Physical AI toolchain, a walking hand

Today: AWS open sources a physical AI toolchain, Helm.ai books $70 million in foundation model contracts, an ETH Zurich hand walks on its fingers, EasyNav 0.5.0 ships for four ROS 2 distributions, and NVIDIA shows agents assembling simulations.

AWS packages the physical AI loop as an open toolchain

AWS published a Physical AI Toolchain, an open source reference workflow for taking a robot from demonstration data to a deployed model. It joins five stages that teams usually wire up by hand: synthetic data generation, training, simulation and validation, edge deployment, and continuous improvement from field data. AWS supplies SageMaker for training and IoT Greengrass for pushing models to the edge, while NVIDIA contributes Isaac Sim, Isaac Lab, Isaac GR00T and Cosmos (The Robot Report, AWS guidance).

Sri Elaprolu, who directs Frontier AI Science and Engineering at AWS, frames the problem as data scarcity: unlike language models, there is no internet-scale corpus for manipulation, so teams must synthesize representative scenarios themselves. The architecture is hardware-neutral, and AWS says it is not a replacement for RoboMaker, the hosted simulation service shut down in 2025. A working sample lives in the aws-samples repository.

Amazon’s own numbers shape the pitch: more than 1 million robots in its fulfilment network and 300 or more Telexistence humanoids in Japanese convenience stores. Microsoft publishes a comparable physical-ai-toolchain, MIT licensed and dating from March 2026, pairing Azure Machine Learning, AKS and Arc with Isaac Sim, Isaac Lab, OSMO and Jetson. Two clouds now package the same loop around the same NVIDIA stack, and the data half is still the hard part.

Helm.ai says its foundation models booked $70M

Helm.ai, founded in 2016 and based in Redwood City, said it has signed $70 million in commercial contracts for its physical AI foundation models over a 12 month period. The customers span automotive OEMs, Tier 1 suppliers and industrial automation companies, with additional work in mining, construction and delivery drones, and the company says it covers SAE Level 2 through Level 4 driving programs on a path to break even (The Robot Report).

The claim behind that is a method Helm.ai calls “deep teaching”: unsupervised training that learns the structure of the physical world separately from the problem of acting in it, so the system needs far less labelled data. CEO Vladislav Voroninski argues this matters more in robotics than in cars, because no robot fleet exists yet to generate the data. That is the argument behind why VLA models matter, separating perception from policy rather than training one network end to end. Watch whether the contract figure converts into disclosed production programs.

An off-the-shelf hand walks on its fingertips

ETH Zurich’s Soft Robotics Lab turned a commercial anthropomorphic hand from Wuji Technology into a self-contained mobile robot, with an 80 gram backpack holding a battery, an inertial measurement unit and a Raspberry Pi Zero. Amirhossein Kazemipour led the work. A hand is not symmetric the way most legged platforms are, since fingers differ in length and the thumb opposes, so off-the-shelf quadruped rewards did not transfer. The team trained policies in a simulator calibrated from hardware measurements, adding “virtual springs” that penalize each finger for straying from its natural crawling pose (IEEE Spectrum, arXiv:2609.17172).

The hardware numbers are concrete. The hand crawled across 14 surfaces, including smooth floors, metal grates, grass and gravel, averaged 9 centimeters per second, and righted itself in 21 of 25 trials. Given an overhead camera it pushed a 41 gram cube to targets up to 40 centimeters away, 15 times in a row. Researchers quoted by Spectrum like the idea and question durability, since walking loads a hand very differently from grasping.

EasyNav 0.5.0 ships a new ROS 2 navigation stack

The Intelligent Robotics Lab released EasyNav 0.5.0, an open source navigation system for ROS 2, at the same version for Humble, Jazzy, Kilted and Lyrical. Packages are in the ROS 2 buildfarm and install after the next sync, while the Pixi route needs no system ROS 2 at all, because the environment brings its own (ROS Discourse). Maps are configuration rather than code: 2D for indoor floors, 3D surfaces for ramps and uneven terrain, voxel maps, or a mix, with localization, planning and control chosen in one parameter file and run as one process. An easynav_nav2_bridge lets existing Nav2 clients send goals unchanged, and four Gazebo PlayGrounds ship as packages for a Kobuki, a Robotnik Summit XL, a PAL TIAGo and omnidirectional bases.

The part worth reading is the safety mode: it validates and freezes the configuration, enforces velocity limits, monitors the real-time cycle and a heartbeat, and reacts to localization failure, sensor loss or a stuck robot while navigation runs. Real-time behavior is what executors and callback groups decide in ROS 2, so teams moving off Nav2 should read the ROS 2 executors and callback groups notes first.

NVIDIA shows agents assembling robot simulations

NVIDIA published Omniverse projects where frontier AI agents take natural-language instructions and wire up simulation applications. In one, an Omniverse product manager had an agent connect the physics (ovphysx), scene update (ovstage), rendering (ovrtx) and UI (ovui) libraries, plus SimReady assets, into an interactive warehouse humanoid simulator. Another built a Robo Olympics scene where simulated Unitree G1 humanoids play sports: the agent wrote controllers, Newton simulated them, Warp accelerated the math and ovrtx rendered the result. NVIDIA reports one hurdle cleared in 64 of 100 simulation trials. A third project compared recorded and simulated camera and lidar output to improve digital twins, and a fourth modelled a car suspension in PTC Onshape and configured it in Isaac Sim so the agent could design a wrench that reaches the bolts (NVIDIA).

Read these as staff demonstrations rather than benchmarks, but note the reported units: iteration counts, walls of trials and measured sensor discrepancies, which is where simulation work usually burns time (Isaac Sim).

One policy, 22 quadruped skills, eight teachers

A new paper trains a single legged policy to perform 22 distinct tasks using eight teacher policies. Train narrow teachers with reinforcement learning, distil them into one student with a combined reinforcement and imitation learning objective, then select tasks adversarially so training concentrates on whichever task performs worst. The authors report better motion quality and command tracking than PPO and distill-then-finetune baselines, and in some cases the student generalizes to tasks it never trained on. They deployed the result on a Unitree B1 quadruped for farm robotics and space exploration, where one platform must walk, dig and survey (arXiv:2610.10601).

The composition claim is the interesting part: walking, digging and hopping combine into behaviors such as crawling without fresh reward engineering. Multi-task legged control has long been limited by gradient conflict between tasks, the same tension that appears when one world action model must cover many embodiments.

Quick Bites

• SafeWorld’s stealth exit is now confirmed at $12.2 million in seed funding, co-led by Shine Capital and a16z Speedrun. The safety simulation company says it works with about half a dozen customers, two thirds of them publicly traded (The Robot Report).
• Kyvern is an Apache 2.0 Python library that records a robot’s own decisions in an Ed25519 signed, hash linked chain, with RFC 3161 anchoring and a PDF report mapped to EU AI Act Articles 12 and 14. The author notes it records and verifies but does not gate commands (ROS Discourse).
• Jabil’s Thomas Brown says humanoid production is leaving the hand-built prototype phase and entering the “tens of thousands”, forcing work on cost structure and manufacturability, with the NVIDIA Jetson line the popular edge compute choice among Jabil’s robotics customers (The Robot Report).
• A ROS 2 driver with 6-DOF SpaceMouse teleoperation for the AgileX PiPER arm ships with a live MuJoCo digital twin and datasets packaged for LeRobot and ACT imitation learning (ROS Discourse).