:27] of rails that determine how the thing is
[11:29] able to move. And then we can use
[11:31] sensors and machine vision and other
[11:32] advanced techniques to kind of speed up
[11:34] or slow it down as it moves. Maybe we
[11:37] can start and stop it, trigger it to
[11:38] make a behavior. Or maybe if it's very
[11:40] advanced, we can change the start and
[11:42] destination locations or maybe some
[11:44] small tweak to how the gripper is
[11:46] oriented. But it has nothing to do with
[11:47] this fancy piece of math that I've put
[11:49] on the board. It doesn't have goals. It
[11:51] doesn't make plans. It doesn't perceive.
[11:54] So this is what I think is the origins
[11:57] of the field of physical AI is building
[11:59] robotic systems and software that do
[12:00] these four things. And this was the
[12:02] first one. Its name is Shaky. It's from
[12:05] the Stanford Research Institute.
[12:06] Originally built in 1966. And would you
[12:09] believe me if I told you this thing had
[12:11] cameras. It used machine vision. It
[12:13] built its own maps. It was able to think
[12:15] for itself and autonomously navigate the
[12:17] lab at SRRI, the AI lab at SRRI.
[12:21] And maybe on that point uh you know what
[12:24] is physical AI kind of forces you to ask
[12:26] the question what is AI? Uh and we'll
[12:28] come back to that in a minute. Here's
[12:30] another one. Uh this was at MIT. Uh this
[12:33] one is called Handy. And I'm also not
[12:35] seeing the video run. So if we could get
[12:36] some assistance on clicking play there.
[12:40] And this was the PhD thesis for Tomas
[12:42] Lozano Perez who was one of my
[12:43] professors when I was at MIT. We want to
[12:45] go back one slide. Here we go. And this
[12:48] is a type of pick and place robot that's
[12:49] able to pick things up and put them
[12:51] down. Hopefully we can get uh the play
[12:53] button to happen. Um but it thinks for
[12:56] itself and so it doesn't have those set
[12:58] of train rails. It absolutely makes has
[13:00] goals. It makes plans. It has
[13:02] perception. Uh and it it learns
[13:04] different types of actions. And um
[13:08] looks remarkably non-commercializable to
[13:10] me, but it was also a miracle of
[13:12] science. And uh Tamas is a giant in the
[13:15] field. Uh doing this type of thing uh
[13:17] whether it was in 1986 or it's today is
[13:20] actually very very difficult. Um moving
[13:22] on to the '9s also at MIT. This is the
[13:25] leg lab. Let's see if we can get that
[13:27] video going. U my co-founder Dan Palooa
[13:29] was in the leg lab around this time. You
[13:32] might be surprised to know that robots
[13:34] have been doing flips for maybe a lot
[13:36] longer than you realized when you saw
[13:37] Atlas do one a few years ago. Uh, and
[13:40] you're probably also not surprised to
[13:41] learn that uh this lab here at MIT is
[13:44] the um it is the the the seeds of Boston
[13:48] Dynamics. Uh so Mark Rabert was a
[13:50] student, the founder of Boston Dynamics
[13:52] along with Tomas uh Perez, who I just
[13:54] showed you his work. And um so this was
[13:57] all going on in the 90s. And I'm sorry
[13:58] the the video Oh, here we go.
[14:01] And let's move on.
[14:08] Here's Dan in as my co-founder in the
[14:12] year 2001, I think, with his humanoid uh
[14:15] robot equipped with legs and everything.
