Walden Robotics Deploys AI Robot Fleet on US Auto Assembly Lines

From Tools to Teammates: Physical AI and Shared Autonomy for Flexible Generalist Robots — Ai4

Ai4

Deploying physical robots directly into active factories with remote human backup delivers immediate labor value while harvesting the on-policy failure data needed to train autonomous foundation models.

AI-assisted summary of From Tools to Teammates: Physical AI and Shared Autonomy for Flexible Generalist Robots — original source. Check the original for full context.

Source published: 2026-08-15

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 15, 2026, this presentation features Rareș Ambruș and Benjamin Burchfiel detailing Walden Robotics' approach to physical AI. According to the page description, the talk examines what it takes to transition capable AI models from the lab to real factory floors through closed-loop field data collection, rapid simulation benchmarking, and shared autonomy.

## TL;DR Rareș Ambruș presents Walden Robotics' full-stack physical AI strategy, which combines custom mobile manipulation hardware with Large Behavior Models (LBMs) to automate complex manufacturing and logistics workflows. Rather than waiting for perfect standalone autonomy, Walden deploys robots immediately to live production lines backed by remote human supervisors who intervene during edge cases, generating high-signal on-policy correction data inspired by the Toyota Production System's principle of Jidoka. The company's hardware evolved from the prototype Walden 0 to the purpose-built Walden 1 platform, supported by agentic generative simulation and automated data curation loops that rapidly turn real-world logs into validated model checkpoints. Through successful production deployments in automotive engine assembly, Walden demonstrates how shared autonomy and vertically integrated co-development turn robots from rigid tools into adaptable physical teammates.

## Atmosphere & style The presentation carries a grounded, pragmatic, and engineering-first tone, anchored by deep academic credibility from MIT and Toyota Research Institute roots. Ambruș speaks with calm, methodical precision, pointedly moving past cherry-picked lab demonstrations to emphasize the messy realities and rigorous unit economics of active factory floors. The delivery feels purposeful and authoritative, blending an ambitious vision for general-purpose physical agents with disciplined, real-world execution metrics.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=4s) Walden Robotics prioritizes immediate real-world industrial deployment over extended lab incubation, proven by a multi-month deployment on an active US automotive engine line with zero line stoppages. - [03:15](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=195s) The company builds on foundational physical AI breakthroughs developed by MIT and TRI alumni, including Diffusion Policy, UMI, OpenVLA, and Large Behavior Models (LBMs). - [03:15](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=195s) Large Behavior Models (LBMs) demonstrate that scaling data volume and diversity significantly improves robot sample efficiency, distractor resilience, and dynamic physical error recovery. - [07:21](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=441s) Hardware evolved from Walden 0 (a modified off-the-shelf platform for rapid data gathering) to Walden 1 (a custom industrial platform with hot-swappable batteries, industrial safety systems, and high edge compute). - [09:32](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=572s) Inspired by Toyota's jidoka principle, shared autonomy and remote assistance enable day-one commercial deployment while harvesting high-value on-policy correction data during edge-case interventions. - [12:49](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=769s) Physical AI training follows a three-tiered data pyramid: internet-scale vision-language pre-training, mid-training on general manipulation trajectories, and top-tier proprietary on-policy field logs. - [12:49](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=769s) Physical AI architectures are rapidly commoditizing into multimodal omni-models, making proprietary field data curation and training curricula the primary competitive moat. - [18:42](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1122s) A rapid iteration pipeline combining agentic generative simulation and automated telemetry segmentation turns raw field logs into validated model checkpoints in a matter of days. - [18:42](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1122s) Hardware-in-the-loop pre-deployment lab benchmarks show a strong positive correlation with real-world production performance on customer factory floors. - [21:03](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1263s) Transforming robots from rigid single-purpose tools into adaptive physical teammates requires vertically integrated co-development of hardware, foundation models, and agentic orchestration software.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=4s) Introducing Walden Robotics: Deploying General-Purpose Physical AI

Rareș Ambruș introduces Walden Robotics, a physical AI company emerging from stealth to tackle the challenge of putting general-purpose robots to work in real-world industrial environments. Walden operates as a vertically integrated, end-to-end physical AI enterprise, developing its own custom hardware, low-level and high-level software stacks, frontier physical AI models, and customer-facing application layers. Ambruș emphasizes that while physical AI has officially crossed a technological inflection point and stands "on the verge of fundamentally impacting our lives," the industry must move past lab demos to prove real-world efficacy, establish sound unit economics, and deeply integrate with industrial customers.

Rather than remaining confined to long research cycles, Walden prioritized immediate real-world deployment. Just two months after founding the company, the team launched their first pilot in March at one of the largest automotive manufacturing plants in the United States. Following the pilot's success, Walden returned to the facility between May and July to place their robots on the active production line full-time. During this multi-month deployment, the robots contributed to manufacturing thousands of automotive engines, operating continuously without slowing down the assembly line or causing a single line stoppage.

```mermaid flowchart LR accTitle: Walden Robotics Vertically Integrated Deployment Pipeline subgraph Core["Full-Stack Development"] HW["Proprietary Hardware"] SW["Software & Platform"] AI["Frontier Physical AI"] end subgraph Deployment["Real-World Deployment"] Auto["Automotive Assembly Line"] Scale["Engine Manufacturing"] end HW --> Auto SW --> Auto AI --> Auto Auto --> Scale ```

- **Company Profile**: Vertically integrated Physical AI startup emerging from stealth, founded by roboticists and AI researchers. - **Production Milestone**: Operated full-time on an active US automotive manufacturing line between May and July, helping build thousands of engines with zero line stoppages.

---

[03:15](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=195s) Heritage from TRI, Diffusion Policy, and Large Behavior Models

The foundation of Walden Robotics rests on extensive academic and applied robotics research. Chief Executive Officer Russ Tedrake, a professor at MIT for roughly two decades and former Vice President of Robotics Research at the Toyota Research Institute (TRI), helped lead influential efforts in the robotics field dating back to the 2015 DARPA Robotics Challenge. Much of Walden's core technical team previously spent a decade at TRI conducting fundamental research in dexterous manipulation and co-authored foundational physical AI frameworks, including Diffusion Policy in 2022, UMI (Universal Manipulation Interface), OpenVLA, and Large Behavior Models (LBMs).

Ambruș highlights their Science Robotics publication on Large Behavior Models (LBMs)—multitask, language-conditioned machine learning models designed to execute complex, autonomous behaviors on physical robots. LBMs demonstrate that as data volume and diversity scale, robots learn dexterous skills faster, require significantly fewer demonstration trajectories, and exhibit higher resilience to visual distractors and distribution shifts. Ambruș shares a benchmark lab demonstration where an LBM-driven robot cores an apple, cuts it into slices, dynamically adjusts to catch an escaping slice sliding off the cutting board, wipes down the knife blade, and safely returns it to its sheath. While visually striking, Ambruș cautions that lab demonstrations can be tuned arbitrarily to inflate performance metrics, reinforcing Walden's philosophy that true validation only happens in messy, unstructured field settings.

- **Research Heritage**: Leadership and engineering roots from MIT and Toyota Research Institute (TRI), with direct contributions to the DARPA Robotics Challenge (2015). - **Core Papers**: Breakthroughs in *Diffusion Policy* (2022), *UMI*, *OpenVLA*, and *Large Behavior Models* (Science Robotics, 2026). - **LBM Characteristics**: Multitask, language-conditioned manipulation models that improve sample efficiency, generalizability, and physical error recovery as training scales.

---

[07:21](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=441s) Hardware Evolution: From Walden 0 to Walden 1

To transition advanced manipulation models from the lab to enterprise applications, Walden is focusing initially on manufacturing and logistics workflows, supported by strategic investments and partnerships with industry giants including Toyota, Boeing, Samsung, NVIDIA, and CoreWeave. Walden's hardware philosophy prioritizes purpose-built utility over novelty: the mobile manipulators utilize a wheeled base rather than bipedal legs, combined with specialized gripper configurations optimized for factory and warehouse tasks.

The initial deployment platform, **Walden 0**, was built on a modified off-the-shelf commercial platform (the Rainbow Robotics RB-Y1), allowing the team to deploy into live facilities immediately while gathering baseline operational requirements. The lessons learned from Walden 0 directly shaped **Walden 1**, the company's first fully custom ground-up platform. Walden 1 introduces substantially higher payload capacities, expanded physical reach and workspaces, enhanced multi-camera perception suites, built-in industrial safety systems, hot-swappable batteries for continuous 24/7 facility shifts, and expanded onboard compute designed to execute larger, low-latency foundation models directly on the edge.

---

[09:32](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=572s) Physical AI Philosophy: Remote Assistance and Jidoka in the Field

Walden's operational paradigm centers on the idea that the single most valuable asset in physical AI is real-world field deployment data gathered under shared autonomy. Rather than waiting for a model to achieve 100% autonomy in isolation, robots are deployed immediately to live customer lines backed by remote human assistance. When a robot encounters an out-of-distribution state or physical obstruction, a human operator seamlessly intervenes via teleoperation to guide the system through the failure mode. This intervention ensures customer operations never stall while generating high-value "on-policy correction" data that captures precisely what the robot should do when faced with novel edge cases.

