Physical AI's new training data is us: human motion is becoming the robot's textbook

Train the robot without the robot

Viam has launched a workcell-to-policy flow for robot developers that removes the most expensive part of robot learning: the robot. The workflow allows developers to create trained policies without previous machine-learning experience and without access to physical hardware, using NVIDIA Isaac Sim as the simulation environment. The platform includes a simulated workcell for policy creation, a framework for episode capture, and deployment of trained policies back to the simulator for evaluation. Users can redesign workcells by hand or with frontier agents, and the system embeds frameworks for task design and episode collection to guide users around common pitfalls. The bet is that the bottleneck in robotics is no longer hardware access but task design, and that a good simulator plus a good workflow is enough to start. m:…3895

Viam's agent integrations benefit from improvements in Astra and Fable, according to the company, which is a reminder that the robot stack now depends on frontier models for spatial reasoning and robot control. The workflow is aimed at developers who know the task but not the machine learning, and its pitch is that the simulation loop, design, capture, train, evaluate, is the product, not the robot itself. If that holds, the number of people who can build a robot policy expands well beyond the labs that own fleets. m:…3895

Humans are the dataset

Ropedia's Xperience 2 release makes the same bet at dataset scale. It includes 1,589 released metric-scale 3D scans and 4,230 egocentric recordings of 74 tasks, and proposes the joint capture of environments and behaviours as a gold standard for physical AI data collection. Scenes can be converted into physically interactive simulation environments, and human motion in the dataset can be retargeted to whole-body robot trajectories. The release aims to address the high cost and difficulty of collecting real-robot data by starting from human-centric recordings. For comparison, the authors note that DROID's 350 hours of demonstrations required 50 operators working for a full year. Human video is cheaper than robot time, and there is far more of it. m:…2588

NVIDIA's SOMA Retargeter is the tool that makes that bet workable. Human motion offers a starting point for teaching humanoid robots how to move, because human demonstrations capture the coordination and timing behind natural movement. But robots cannot simply copy the same joint motion from a human: robot proportions, joint structure and range of motion differ from humans. The SOMA Retargeter is designed to adapt human motion for robots, bridging the gap between demonstration and execution. The dataset gives you the motion; the retargeter makes it fit the body that has to perform it. m:…1811

The contact problem is the hardest part of that transfer, and a preprint on DexRemo tackles it directly. DexRemo provides a nine-scenario task suite for removal manipulation, including adhesive detachment, surface cleaning and occluded retrieval, built on a three-finger hand with force sensing. Its finding is that wrist-centered force frames improve stability and generalization in contact-rich manipulation compared with local-frame force inputs, and that multi-finger force observations are more effectively fused when expressed in a shared wrist-centered frame, giving force components consistent geometric meaning. The authors also note that large foundation models like GPT-6 Astra and Claude Fable 5.1 remain brittle on long-horizon, contact-rich tasks in real-world environments. Video and motion get the robot to the object; force sensing gets it through the contact. m:…6580 m:…2588

And on public roads

Zoox, meanwhile, has begun robotaxi operations in Atlanta, Georgia, concentrated in the Northwest region, specifically Atlantic Station, Woodward and Home Park. Current ride access is restricted to Zoox employees, their families and friends, and the company plans to open a public waitlist in the coming months. The company began operations in Atlanta in the spring of 2025 to map local driving behaviours and infrastructure, and says it engages with local and state officials before beginning testing in any new market. The vehicle is purpose-built without a steering wheel or driver's seat, with carriage-style seating for up to four passengers. Zoox claims to be the only company currently deploying a fully driverless, purpose-built robotaxi on public roads in the US, a claim that is Zoox's own and should be read as such. m:…0243

The thread across all five items is that physical AI is shifting from collecting robot data to repurposing human data: human motion, human video, human demonstrations in simulation. The hardware still matters, and the contact problem is unsolved, but the training pipeline is starting to look like the language-model pipeline: a large, cheap, human-generated corpus, a retargeting step, and a simulator to close the loop. Zoox is the reminder that the other end of the pipeline is a vehicle on a public road, where the tolerance for error is not a benchmark score but a city block. m:…3895 m:…2588 m:…0243