[14:18] So anyway, let's recap the history. From
[14:20] 1954 to 2001, uh the program of giving
[14:24] robots the ability to think for
[14:25] themselves, took years to develop, took
[14:27] tremendous amounts of DARPA money, uh
[14:30] delivered uh zero commercially useful
[14:33] robots, too expensive, too brittle, not
[14:35] capable enough. But then something
[14:37] changed. So this is MIT about the time I
[14:39] arrived. Uh so this is Russ Tedre's PhD
[14:42] thesis called toddler and this is the
[14:44] first time anyone tried to get one of
[14:46] these physical AIs to learn. Uh so this
[14:49] uses reinforcement learning and the the
[14:52] robot actually learns not just uh in
[14:54] simulation but from real physical data
[14:56] from experience. Um and this was a
[14:59] breakthrough uh primitive though it may
[15:01] be in 2004
[15:03] and I joined the lab in05
[15:06] uh and maybe there was yeah and so I
[15:09] want to talk a little bit about learning
[15:10] from physical data because you can learn
[15:11] from other things reinforcement learning
[15:13] had been playing games for a long time
[15:14] back gamon and chess and things like
[15:16] this but learning from physical data is
[15:18] its own thing and now in 2025 there's so
[15:21] much excitement around AI I think
[15:22] there's a new set of um understanding
[15:25] about the ontology here so let's say
[15:27] there's four different ways to learn
[15:28] from physical data. There's generative
[15:29] models that try and hallucinate uh is
[15:32] the word we all have kind of come to
[15:33] know. Uh there's discriminate which is
[15:35] is it a cat or is it a dog. There's
[15:37] reinforcement learning which hopefully
[15:38] some of you have some experience with
[15:40] like fine-tuning and large language
[15:41] models. And then there's traditional
[15:43] inference and regression. And all four
[15:45] of these methods allow you to turn real
[15:46] data from the physical world and speak
[15:49] to one of those four things that a
[15:51] physical AI has to do. Set goals, plan,
[15:54] take action. Um and so here were three
[15:57] two projects that I was involved with
[15:58] and one that I was not that spoke to
[16:00] that. Um so uh this is perching glider
[16:03] on the left and what we were doing was
[16:05] taking these little foam airplanes and
[16:07] stalling them where the mathematics of
[16:09] the of the physics of that machine
[16:12] become completely untenable. And so what
[16:14] we're forced to do is learn using
[16:15] inference what those dynamics are so
[16:17] that we can then design um physical AIs
[16:20] that control them to land on a wire. So
[16:22] they perched like a bird. Uh then I
[16:24] spent a short stint at Honda research uh
[16:26] working on Osimo and I worked on the
[16:28] vision system for Osimo and we were
[16:30] actually one of the first groups to use
[16:32] GPU accelerated synthetic data training
[16:34] pipelines in '08 by building a
[16:36] discriminator that would ask whether or
[16:38] not we expected Osimo to recognize that
[16:41] image or not. And if we thought it
[16:42] wasn't going to recognize it, we added
[16:44] more data to the pipeline. Um and that
[16:46] was incredibly uh successful. And then
[16:48] uh on the final piece and this is a
[16:50] something I had nothing to do with with
[16:52] at all. Uh Peter Aiel's PhD thesis at
[16:54] Stanford uh in the late as if you have
[16:57] to YouTube anything in the talk YouTube
[16:58] this. It's unbelievable. It's a
[17:00] helicopter that can do acrobatic flight
[17:02] at the level of a human being. And the
[17:04] trick was something we now call
[17:05] imitation learning or learn from
[17:07] demonstration. It was to use an
[17:08] inference model trained on data of a
[17:10] human being controlling uh such a
[17:12] helicopter to learn the goal because
[17:15] sometimes describing what the goals are
[17:16] for physical AI is very very difficult.
[17:20] So that kind of get catches up to about
[17:22] 2010 what's been going on recently. Uh
[17:24] so these are also videos if we can get
[17:26] them to play. On the left hand side this
[17:28] is a breakthrough from just two years
[17:29] ago. This is called diffusion policy and
[17:32] this is using the exact same noise to
[17:34] data format that we know about uh from
[17:37] from like midjourney and stable
[17:39] diffusion and any of the video gener
[17:40] video generation models but using it to
[17:43] start with noise about how a robot might
[17:44] act uh and then having the robot act and
[17:47] it's all trained on real physical data.
[17:49] And then on the other side we have
[17:51] something you may not be surprised
[17:52] exists but this is a full hallucinated
[17:54] model from our friend uh Yulundu over at
[17:57] Harvard. Uh, and you know, everything
[17:59] you see on the right side is completely
[18:01] synthesized, right? It's a
[18:02] hallucination, but it's an important
[18:04] hallucination that's been designed to
[18:05] allow the robot to imagine what it's
[18:07] doing, uh, and then therefore make
[18:09] better choices about how to open the
[18:11] drawer and put the banana in it and so
[18:12] forth.