```mermaid flowchart TD accTitle: Jidoka-Inspired Shared Autonomy Data Flywheel subgraph DeploymentSite["Live Deployment Site"] Work["Autonomous Execution"] LineSignal["Line Signals & Triggers"] end subgraph SharedAutonomy["Remote Assistance (Jidoka)"] Anom["Anomaly / Edge Case"] Teleop["Human Operator Intervention"] end subgraph Flywheel["Continuous Learning Loop"] Data["On-Policy Correction Data"] Retrain["Model Retraining & Fine-Tuning"] Deploy["Over-the-Air Checkpoint Update"] end Work -->|Uncertainty / Failure| Anom LineSignal --> Anom Anom --> Teleop Teleop -->|Restores Line Flow| Work Teleop --> Data Data --> Retrain Retrain --> Deploy Deploy --> Work ```

This model is directly inspired by *jidoka*—the foundational Toyota Production System principle of "automation with a human touch," where machines are designed to detect anomalies, stop compounding errors, and request human guidance. It also mirrors the operational playbook of leading autonomous vehicle companies like Waymo, where remote operators oversee dozens of vehicles simultaneously. By deploying with remote assistance from day one, Walden achieves three compounding benefits: delivering immediate labor value to customers, establishing operational infrastructure for scaling large fleets, and harvesting the highest-quality demonstration and recovery data directly at the frontier of robot capability.

- **Jidoka Principle**: Machines autonomously detect faults, halt failure escalation, and request assistance, enabling a single operator to supervise an expanding fleet of robots. - **On-Policy Corrections**: Human interventions provide high-signal training demonstrations precisely at the edge of the model's failure boundaries. - **Three-Fold Compounding**: Rapid field deployment generates critical training data, builds fleet operations infrastructure, and delivers immediate economic ROI.

---

[12:49](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=769s) The Data Pyramid and Architectural Convergence

Ambruș conceptualizes physical AI training as a three-tiered data pyramid. The bottom base consists of massive, web-scale vision-language pre-training (VLMs). The middle layer is a mid-training stage where the model converts into a Vision-Language-Action (VLA) model by absorbing general robotic manipulation trajectories and egocentric human video datasets. The top tier—and the most decisive differentiator—consists of specialized, on-policy field deployment data gathered directly from Walden's physical platforms operating in targeted enterprise domains.

```mermaid flowchart TD accTitle: Physical AI Data Diet Pyramid Top["Top Tier: Proprietary Field Deployment Data (On-policy corrections & high-value site demonstrations)"] Mid["Mid-Training Tier: General Robotic Datasets (Cross-embodiment datasets, egocentric video, VLA fine-tuning)"] Base["Base Pre-Training Tier: Internet-Scale Vision & Language (Foundation VLMs, semantic reasoning, visual priors)"] Top --> Mid Mid --> Base ```

As deployment fleets expand, the compounding volume of high-quality field data will dominate model performance, progressively reducing reliance on generic mid-training datasets. Drawing a parallel to digital AI—where post-training reinforcement learning and synthetic data loops feed back into foundational pre-training—Ambruș anticipates physical AI deployment logs will recursively train next-generation base models.

Addressing the architectural rivalry between pure VLAs and predictive World Models, Ambruș characterizes the debate as a "false duality." The field is rapidly converging toward unified omni-models that are natively multimodal across inputs and outputs. Walden's current architecture conditions policies on image observation histories, kinematic proprioception chains, and language instructions to output high-frequency robot actions, with active expansion into video generation, tactile sensing, and audio modalities. Ambruș notes that model architectures are rapidly commoditizing; the true competitive moat lies in data provenance, curation, and the training curriculum.

---

[18:42](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1122s) Rapid Iteration Engine: Generative Simulation and Deployment Benchmarking

To maximize the return on field deployments, Walden built an engineering pipeline capable of transforming raw field logs into updated, validated model checkpoints in a matter of days. Extracting value from raw operational data relies heavily on sophisticated data engineering. Walden utilizes proprietary introspection interfaces that automatically segment and annotate multi-modal telemetry into hierarchical task descriptions with arbitrary granularity. This enables engineers to slice and filter data by motion smoothness, execution velocity, trajectory quality, and task progression metrics, creating curated data mixes tailored for mid-training and post-training runs.

```mermaid flowchart LR accTitle: Walden Physical AI Iteration and Evaluation Loop subgraph DataEng["Data Engineering & Curation"] Raw["Raw Field Telemetry"] Taxonomy["Hierarchical Annotation & Quality Slicing"] end subgraph SimEngine["Agentic Simulation"] GenAI["Generative Scene & Task Synthesis"] ScaleEval["High-Throughput Policy Evaluation"] end subgraph Release["Deployment Pipeline"] Bench["Pre-Deployment Lab Benchmarks"] Docker["Containerized Policy Deployment"] end Raw --> Taxonomy Taxonomy --> ScaleEval GenAI --> ScaleEval ScaleEval --> Bench Bench --> Docker ```

Complementing real-world logs is a large-scale simulation engine driven by agentic generative workflows. The system procedurally generates full architectural floor plans, diverse room geometries, realistic environmental clutter, and novel manipulation assets to test hypotheses and evaluate policies at scale without physical risk.

For reliable real-world rollouts, Walden utilizes a fully containerized, Docker-based workflow that isolates physical AI policies from the low-level machine runtime. Prior to shipping to customer sites, every candidate checkpoint undergoes rigorous pre-deployment benchmark evaluations across hundreds of hardware-in-the-loop lab trials, measuring autonomy rates, intervention frequencies, domain transfer robustness, and execution speeds. Crucially, Walden has established a strong positive correlation between lab benchmark metrics and real-world performance on industrial customer floors.

- **Data Engineering**: Dynamic slicing based on motion smoothness and task progress metrics, separating coarse trajectories from high-precision demonstrations. - **Agentic Simulation**: Generative procedural synthesis of complex 3D environments, tasks, and manipulation assets for scalable synthetic evaluation. - **Dockerized Policy Execution**: Containerized policy deployments decouple high-level AI models from core low-level robot control software. - **Predictive Lab Benchmarks**: Rigorous hardware-in-the-loop pre-flight evaluations demonstrate strong correlation with real-world production performance.

[20:21](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1221s) Live Deployment and Shared Autonomy in Action

Having a tight correlation between internal benchmarking and real-world execution changes the development cycle: the team can iterate through dozens of model variants in the lab, pick the single best checkpoint with high confidence, ship it into production, and watch it perform reliably on the floor. To demonstrate this in practice, the speaker presents video footage taken directly from an active industrial deployment where a robot is tasked with picking up unoriented objects and placing them onto a moving conveyor belt.

While the action might look straightforward to an outside observer, the speaker stresses that the physical nuances, subtle edge cases, and continuous environmental variations make the operation exceptionally difficult in practice. The presentation highlights a live instance of shared autonomy: when the robot encounters a tricky state, a remote human supervisor steps in momentarily via teleoperation, guides the robot through the bottleneck, and immediately hands control back over, allowing the robot to seamlessly resume executing the task on its own.

```mermaid flowchart LR accTitle: Shared Autonomy and Remote Intervention Loop subgraph Autonomous["Autonomous Execution"] A["Robot executes pick-and-place"] --> B{"Edge case / failure risk?"} B -- "No" --> A end subgraph Intervention["Shared Autonomy"] B -- "Yes" --> C["Remote operator steps in"] C --> D["Human clears bottleneck"] D --> A end ```

- Conveyor staging deployment: Robots manipulate varied, unstructured objects and place them onto moving conveyor belts in live operational settings. - Real-time intervention: Remote assistants intervene dynamically without taking the robot completely offline, ensuring uninterrupted task throughput while capturing valuable recovery data.

[21:03](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1263s) From Tools to Teammates: Co-Developing the Full Robotics Stack

In concluding the talk, the focus turns to Walden Robotics' foundational thesis: driving real-world deployments where robots perform genuinely useful physical labor powered by Large Behavior Models (LBMs) and supported by human-in-the-loop remote assistance. Rather than waiting for perfect autonomous agents before entering the field, getting robots into customer environments immediately serves a dual purpose—it delivers concrete commercial value on day one while capturing the highest-fidelity, most informative edge-case data available to train subsequent model checkpoints.

This deployment paradigm fundamentally alters the relationship between humans and automated machinery, shifting away from rigid, single-purpose industrial tools toward flexible, "capable assistants" that act as physical teammates. Achieving this level of fluid capability requires co-developing the entire vertical system in unison: tightly coupling the physical hardware, the underlying AI models, and the agentic software harness that orchestrates autonomy, monitoring, and operator handoffs so that the entire platform compounds in capability with every iteration.

- Core mission: Delivering real customer value from day one through Large Behavior Models (LBMs) backed by on-demand remote human assistance. - Full-stack co-development: Jointly designing hardware, the agentic harness, and AI architectures to eliminate integration bottlenecks and accelerate field iteration. - Paradigm shift: Moving from hardcoded, single-task machines toward adaptive physical assistants capable of operating dynamically alongside human workers.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=4s) on the verge of fundamentally impacting our lives — Rareș Ambruș

> Physical AI is opening a path to robots that behave like versatile agents in the physical world rather than fixed, single-purpose automation.

> [09:32](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=572s) automation with a human touch — Rareș Ambruș

> [09:32](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=572s) on-policy correction — Rareș Ambruș

> [12:49](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=769s) false duality — Rareș Ambruș

> [21:03](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1263s) capable assistants — Rareș Ambruș

> From Tools to Teammates: Physical AI and Shared Autonomy for Flexible Generalist Robots

> [03:15](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=195s) Diffusion Policy — Rareș Ambruș

## Mentioned Rareș Ambruș, Benjamin Burchfiel, Russ Tedrake, Walden Robotics, Ai4, Toyota Research Institute (TRI), MIT, DARPA Robotics Challenge, Rainbow Robotics, RB-Y1, Walden 0, Walden 1, Toyota, Boeing, Samsung, NVIDIA, CoreWeave, Waymo, Diffusion Policy, Universal Manipulation Interface (UMI), OpenVLA, Large Behavior Models (LBMs), Science Robotics, Docker

## Video description Benjamin Burchfiel - PlaceholderPlaceholder Rares Ambrus - Co-Founder & Head of AI, Walden Robotics

Physical AI is opening a path to robots that behave like versatile agents in the physical world rather than fixed, single-purpose automation. In this session we'll look at what it takes to carry a capable model from the lab onto a real factory floor, including the loop where field data, evaluations, and live deployments feed each next checkpoint, the role simulation plays in making that iteration fast, and how immersive shared autonomy lets robots take on real work from day one.

[Source](https://www.youtube.com/watch?v=hiUbHN7VNTA)

中文简报

Walden Robotics 在美国汽车装配线部署 AI 机器人车队将实体机器人直接部署到正在运转的工厂中并辅以远程人工支持，在提供即时劳动力价值的同时，收集训练自主基础模型所需的同策略失败数据。 Rareș Ambruș 介绍了 Walden Robotics 的全栈物理 AI 战略，该战略将定制的移动操作硬件与大型行为模型（LBM）相结合，以实现复杂制造和物流工作流的自动化。Walden 没有等待实现完美的独立自主，而是直接将机器人部署到实际生产线上，并在边缘情况下由远程人类主管进行干预；该做法受丰田生产方式的“自働化”（Jidoka）原则启发，能够产生高信号的同策略纠偏数据。该公司的硬件从原型 Walden 0 演进为专用的 Walden 1 平台，并由智能体生成式仿真和自动化数据策展循环提供支持，可迅速将现实世界的运行日志转化为经过验证的模型检查点。通过在汽车发动机装配中的成功生产部署，Walden 展示了共享自主与垂直整合的协同开发如何将机器人从僵化的工具转变为适应性强的实体协作伙伴。 Walden Robotics 优先考虑立即进行现实世界的工业部署，而非长期的实验室孵化；其在美国一条运转中的汽车发动机生产线上进行了数月部署且未发生停线，证明了这一点。 (0:04) 该公司立足于 MIT 和 TRI 校友取得的物理 AI 基础性突破，包括 Diffusion Policy、UMI、OpenVLA 和大型行为模型（LBM）。 (3:15) 大型行为模型（LBM）表明，扩大数据规模和多样性可显著提高机器人的样本效率、抗干扰能力以及动态物理错误恢复能力。 (3:15) 硬件从 Walden 0（用于快速收集数据的改装现成平台）演进为 Walden 1（配备热插拔电池、工业安全系统和高边缘算力的定制工业平台）。 (7:21) 受丰田自働化（jidoka）原则的启发，共享自主与远程协助实现了首日商业部署，同时在边缘情况干预期间收集高价值的同策略纠偏数据。 (9:32)

Reading summary

Walden Robotics argues that physical AI must move beyond isolated lab demonstrations into immediate industrial deployment, using vertically integrated systems and human-in-the-loop shared autonomy to keep factory lines moving while generating high-value operational training data. Transitioning from Lab Benchmarks to Live Manufacturing (0:04) Walden Robotics, led by CEO and former MIT professor Russ Tedrake alongside researchers from the Toyota Research Institute (TRI), contends that physical AI has reached a technological inflection point that requires rigorous field validation over lab benchmarks. While research roots in frameworks like Diffusion Policy, OpenVLA, and Large Behavior Models (LBMs) demonstrated complex dexterous manipulation in controlled environments—such as dynamic slicing and knife handling—speaker Rareș Ambruș cautions that lab environments can be tuned arbitrarily to inflate performance metrics. To prove real-world utility, Walden launched a pilot at a major US automotive plant two months after founding, followed by a continuous deployment between May and July where robots assisted in building thousands of engines without causing an assembly line stoppage. Shared Autonomy and the Jidoka Data Flywheel (9:32) Rather than delaying rollouts until models achieve total autonomy, Walden deploys robots alongside remote human assistance inspired by the Toyota Production System principle of jidoka—automation designed to detect anomalies and request human guidance. When a robot encounters an out-of-distribution state or bottleneck, such as placing unoriented objects onto a moving conveyor belt, a remote operator temporarily intervenes via teleoperation to clear the error before returning control to the autonomous policy. Ambruș notes this architecture mirrors autonomous vehicle fleet supervision like Waymo's. It provides immediate operational value to the facility while capturing critical 'on-policy correction' data that records exact recovery behaviors at the boundary of the model's capabilities. Iterative Hardware Engineering for Industrial Uptime (7:21) To withstand industrial manufacturing and logistics environments, Walden prioritizes purpose-built hardware over novelty, opting