[18:14] And so, you know, now in 2025, all of
[18:17] these things that were extremely hard in
[18:19] physical AI that required a lot of long
[18:21] nights and PhDs to come up with ways to
[18:23] write software about setting goals and
[18:25] perceiving and planning and acting, we
[18:27] now have tools for learning from data
[18:29] using those four types of models,
[18:31] generative, discriminative, uh
[18:33] reinforcement learning, and uh and
[18:35] inference to speak to literally the
[18:37] entire stack. And that is brand new uh
[18:39] just in the last few years. And in fact,
[18:41] if you get all four pieces going with
[18:43] learned models from data, we call it
[18:45] end-to-end robotics, which is kind of an
[18:47] emer emerging category of how to control
[18:49] them. Uh, and so this is all our robot
[18:51] uh at Pickle. And you can see um what
[18:53] was once a very difficult machine vision
[18:55] problem. We have seven RGB cameras on
[18:57] the robot. We now make pretty short work
[18:59] of understanding where the floors uh uh
[19:02] uh where's the floor, where's the
[19:03] labels, everything else, building 3D
[19:05] reconstructions of the scene,
[19:06] hallucinating different ways to move.
[19:08] And on the bottom right, we're actually
[19:10] using uh a ways to learn very subtle
[19:13] differences in the motion. You almost
[19:14] can't see it differently with your eye,
[19:16] but when the robot moves, um sometimes
[19:18] it drops the package and sometimes it
[19:19] doesn't based on very small variations
[19:21] that we now have to learn to do
[19:22] correctly, which lets us move much more
[19:24] quickly and ultimately deliver more
[19:26] value to the customer. Um and so I'm
[19:29] starting to run out of time, but I just
[19:30] want to run this clip for a few seconds
[19:32] to show you what we're doing today in
[19:33] production.
[19:35] A lot more sophisticated than even just
[19:36] a couple of years ago. and I think is
[19:38] just proof that the formula uh for the
[19:40] secret sauce of how to break into this
[19:42] market uh is really starting to work
[19:45] although it did cost us tens of millions
[19:47] of dollars to get a product up to this
[19:48] level of performance. Um
[19:52] so in conclusion uh I think that it's
[19:55] fair to say that at least in my area of
[19:57] physical AI robotic truck unloading we
[19:59] have now crossed the curve from early
[20:01] early adopters uh into uh uh kind of
[20:04] more broad adoption. the steep end of
[20:06] the curve. Um, and this is just the
[20:09] first killer application, I think, uh,
[20:11] in physical AI. There will be many, many
[20:13] more behind it and many, many other
[20:14] startups. I encourage you to start one.
[20:16] Uh, it's incredibly exciting time to be
[20:17] in this field. Certainly the most
[20:19] exciting in the 25 years that I've been
[20:20] doing it. Um, I'd like to take a quick
[20:23] tweak to my secret sauce and update it.
[20:25] Um, so instead of putting at the bottom
[20:27] learning from the customer data, let's
[20:29] let's say that the most important thing
[20:31] you can do now is to fine-tune uh on
[20:34] customer problem data to solve the rest
[20:36] of the problems in your stack. Um, and
[20:39] one of the great things about the last
[20:40] few years is we don't have to train
[20:42] models from scratch anymore. We start
[20:43] with foundation models and then we
[20:45] fine-tune them on our customer data to
[20:47] make the robots work right. And what
[20:48] happens when you get this formula right?
[20:50] Um, well, you get a company like Pickle
[20:52] Robot. And I'm incredibly proud of where
[20:54] we've arrived to uh in this moment in
[20:55] the back half of 25. We've deployed to
[20:58] the first 10 of what are hundreds of
[21:00] customers in our pipeline. Everyone from
[21:02] giant companies that you have no problem
[21:04] guessing guys like Amazon and whatever.
[21:06] Um all the way down to like fun brands
[21:08] you might not realize have a use case
[21:09] for this. Um you know like Zara or Puma
[21:13] or Hely Hansen or whatever. Uh we were
[21:15] >> so much AJ that is all the time that we
[21:17] have. I really appreciate