for wheeled dual-arm mobile manipulators rather than bipedal humanoid platforms. The company began field tests with Walden 0, a modified commercial Rainbow Robotics RB-Y1 platform used to rapidly collect baseline facility requirements. The operational lessons informed Walden 1, a fully custom platform engineered with higher payload capacities, an expanded kinematic reach, industrial safety systems, expanded edge AI compute for low-latency foundation models, and hot-swappable batteries designed for 24/7 continuous shift work. The hardware effort is supported by strategic partnerships and investments from industry firms including Toyota, Boeing, Samsung, NVIDIA, and CoreWeave. Data Hierarchy, Omni-Models, and Simulation Pipelines (12:49) Ambruș outlines a three-tier data pyramid for physical AI: broad vision-language model (VLM) pre-training at the base, general robotic trajectories and egocentric video mid-training in the center, and proprietary on-policy field logs at the top. Characterizing the debate between Vision-Language-Action (VLA) architectures and World Models as a false duality, Ambruș expects the field to converge on unified multimodal omni-models, making data quality and curation the primary competitive differentiator. To iterate on checkpoints within days, Walden pairs hierarchical telemetry annotation with agentic generative simulations that procedurally generate 3D floor plans and clutter. Candidate policies are evaluated in containerized, Docker-based lab benchmarks, which Walden reports correlate strongly with real-world customer floor performance. Walden Robotics positions its full-stack co-development strategy as a path to evolve automated machinery from single-purpose industrial tools into flexible, collaborative teammates. The model depends on maintaining a reliable pipeline where shared-autonomy interventions continuously expand autonomous reliability without burdening human supervisory bandwidth as robot fleets scale.

精炼阅读摘要

Walden Robotics 联合创始人 Rareș Ambruș 认为，物理人工智能（Physical AI）已跨越技术拐点，当前的核心任务并非追求孤立的实验室演示，而是通过软硬件全栈垂直整合与远程共享自主机制，将通用机器人直接部署至复杂工业现场并形成持续学习飞轮。学术积累与通用大行为模型 (3:15) Walden Robotics 的技术基石源于丰田研究院（TRI）和麻省理工学院（MIT）十余年的灵巧操作研究。其核心团队曾参与 2015 年 DARPA 机器人挑战赛，并主导或共同开创了 Diffusion Policy（2022年）、通用操作接口（UMI）、OpenVLA 以及大行为模型（Large Behavior Models, LBMs）。在发表于《Science Robotics》（2026年）的成果中，LBMs 展现为多任务、语言引导的机器学习架构，随着训练数据规模与多样性的扩展，机器人学习灵巧动作的样本效率大幅提升，所需示教轨迹显著减少，并对视觉干扰与分布偏移表现出更强的恢复弹性。 Ambruș 引用了实验室中机器人连续完成去核、切片、动态捕获滑脱苹果片、擦拭并收纳刀具的评测基准，但同时指出，实验室测试环境容易经过人为针对性微调而虚标表现。他强调，仅凭受控实验无法证明真实价值，算法真正的泛化能力与鲁棒性必须在非结构化且充满混乱的真实工业场景中接受检验。面向工业制造的专用硬件演进 (7:21) 为使操作模型顺利进入制造与仓储物流环境，Walden 获得了丰田、波音、三星、NVIDIA 和 CoreWeave 等产业资本的支持。团队在硬件形态上放弃了双足设计，选用兼具机动性与承载稳定的轮式底座及专用夹爪配置。为了在成立仅两个月内实现客户现场试运行，Walden 0 采用了改装的 Rainbow Robotics RB-Y1 商用平台，用于快速采集生产基线指标。基于早期现场暴露的工程限制，Walden 研发了首款完全自主设计的下一代量产平台 Walden 1。该机型显著提升了负载能力与机械臂运动工作空间，集成了多相机感知套件、工业级安全系统与支持 24 小时连续运行的热插拔电池，并扩充了机载边缘端算力，以便在本地直接运行低延迟的大规模基础模型。人机协同与自动化“自働化”机制 (9:32) 在生产运营中，Walden 借鉴丰田生产方式（TPS）中的“自働化”（Jidoka，带有人字旁的自动化）哲学，实施结合远程人工协助的共享自主系统。团队并不等待算法达到百分之百的理论自主，而是在机器人遭遇异常状态、物理阻碍或分布外边缘情况时，由机器主动暂停报错并请求人类干预。远程操作员介入接管控制消除瓶颈后，控制权即刻交还系统，保障装配流水线持续运转。 Ambruș 指出，该模式借鉴了 Waymo 等自动驾驶公司的车队监管逻辑。它不仅消除了单点故障导致的停线风险，还产出了极高价值的“策略内纠偏”（on-policy correction）轨迹数据，精确记录模型在失效边界的应对行为。在 5 月至 7 月期间，该系统在美国大型汽车制造厂全职运行于活跃生产线，协助组装了数千台汽车发动机，实现了零停线的连续作业记录。数据金字塔与多模态端到端架构 (12:49) Ambruș 将具身智能的数据架构归纳为三层金字塔：底层是互联网规模的视觉-语言预训练模型（VLMs），中层是通过跨本体通用机器人轨迹与第一人称人类视频微调的视觉-语言-动作模型（VLA），顶层则是从实际企业部署中收集的专有现场策略数据。Ambruș 预计，随着部署规模扩大，高价值工业数据将在训练中占主导地位，逐步降低对通用中层数据的依赖，并通过持续回流推动基础模型的递归迭代。针对学术界关于纯 VLA 策略与预测性世界模型（World Models）的路线分歧，Ambruș 称其为“虚假二元对立”。他指出物理 AI 架构正在向原生统一的多模态全功能模型汇聚。Walden 当前的系统整合了历史图像输入、运动学本体感受序列和语言指令以输出高频动作，并逐步融合触觉、音频与视频生成模态。他认为模型结构已逐步商品化，企业的核心技术护城河在于专有数据源、精细化标注清洗机制与训练课程设计。工程闭环与容器化快速迭代流水线 (18:42) 为缩短算法验证周期，Walden 搭建了在数日内将遥测数据转化为发布检查点的工程闭环。系统通过专有接口对多模态传感器日志进行分层任务标注，根据动作平滑度、运行速度和任务完成度指标进行数据切片与过滤，为模型微调提供高质量数据集。在合成数据层面，团队引入智能体式生成仿真引擎，程序化合成复杂的三维车间布局、杂乱障碍与操作对象，在无物理风险的前提下进行高通量策略评估。在发布运维阶段，Walden 采用 Docker 容器化方案将高层 AI 策略与底层机器人硬件运行时彻底解耦。所有模型在推送到客户工厂前，均需在硬件在环实验室基准中经历数百次自动化测试，评估其自主率、干预频率和泛化鲁棒性。实测表明，实验室内测指标与汽车厂流水线传送带抓取等工序的实际表现具有强正相关性，确保了算法迭代的稳定交付。 Walden Robotics 倡导通过全栈软硬件共研与即刻投入真实产线，将机器人从刻板的单任务工业工具转变为可协作的自主助手。然而，其长期经济性与技术泛化壁垒，仍取决于车队规模扩张时远程操作员接管比例能否持续下降，以及特定制造业现场采集的高成本数据能否低成本迁移至更多通用物流与制造场景。

Detailed notes — zh

## Background 该视频由 Ai4 频道于 2026 年 8 月 15 日发布在 YouTube 平台上，记录了 Walden Robotics 联合创始人兼 AI 负责人 Rares Ambrus 的主题演讲。根据页面简介，内容聚焦于物理 AI（Physical AI）如何赋能通用机器人走出实验室并入驻真实工厂车间，重点探讨现场数据飞轮、仿真加速以及沉浸式共享自主如何让机器人从第一天起就承担实际生产任务。 ## TL;DR Walden Robotics 在结束隐身模式之际，系统阐述了其以全栈垂直整合范式推动通用机器人落地工业现场的最新进展与技术架构。依托团队在扩散策略与大行为模型（LBM）上的深厚积淀，公司自研了专为端到端具身智能打造的 Walden One 轮式双臂平台，并在美国顶尖汽车制造厂达成了连续数月全职组装发动机且“零停线”的工业级验证。针对非结构化环境中的分布外工况，团队引入融合丰田自働化（Jidoka）哲学的共享自主机制，由远程专家即时介入兜底并实时沉淀高价值策略内纠偏数据。配合分层数据金字塔、全模态模型演进与智能体驱动的生成式仿真评测闭环，团队正加速推动机器人从被动自动化工具蜕变为能够自主协同的物理 AI 队友。 ## Atmosphere & style 整场演讲洋溢着兼具顶级学术底蕴与严苛工业实战经验的沉稳与自信，风格务实、硬核且极具说服力。演讲者以风趣的实验室切苹果轶事开篇，随后迅速切入高节拍汽车装配产线的严酷现实，叙事节奏张弛有度、层层递进。现场展示剥离了纯概念式的技术炒作，将镜头聚焦于真实的边缘接管、底盘架构取舍与全栈工程落地细节，给观众带来一种具身智能正在物理世界工业硬地表扎实生根的震撼与确定感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=4s) Walden Robotics 走出隐身模式，以自研硬件、底层系统、基础模型与顶层应用的全栈垂直整合架构，推进通用机器人具身智能落地。 - [00:04](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=4s) 公司成立仅两个月便进驻全美顶级汽车制造厂，连续数月全天候制造数千台发动机，达成工业级零停线（Zero Downtime）记录。 - [03:15](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=195s) 团队拥有深厚的 MIT 与丰田研究院（TRI）技术底蕴，基于扩散策略与《Science Robotics》发表的大行为模型（LBM），实现具备强抗扰与泛化能力的灵巧操作。 - [07:55](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=475s) 硬件演进上从基于商用本体改装的 Walden Zero 过渡到自研量产平台 Walden One，放弃双足改用轮式双臂，大幅强化负载、作业空间与机载边缘算力。 - [09:32](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=572s) 采用借鉴丰田 TPS 自働化（Jidoka）哲学的共享自主与远程辅助（Shared Autonomy）机制，遇异常即停并呼叫人工，实现不停机交付生产力。 - [09:32](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=572s) 远程人工介入接管实时产生高价值的“策略内纠偏（On-Policy Correction）”数据，直接揭示模型认知盲区，驱动模型自进化飞轮。 - [12:49](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=769s) 物理 AI 数据金字塔由底层 VLM 常识、中层多机器人与第一人称视频、顶层高壁垒部署现场轨迹构成，部署数据闭环周期被压缩至数天以内。 - [15:33](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=933s) 自研深度内省分析界面，从语义目标（What）与运控质量（How）双维度精细化切片清洗现场数据，以高质量数据流替代单一网络架构竞争。 - [15:33](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=933s) 破除 VLA 与世界模型的虚假对立，构建融合图像、本体感觉与文本输入的原生全模态架构，并拓展触觉与音频等信号。 - [18:42](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1122s) 利用基于智能体工作流的生成式仿真系统进行超大规模合成数据采样，同时采用 Docker 容器化架构实现策略与运控解耦的热更新。 - [18:42](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1122s) 建立高保真预部署基准评测集，使实验室测试结果与真实工厂作业表现高度正相关，保障模型上线的确定性信心。 - [21:03](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1263s) 机器人正在从被动的单用途自动化工具转变为自主协同的物理 AI 助手，依托硬件、Agent Harness 与模型的深度全栈共研持续迭代。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=4s) 破茧而出：Walden Robotics 的具身智能使命与工业落地突破 Walden Robotics 联合创始人兼 AI 负责人 Rares Ambrus 登台分享了团队在将通用机器人引入真实物理世界方面的最新进展。公司刚刚结束隐身模式（Stealth Mode），正以全栈垂直整合的具身智能（Physical AI）公司姿态推向市场——涵盖自研硬件底座、系统软件、前沿具身基础模型以及顶层行业应用。在 Ambrus 看来，由物理 AI 驱动的通用机器人已经迎来技术拐点，正处于深刻重塑人类生产生活的边缘。然而，要真正跨越从实验室到工业现场的鸿沟，企业必须直面真实世界的考验：不仅要在非结构化环境中证明系统的稳定性，还要算清商业化落地的单机经济模型（Unit Economics），并与一线客户建立深度的工程互动。 Walden Robotics 从成立之初便确立了“快速实战部署”的核心策略。公司成立仅两个月（2026年3月），便将机器人送入美国最大的汽车制造工厂之一展开实地试点。该试点取得了令人瞩目的成效：在随后的5月至7月间，机器人正式进驻产线进行全天候全职作业。在此期间，这批机器人直接参与了数千台汽车发动机的装配生产，不仅完全跟上了产线节拍，而且创造了“从未拖慢产线、从未导致停线”的优异记录，验证了通用机器人在高节拍严苛工业场景下的可用性。 - **全栈整合定位**：同时掌控硬件设计、底层系统软件、前沿物理 AI 模型及垂直应用开发。 - **产线实战验证**：成立2个月即开启全美顶级汽车制造厂试点；连续数月全职上岗制造数千台发动机，达成零停线（Zero Downtime）工业级标准。 - **视频演示**：现场展示了机器人最新能力片段，并公布了下一代硬件平台的研发轮廓。 --- ### [03:15](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=195s) 从实验室到前沿研究：核心团队积淀与大行为模型（LBM） Walden Robotics 拥有深厚的技术积淀。首席执行官 Russ Tedrake 是麻省理工学院（MIT）任教逾20年的资深教授，曾长期担任丰田研究院（TRI）机器人工程与研究副总裁，早在2015年就带领团队在 DARPA 机器人挑战赛（DRC）中确立了行业影响力。核心研发团队的多位骨干同样来自 TRI 及自动驾驶前沿领域，在灵巧操作（Dexterous Manipulation）与机器学习基础研究方面深耕十余年，主导或深度参与了扩散策略（Diffusion Policy）、UMI、OpenVLA 以及大行为模型（Large Behavior Models, LBM）等里程碑式研究。在回顾技术演化时，Ambrus 重点提到了发表于《Science Robotics》的 LBM 研究。该类多任务、以语言为条件的大模型能够自主执行复杂的物理交互动作，并展现出卓越的样本效率、抗干扰能力与分布偏移适应性。他在现场展示了一段经典的实验室“切苹果”长序列任务：机器人自主完成去核、对半切开、细分切片，并在果块即将滑落砧板的瞬间实现自主纠偏将其推回，最后完成刀具擦拭与归位。Ambrus 风趣地调侃自己在测试期间“吃了海量的苹果”，但也清醒地指出：实验室里惊艳的 Demo 往往可以通过微调任务条件来人为拔高表现，而 Walden 的真正目标是走出温室，让模型在真实的工业土壤中扎根。 - **核心团队基因**：汇聚 MIT 机器人专家、DARPA 挑战赛老将及 TRI 灵巧操作核心团队，兼具前沿学术突破与自动驾驶级海量数据管线量产工程经验。 - **理论突破基石**：以2022年兴起的扩散策略及2026年《Science Robotics》发表的大行为模型（LBM）为基础，实现高泛化、强鲁棒的物理世界技能学习。 - **生态与资本支持**：获得丰田（Toyota）、波音（Boeing）、三星（Samsung）、英伟达（NVIDIA）、CoreWeave 等战略投资者与行业巨头的支持，首期聚焦制造业与现代物流。 --- ### [07:55](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=475s) 硬件演进：从 Walden Zero 到专为具身智能打造的 Walden One 在硬件路线的选择上，Walden 团队采取了务实而敏捷的演进路径。第一代平台 Walden Zero 基于 Rainbow Robotics 的 RB-Y1 双臂轮式机器人底座打造。团队选择商用成熟本体的初衷是为了以最快速度进入生产现场收集真实数据，并在现场部署过程中针对工业严苛环境进行了大幅度的软硬件改装。这些一线实战经验被全量沉淀并反哺到了全新自研的 Walden One 平台设计中。 Walden One 放弃了双足设计，采用了高稳定性的轮式移动底盘与专属定制的双臂夹爪构型。这一形态直接来源于对目标市场客户需求的真实体察。整机完全围绕端到端具身智能（End-to-End Autonomy）的严苛需求从零设计：大幅提升了有效负载（Payload）、作业空间包络（Workspace）与整机感知能力，配备了支持全天候连续不间断作业的热插拔电池系统，并集成了工业级主动安全系统。最令 AI 团队振奋的是，Walden One 配备了极其强劲的机载边缘计算单元，为在本地高速运行前沿大模型扫清了算力瓶颈。 | 对比维度 | 第一代过渡平台 (Walden Zero) | 下一代量产平台 (Walden One) | |:--- |:--- |:--- | | **研发基础** | 基于 Rainbow Robotics RB-Y1 改装 | 自研定制端到端具身智能平台 | | **移动形态** | 轮式双臂底座 | 工业优化轮式移动底座，无腿部复杂冗余 | | **作业负载与空间** | 标准商用水平 | 显著增强的有效负载与扩展操作空间 | | **算力与供电** | 基础机载算力，常规充电 | 强大机载边缘计算，支持电池热插拔连续作业 | | **安全与感知** | 实验室级防护改装 | 工业级前置增强感知与多重安全冗余机制 | --- ### [09:32](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=572s) 人机协同自律哲学：远程辅助、Jidoka 原则与数据飞轮针对如何让物理 AI 在成熟度尚不完美时即刻创造商业价值，Walden Robotics 给出的解法是“共享自主（Shared Autonomy）与远程辅助（Remote Assistance）”。机器人进入客户产线第一天起即自主执行作业，当遭遇未见工况或置信度不足的边缘场景时，系统会主动呼叫远程操作员接入。人类专家介入接管不仅确保了客户产线的绝对连续性，更在现场实时产出了极高价值的“策略内纠偏（On-Policy Correction）”监督数据。随着模型在此类数据上的持续微调，单名操作员所能管理的机器人集群规模呈指数级提升。 Ambrus 强调，这一设计哲学深度契合丰田生产方式（TPS）的核心支柱——“自働化（Jidoka，带有人字旁的自动化，即 Automation with a human touch）”。其核心在于赋予机器自主识别缺陷与异常的能力，遇险即停并主动求助，坚决防止缺陷向下游工序传递。同时，这也借鉴了头部自动驾驶公司（如 Waymo）成熟的远程车队运营模式。通过将远程辅助视为主力系统架构中的“一等公民”，企业实现了三重正向复合飞轮：采集最高价值的真实边界数据、构建庞大机群的运营调度能力，并在第一天就为客户交付切实的生产力。 ```mermaid flowchart TD accTitle: Shared Autonomy and Data Flywheel Engine accDescr: Diagram illustrating the closed-loop system connecting autonomous factory deployment, remote human assistance, and continuous model improvement. subgraph Factory ["客户工业产线 / 现场部署"] R["Walden 机器人集群 (自主执行装配/搬运)"] Edge["遭遇异常或分布外工况 (OOD)"] R -->|"识别异常"| Edge end subgraph Teleop ["自働化 (Jidoka) 远程辅助中心"] H["远程专家无缝介入接管 (1:N 弹性监控调度)"] Edge -->|"主动求助"| H H -->|"恢复作业"| R end subgraph Flywheel ["物理 AI 数据飞轮与模型演进"] D["高价值策略内纠偏数据 (On-Policy Corrections)"] M["大行为模型 (LBM / VLA) 自进化与版本迭代"] H -.->|"沉淀真实边界数据"| D D -->|"模型微调与预训练"| M M -->|"新权重下发机载算力"| R end ``` - **Jidoka 制造哲学赋能**：机器自主捕获异常并暂停，远程人工介入消除隐患，杜绝产线停机与不良品产生。 - **高密度价值数据**：现场干预数据直接揭示了模型当前能力的“认知盲区”，构成最优质的强化与微调素材。 - **三维复合价值增长**：产线连续交付客户价值、团队打磨大规模机群运营体系、沉淀不可替代的专有场景数据资产。 --- ### [12:49](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=769s) 数据金字塔与自进化闭环架构物理 AI 模型的训练依赖于层次分明的“数据金字塔”结构。金字塔底层是预训练阶段（Pre-training），主要依托海量视觉-语言模型（VLM）吸收通识语义与常识；中层为中训练阶段（Mid-training），通过引入通用机器人跨平台数据与第一人称视角的以自我为中心（Egocentric）视频，将通用 VLM 转化为具备空间感知与物理动作雏形的视觉-语言-动作模型（VLA）；而金字塔顶端最昂贵、最具决定性壁垒的数据，则是直接来自 Walden 专属机器人在真实业务现场积累的交互轨迹。 Ambrus 预判，如同数字大模型（Digital AI）正在经历合成数据与后训练（Post-training）自我提升的循环一样，物理 AI 也将迅速重构传统的数据范式。未来，来自昨日工业部署现场的高质量物理数据将向下反哺整个基础模型的预训练与中训练阶段，而通用的跨机型中训练数据权重将逐步被自有场景数据取代。为了抓住这一趋势，Walden 构建了一套从产线数据捕获、自动清洗、仿真回归到权重下发的闭环自进化系统，将现场反馈转化为模型迭代的周期压缩至短短数天以内。 - **三层数据金字塔**： 1. **顶层（高壁垒）**：自有本体在部署现场产出的高精度、长序列真实工况与纠偏数据； 2. **中层（能力转化）**：通用多机器人具身数据集与以自我为中心的视频轨迹； 3. **底层（常识认知）**：海量互联网多模态视觉与语言预训练权重。 - **数据回流范式**：现场高价值轨迹反哺预训练与中训练，推动物理 AI 基础模型发生代际演进。 - **敏捷闭环指标**：将“现场发现问题 $\rightarrow$ 数据入库清洗 $\rightarrow$ 模型更新 $\rightarrow$ 回测上线”的全链路时间缩短至数天级别。 --- ### [15:33](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=933s) 数据工程精细化与统一多模态模型架构海量部署数据并不能简单地直接倾倒进训练集群，深度的“数据工程（Data Engineering）”才是让模型在工业现场稳定工作的核心底牌。Walden 自研了深度内省分析界面（Introspection Interface），从“做什么（What）”与“怎么做（How）”两个核心维度解构现场数据：前者通过分层任务描述体系在任意时间粒度上标注动作语义；后者则精细化量化动作完成度、作业节拍、运控平滑度及轨迹质量。算法工程师可以根据不同训练阶段的诉求进行灵活的数据切片（Data Slicing），例如在特定训练期过滤掉抖动数据，或提取极端工况下的高质量恢复动作。 ``` 原始部署轨迹数据 ───► [分层任务语义标注 (What)] ───► [运动质量/平滑度评估 (How)] ───► [多维数据切片与过滤] ───► 课程学习与分阶段训练 ``` 在模型架构路线上，针对业界关于“视觉-语言-动作模型（VLA）与世界模型（World Models）孰优孰劣”的激烈争论，Ambrus 直言这是一种“虚假对立（False Duality）”。行业大趋势必然走向在输入端与输出端均实现原生全模态支持的统一模型（Omni-models）。Walden 当前的模型架构以包含历史记忆的图像帧、机器人本体运动学状态（Proprioception）及自然语言作为联合条件输入，实时输出精准机械控制动作，并正积极拓展视频生成预测、触觉（Tactile）与音频等多模态信号。他指出，基础神经网络架构正在迅速通用化与商品化，物理 AI 真正的胜负手在于高质量数据在模型中的流转方式与课程学习（Curriculum）设计。 - **多维内省数据切片**：结合层次化语义标签与底层物理运控质量（速度、加速度、平滑性）进行动态筛选，精准适配预训练与后训练。 - **超越架构之争**：破除 VLA 与世界模型的对立思维，打造整合历史视觉、运动学本体感受、文本指令乃至触觉听觉的原生全模态架构。 - **差异化核心**：模型网络结构趋向同质化，“高质量现场数据流 + 科学的训练课程策略”构成核心护城河。 --- ### [18:42](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1122s) 仿真加速与确定性部署管线仿真系统构成了 Walden 研发与交付体系的另一大核心支柱。为了摆脱物理世界测试吞吐量的限制，团队构建了基于智能体工作流（Agentic Workflow）的生成式仿真系统。该系统能够全自动生成结构各异的厂房环境、摆放多样化的工业资产零部件并派发不同复杂度的任务，从而大规模开展自动化数据合成、策略演练与自动化评测。在部署工程链路方面，Walden 采用了基于 Docker 容器化的隔离架构，使策略模型与底层的机器人硬件抽象、现场工业通信网络彻底解耦，实现了策略权重的无缝热更新。在策略推向客户产线之前，必须通过实验室内部严格的预部署基准评测集（Pre-deployment Benchmark Suite），全方位测量自主率、干预率、分布迁移适应性及运行节拍。Ambrus 欣慰地指出，经过长期调优，实验室基准测试的表现与客户真实工厂内的现场表现之间已经建立了极其强劲的正相关性，使得研发团队在每一次模型推上线时都具备确定性的质量信心。 - **智能体驱动的生成式仿真**：自动化生成逼真工业场景、物理资产与长序列任务，实现超大规模合成数据采样与边界压力测试。 - **容器化策略解耦**：利用 Docker 隔离 AI 策略与底层运控通信，确保模型秒级热更且不影响工业系统的底层安全性。 - **高保真评估对齐**：构建了涵盖自主率、接管频率、动作速度等指标的预部署实验室评测体系，与真实客户场景表现高度正相关。 ### [20:21](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1221s) 真实场景部署演示与人机协作干预在将模型正式部署到真实环境前，团队能够先在实验室内部快速迭代并评估大量的候选变体，最终筛选出表现最优的检查点（checkpoint）进行交付。演讲者展示了一段来自实际部署现场的运行视频：机器人正在执行一项高难度任务，需要精准抓取特定物体并稳定放置到传送带上。尽管这类任务在物理层面的微妙差异与操作难度极难仅凭肉眼完全体会，但其复杂度和对操作精度的要求极高。视频中特别展示了一个典型的远程干预（Remote Intervention）实例：当机器人遇到棘手状态时，远程操作人员能够即时切入并协助机器人化解卡点；一旦脱离困境，系统便立即无缝切回，由机器人继续自主执行后续流程。这种共享自主（Shared Autonomy）机制确保了产线任务的连续性与高可靠性。 ```mermaid flowchart LR accTitle: 远程协作与自主执行闭环架构 subgraph Lab["实验室研发与迭代"] A["多模型变体评估"] --> B["选定最优 Checkpoint"] end subgraph Deployment["现场部署与运行"] B --> C["自主执行作业 (抓取并放置至传送带)"] C -->|遇到极端工况| D["远程专家即时介入"] D -->|问题解决后交还控制权| C end subgraph Flywheel["数据回流"] C -.->|现场高价值数据| A D -.->|干预示范数据| A end ``` ### [21:03](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1263s) 总结与展望：从被动工具演进为全栈共研的物理 AI 助手演讲最后总结了 Walden Robotics 的核心战略与愿景：团队始终高度聚焦于真实世界的现场部署，让机器人依靠大行为模型（LBMs, Large Behavior Models）并辅以远程协助来完成真正创造商业价值的实际工作。团队坚信，只有深入一线真实场景进行部署，才能源源不断地捕获最高质量的实际工况数据，并第一时间为客户创造实际效益。更为关键的是理念上的跃升——团队正在将机器人从单纯的被动自动化“工具”（mere tools），转变为能够自主协同的“得力助手与队友”（capable assistants）。为了实现这一目标，团队正在对整个系统实施全栈协同研发（Co-development），将底层硬件、智能体控制框架（Agentic Harness）与物理 AI 模型深度打通，从而形成飞轮效应，持续推动机器人物理能力的快速迭代与进化。 - **核心驱动架构**：以大行为模型（LBMs）为自主核心，配合远程人机协同（Remote Assistance）兜底，保障从第一天起即可上岗作业。 - **数据价值飞轮**：真实场景不仅是检验模型能力的终极靶场，也是反哺模型迭代的最关键高价值数据来源。 - **全栈系统协同**：硬件架构、Agent 调度框架（Agentic Harness）与 AI 模型深度共研，全面推动物理智能体（Physical AI）的演进。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=4s) 由物理 AI 驱动的通用机器人已经迎来技术拐点，正处于深刻重塑人类生产生活的边缘。 — Rares Ambrus > [00:04](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=4s) 在这期间，这批机器人直接参与了数千台汽车发动机的装配生产，不仅完全跟上了产线节拍，而且创造了‘从未拖慢产线、从未导致停线’的优异记录。 — Rares Ambrus > [03:15](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=195s) 实验室里惊艳的 Demo 往往可以通过微调任务条件来人为拔高表现，而 Walden 的真正目标是走出温室，让模型在真实的工业土壤中扎根。 — Rares Ambrus > [09:32](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=572s) 机器人进入客户产线第一天起即自主执行作业，当遭遇未见工况或置信度不足的边缘场景时，系统会主动呼叫远程操作员接入。 — Rares Ambrus > [09:32](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=572s) 自働化（Jidoka，带有人字旁的自动化，即 Automation with a human touch）的核心在于赋予机器自主识别缺陷与异常的能力，遇险即停并主动求助，坚决防止缺陷向下游工序传递。 — Rares Ambrus > [12:49](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=769s) 未来，来自昨日工业部署现场的高质量物理数据将向下反哺整个基础模型的预训练与中训练阶段，而通用的跨机型中训练数据权重将逐步被自有场景数据取代。 — Rares Ambrus > [15:33](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=933s) 视觉-语言-动作模型（VLA）与世界模型（World Models）孰优孰劣的争论是一种‘虚假对立’，行业大趋势必然走向在输入端与输出端均实现原生全模态支持的统一模型。 — Rares Ambrus > [15:33](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=933s) 基础神经网络架构正在迅速通用化与商品化，物理 AI 真正的胜负手在于高质量数据在模型中的流转方式与课程学习设计。 — Rares Ambrus > [18:42](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1122s) 实验室基准测试的表现与客户真实工厂内的现场表现之间已经建立了极其强劲的正相关性，使得研发团队在每一次模型推上线时都具备确定性的质量信心。 — Rares Ambrus > [21:03](https://www.youtube.com/watch?v=hiUbHN7VNTA&t=1263s) 团队正在将机器人从单纯的被动自动化‘工具’，转变为能够自主协同的‘得力助手与队友’。 — Rares Ambrus ## Mentioned Walden Robotics, Rares Ambrus, Russ Tedrake, Benjamin Burchfiel, Toyota Research Institute (TRI), Massachusetts Institute of Technology (MIT), DARPA Robotics Challenge (DRC), Science Robotics, Rainbow Robotics, RB-Y1, Walden Zero, Walden One, Toyota, Boeing, Samsung, NVIDIA, CoreWeave, Waymo, Diffusion Policy, UMI, OpenVLA, Large Behavior Models (LBM), Vision-Language-Action (VLA), Vision-Language Models (VLM), Docker ## Video description Benjamin Burchfiel - PlaceholderPlaceholder Rares Ambrus - Co-Founder & Head of AI, Walden Robotics Physical AI is opening a path to robots that behave like versatile agents in the physical world rather than fixed, single-purpose automation. In this session we'll look at what it takes to carry a capable model from the lab onto a real factory floor, including the loop where field data, evaluations, and live deployments feed each next checkpoint, the role simulation plays in making that iteration fast, and how immersive shared autonomy lets robots take on real work from day one. Physical AI is opening a path to robots that behave like versatile agents in the physical world rather than fixed, single-purpose automation. In this session we'll look at what it takes to carry a capable model from the lab onto a real factory floor, including the loop where field data, evaluations, and live deployments feed each next checkpoint, the role simulation plays in making that iteration fast, and how immersive shared autonomy lets robots take on real work from day one. [Source](https://www.youtube.com/watch?v=hiUbHN7VNTA)

Source quotes

[0:04] Rares Ambrus: 由物理 AI 驱动的通用机器人已经迎来技术拐点，正处于深刻重塑人类生产生活的边缘。 [0:04] Rares Ambrus: 在这期间，这批机器人直接参与了数千台汽车发动机的装配生产，不仅完全跟上了产线节拍，而且创造了‘从未拖慢产线、从未导致停线’的优异记录。 [3:15] Rares Ambrus: 实验室里惊艳的 Demo 往往可以通过微调任务条件来人为拔高表现，而 Walden 的真正目标是走出温室，让模型在真实的工业土壤中扎根。 [9:32] Rares Ambrus: 机器人进入客户产线第一天起即自主执行作业，当遭遇未见工况或置信度不足的边缘场景时，系统会主动呼叫远程操作员接入。 [9:32] Rares Ambrus: 自働化（Jidoka，带有人字旁的自动化，即 Automation with a human touch）的核心在于赋予机器自主识别缺陷与异常的能力，遇险即停并主动求助，坚决防止缺陷向下游工序传递。 [12:49] Rares Ambrus: 未来，来自昨日工业部署现场的高质量物理数据将向下反哺整个基础模型的预训练与中训练阶段，而通用的跨机型中训练数据权重将逐步被自有场景数据取代。 [15:33] Rares Ambrus: 视觉-语言-动作模型（VLA）与世界模型（World Models）孰优孰劣的争论是一种‘虚假对立’，行业大趋势必然走向在输入端与输出端均实现原生全模态支持的统一模型。 [15:33] Rares Ambrus: 基础神经网络架构正在迅速通用化与商品化，物理 AI 真正的胜负手在于高质量数据在模型中的流转方式与课程学习设计。 [18:42] Rares Ambrus: 实验室基准测试的表现与客户真实工厂内的现场表现之间已经建立了极其强劲的正相关性，使得研发团队在每一次模型推上线时都具备确定性的质量信心。 [21:03] Rares Ambrus: 团队正在将机器人从单纯的被动自动化‘工具’，转变为能够自主协同的‘得力助手与队友’。

Source quotes

[0:04] Rareș Ambruș: on the verge of fundamentally impacting our lives Physical AI is opening a path to robots that behave like versatile agents in the physical world rather than fixed, single-purpose automation. [9:32] Rareș Ambruș: automation with a human touch [9:32] Rareș Ambruș: on-policy correction [12:49] Rareș Ambruș: false duality [21:03] Rareș Ambruș: capable assistants From Tools to Teammates: Physical AI and Shared Autonomy for Flexible Generalist Robots [3:15] Rareș Ambruș: Diffusion Policy

Original transcript

[0:04] Hi everyone. Um my name is Rahes Ambush. I'm super excited to be here. Um I'm a co-founder and head [0:10] of AI at Walden Robotics. And um um I would love to tell you a little bit about us. Um who we are [0:16] and our approach to putting general purpose robots to work in um in the real world. [0:27] That's not that's not coming from my laptop. Um so very exciting time for us. We were in stealth. We [0:34] just came out of stealth um a couple of weeks ago and um a little bit about us. We're a vertically [0:41] integrated end-to-end uh physical AI company. We build the hardware, we build the software, [0:48] we build frontier class physical AI and the application layer. And we believe that [0:54] these general purpose robots which are powered by physical AI are basically on the verge of um [1:02] fundamentally impacting our lives and it feels like this technology has crossed an inflection [1:07] point but there's still a lot of things left to do. We have to prove that this technology works um [1:13] in the real world so not just in the lab. We have to build a business and understand the [1:17] unit economics of the business. uh we have to um understand our customers and deeply interact [1:24] with them. And as a company, our focus so far has been on uh deploying very fast, [1:30] putting our robots out there in the real world and having them do uh useful work and valuable work. [1:36] And so um 2 months after we started the company, we had our first pilot at one of the largest [1:43] automotive manufacturing plants in the US. This was in March. Uh the pilot went super well. So we [1:49] went back a couple of months later and we had the robots operating full-time in the production line. [1:55] Uh and I'm super happy to say that between um May and July this year, the robots contributed [2:01] to manufacturing thousands of engines and um they never slow down the line and they never u [2:09] cause the line to stop which is awesome and we are super uh happy and proud of being able to [2:14] deploy so quickly and be successful in this in this uh deployment. It's a big milestone for us. [2:21] Uh and here's a small snippet of what our robot can do. It's from the launch video. [2:25] It was posted online a couple of weeks ago just to give you an idea of the robots [2:29] um the capabilities uh the form factor the applications and so on. [3:08] and this is a small snapshot of the next generation platform that we're uh building. [3:15] Okay, awesome. So, I want to take a minute and just tell you a little bit about the [3:19] Wald and founding team and our expertise. So we've been working on robotics and AI for um [3:25] for a long time now and I only I can only show a few sort of snippets of our work. [3:30] Um so our CEO Russ Tedric he is an MIT professor for 20 years or so. He was a VP of [3:37] uh robotics engineering robotics research at Toyota research institute. Um and he's [3:42] been very influential in the field. One of the first uh one of the things to call out is his [3:46] contributions to the DARPA robotics challenge. So this is goes all the way back to 2015. [3:52] I had a video here, but uh in the interest of time, I'm not going to show it. U it's pretty [3:56] impressive what these robots could do all the way back in 2015. Um a few other members of the [4:01] core team came from Toyota Research Institute. Uh we spent the last 10 years or so uh doing [4:06] fundamental research on Dexter's manipulation. Um, we co-authored some influential papers that I'm [4:12] sure some of you uh know like diffusion policy, UMI, open VLA, large behavior models or LBN [4:20] and um these are some stills of uh the robot doing some of these dextrous tasks and they were [4:26] made possible by uh this discovery of diffusion policy which was around 2022 a new technique that [4:33] allowed us to basically uh program more and more complex and dextrous behaviors in these robots. uh [4:40] my own background and the background of others from the team is in machine learning. So [4:43] fundamental research in machine learning for a wide range of applications uh and domains like [4:48] robotics, automated driving, uh perception and so on. And a few of us also have worked [4:54] in a autonomous driving. So shipping models into production, building uh robust and scalable data [5:00] pipelines and uh the stuff that we're still doing today, but in a slightly different uh [5:06] fashion. Uh so one last result to call out is our work on large behavior models. Uh this was [5:12] published in 2026 so earlier this year in science robotics. Uh what we wanted to do here is really [5:17] understand um how these how these models work quantify their performance in a wide range of [5:22] settings. For context LBMs are multitask language conditioned uh ML models that can execute complex [5:30] behaviors autonomously on robots. And so through this work we we could we we could understand and [5:35] conclude that when we have these LBMs and we know how to use them we can learn new skills [5:40] dextrous and complex skills on robots u [snorts] uh faster with less data. We can be more robust to [5:47] distractors and distribution shift and things like that. And the performance of these LBMs increases [5:52] as we scale data and diversity. And I want to show you here one uh one of my favorite tasks um [5:59] that we basically programmed on on this on this robot. It basically the robot has to cord this [6:04] apple, slice it in half, slice the halves into sl into different pieces uh and then put the knife [6:09] away. And just to give you an idea of the level of dexterity, uh generalization, adaptability, um [6:17] that's required. Obviously, no two apples are the same. The slices fall in different configurations. [6:24] There's an interesting recovery here. One slice of the apple was about to fall off the chopping [6:28] board. The robot uh basically slid it back. Going to slice the second half um and then [6:37] wipe the knife and then also put it away back in this uh in this cover. [6:45] So, this is I've seen this video many times. I also ate a lot of apples while we were while we [6:51] were getting this done. And uh you know I still find it very impressive. It's awesome. But it's [6:56] you know it's a demo. It's a demo in the lab. And so the question is does this work you know as well [7:02] u in the real world and once you spend a lot of time with these systems and [7:07] with the robots and the policies you kind of develop a deep understanding for how they work [7:11] and you also realize that you can sort of tweak these tasks to arbitrarily make the [7:16] performance you know better or worse in the lab. And so fast forward to today, [7:21] uh, Walden Robotics, we're deploying general purpose robots powered by AI. [7:26] Uh, we're building the hardware, the software, the AI, the application layer. [7:31] We are focusing on manufacturing and logistics in the beginning. We're deploying super fast. [7:36] We've done a lot of work in the lab and now we're learning what it takes to get this uh, [7:41] technology from the lab into the real world, into the field. And we have in investment and [7:47] strategic partnerships from Toyota, Boeing, Samsung, Nvidia, Coreeave and and many others. [7:55] Um, a few words about our hardware. So, this is the first uh hardware revision. This form factor, [8:01] it's uh no legs, wheeled base, this gripper configuration. It's based on our understanding [8:07] of the market um and the kind of tasks we're seeing that people need uh done or help with. [8:13] And we've designed this and we expect it to be able to do a wide range of things in the domains [8:18] that we chose to deploy in. Uh compared to Walden Zero, which is our current platform, this one has [8:24] um increased payload, advanced safety uh features, and enhanced perception system. [8:30] It's got an expanded uh workspace, um hot swappable batteries for continuous operation. [8:37] Um it's basically designed from the ground up for effective end-to-end autonomy. uh maybe the part [8:42] that I care about the most. It's got much better onboard compute, so can run better models faster. [8:49] And okay, just to show you, here's our current um or our first uh hardware uh Wen Zero. [8:57] It's based on um Rainbow RBY1 robot. We decided to go with an off-the-shelf commercial robot because [9:03] we could get it and we could deploy fast. We could get these robots deployed. We could have [9:07] them running. We did have to make significant modifications. uh to this platform to get it [9:12] deployment uh ready. Um and we're taking those lessons and putting it into wall in one. And [9:19] it's awesome. It feels like a superpower to be able to design the system, the hardware, [9:24] and uh the software together. And we're very happy with um how quickly we're moving there. [9:32] Okay. So, let's talk a bit about Walden's physical AI. [9:37] Our philosophy here is that the most valuable data is data that we get from um the field [9:44] and we deploy our robots backed by remote assistance. Um so the robots start working [9:51] autonomously uh at the deployment sites and whenever they encounter um [9:56] a new situation or something that they can't deal with on their own, a person is ready to step in, [10:02] help the robot along um make sure that useful work continues to be done for our customers and also [10:08] that intervention or that correction provides very useful data to keep training our models. And as [10:13] our models learn and improve u the ratio between operators and robots continuously improves um also [10:23] um and this pattern is deeply rooted u in the Toyota production system. So in manufacturing [10:29] we've learned decades ago that the best automation is able to detect its own flaws and Toyota calls [10:35] this principle jidoka or automation uh with a human in the loop or with a human touch. [10:42] And um according to that principle, machines detect their own faults uh ask for help, stop [10:48] operating, and don't compound the problem that's uh that's happened or uh that's about to happen. [10:54] And self-driving, by the way, follows a similar playbook. So um you know, way more self-driving [10:58] cars, they're powered by uh or like they're backed by a remote operator. Whenever the car [11:03] encounters a new situation, um the remote operator is ready to step in, assist, and rumor has it that [11:09] one operator can control u dozens of cars. So, we're inspired by the same principles and we build [11:16] uh this idea of remote assistance um as a first class citizen and as a way to basically provide [11:21] fast and seamless resolution to whatever new situations we encounter in the field. [11:27] Um talking a little bit more about this idea of remote assistance. We think of it as a way to um [11:33] create a feedback ecosystem. So for any deployment we might have a number of robots [11:38] operating autonomously. They interact with the customer. They interact with the team. [11:42] They learn, they get feedback. They may also interact with um the production line [11:47] parts that have been automated uh you know like digital signals and things like that. [11:52] Whenever a robot encounters a new situation, a remote assistant is ready to step in. The [11:59] the intervention is extremely valuable data. What we call on policy correction. It's it basically [12:05] teaches the robot here's a novel situation or here's something that you don't know how to do. [12:09] This is what you should be doing. And that data is super super valuable to um to improve the model. [12:17] And it also allows us to deploy at the frontier of our capabilities. So we can [12:21] push the boundaries of what our systems can do. Um and deployment compounds in three ways. [12:29] We generate the most valuable data. We build operational capability to basically staff a large [12:36] fleet of robots uh in the field. Um and the robots do useful work, provide value for our customers. [12:44] That's our product. Robots that do useful works for our customers for their applications. [12:49] um talking a little bit about the data diet of these models. So I think of this as a pyramid. [12:55] Um I think it's a common metaphor these days. The base layer is what we call the [13:00] pre-training layer. That's vision and language may come in the form of a VLM pre-trained model. [13:05] On top of that we have a mid-training stage. We turn this VLM into a VLA uh by adapting it [13:11] on training it on uh general robot data uh egocentric data and so on. [13:16] And then the top level of the pyramid, most valuable data on our own uh robot. [13:23] And thinking about what this might uh look in the future, we really see this data from deployment as [13:28] as becoming increasingly uh valuable and a differentiator. And as the fleet grows, [13:35] as the diversity of the tasks that we do in the field grows, this data is going to uh [13:41] compound and provide more and more uh capabilities for our models. uh and we also expect that the [13:46] value and the impact of the mid-training um slice of the pyramid is going to uh diminish. [13:54] I also think about this uh sort of draw a parallel with what's happening in digital AI where we see [13:59] these uh synthetic and post-training data feedback loops where the data that's generated by models [14:04] then goes back and is used to train the next generation models and so we see something like [14:08] this or we expect something like this is going to happen in physical AI as well where this [14:13] data from our deployments yesterday is going to make its way into the pre-training layer the [14:18] mid-training layer help train the next generation of these foundation models uh from the ground Um [14:25] in the last few minutes I want to talk a little bit about the deployment um and what we had to uh [14:31] to build and architect in order to be successful. So we developed this pipeline uh which is uh [14:37] basically a pipeline built to iterate very quickly and to improve um to improve continuously. [14:45] And the goal of this self-improving loop is to maximize the value we get from every second that u [14:51] our robot spends in the field during deployments. Um and we also want to minimize the time it takes [14:57] us to go from data that we see in the field uh observations feedback from customers and our [15:04] own team to updates to our models. So we can close this loop in a matter of days uh today [15:11] and we are operationalizing and instrumenting and observing every step of the pipeline and trying to [15:17] continuously improve that drive the time down so that we can be responsive to what we see in the [15:22] field what our customer wants what the robots can uh can do and what they're learning from. And I'm [15:27] going to spend a little bit of time talking about each one of these uh pillars in the next slides. [15:33] So I mentioned that data from deployment is the most important but [15:38] understanding um what that data is and knowing how to use it is a bit of an art [15:43] um a required skill and it's something I call data engineering. [15:48] Our philosophy here is that we basically we learn from everything and here's an example of one of [15:55] our introspection interfaces just to give you an idea. the robot is doing a particular task in the [16:00] middle and we have a few uh different rows on the bottom showing uh what's happening in the scene [16:08] and we care about two things in particular. We care about what um the robot is doing [16:14] and we do we get this through u annotation by adding hierarchical task uh descriptions [16:22] with arbitrary granularity which I think is an important uh feature and we also care about how [16:27] the robot is executing a particular task. So um we care about uh measures of task progression, [16:34] quality, speed, uh smoothness of motion and things like that which we can then use uh later [16:40] in the pipeline. And these concepts allow us to taxonomize this data to put it in different [16:46] categories uh in different buckets and based on these categories and buckets we then decide how um [16:53] and when to use this data when we train. So as an example from one of our deployment days depending [16:59] on how we query the database we might end up with 100% of the data or we could do filtering based [17:04] on some finer fine grained notions of motion smoothness for example. So we would discard [17:10] part of the data or we could do coarse filtering and discard more of the data and understanding [17:15] how to get these slices. So how to slice the data, how to end up with different portions, [17:21] how to use them, at what point in training, during mid-training, during post- trainining, that's sort [17:26] of an an art and it's part of the secret sauce that makes these robots work well in production. [17:33] um in terms of architecture. So I recently see a very vigorous debate um [17:41] um between these two types of models VAS and world models um and about which one is best suited for [17:47] u training being the starting point for a robot foundation model and um we think of that as [17:54] sort of a false duality. uh we think that the field is moving increasingly towards [17:58] unified models, omnimodels that are natively multimodel on the input and on the output. [18:04] And so our current architecture today we condition uh the policy on um images with history on robot [18:12] propriception. So the kinematic chain um language we output actions so controlling the robot and we [18:20] uh are exploring with extending that to a wide range of modalities video tactile audio and so on. [18:27] And we also see that the field is moving increasingly towards placing less importance [18:32] um on these architectures. They're becoming more and more common and maybe commoditized. [18:38] Um and the real differentiator is the data that basically flows through this model and [18:42] the curriculum. How to use it and when. Uh very briefly, simulation is one of our main pillars in [18:49] this pipeline. We use simulation to iterate very fast. We test many hypotheses. I can't imagine [18:54] moving as fast as we do without simulation. We've uh implemented a self-improvement loop for our [19:01] policies via an agentic workflow. So we generate scenes, tasks, assets um in in a generative way [19:11] and we use that for data generation and evaluation. So I want to show you an example of [19:16] what this what this pipeline looks like. Uh so we could generate you know a scene like this which is [19:21] a room, various assets, various objects. In the interest of time, I'll just skip forward [19:27] a little bit. We can generate a whole floor plan. We can place a robot here. Uh, you know, have the [19:33] robot drive around, have the robot do stuff for data collection via teleoperation, for example, [19:38] or for evaluation. And we can obviously, uh, scale that up massively, uh, which we do [19:47] in terms of deployment. Uh, just very briefly. So, we use a docker based workflow, which is [19:51] awesome. It means the policy is insulated from the rest of the system. We can iterate very quickly. [19:57] We have a pre-eployment testing benchmark suite. Um in the lab we measure many [20:03] signals anything that you can think of uh autonomy interventions transfer um speed [20:10] and also I'm very happy to say that we've managed to establish very strong positive correlation [20:15] between what we see in the lab how the policies are performing and what we see at customer sites [20:21] during deployment which is uh which is awesome because it allows us to iterate in the lab [20:26] go through many variants choose the best one and that's the one we ship and it performs the [20:32] Here's a video from uh our deployment. The robot is doing this uh particular task, picking up these [20:38] um objects, putting them on a conveyor belt. The nuance uh and the difficulty is um tough to convey [20:46] uh but it's it's very hard. You just now saw an instance of an remote intervention. Uh someone [20:53] uh stepped in, helped the robot along, and then the robot continues to do the task autonomously. [21:03] Okay, so very briefly to conclude, we are very focused on deploying our robots in the real world, [21:09] doing useful work um powered by our LBMS and backed by um remote assistants. [21:17] We're very focused on deployments as a way to get the most useful data [21:21] and also generate value for our customers. And we are increasingly thinking of our robots as um [21:29] capable assistants rather than mere tools. And we are co-developing the entire system. So hardware, [21:35] the agentic harness, um the AI so that we keep iterating and improving the capabilities. [21:44] Thank you.