AWS's Open-Source Physical AI Toolchain Names Five Stages and Four NVIDIA Parts, but Its Solutions Page Lists No Licence or Repository
Hardware / news
AWS's Open-Source Physical AI Toolchain Names Five Stages and Four NVIDIA Parts, but Its Solutions Page Lists No Licence or Repository
The Robot Report calls it open source and hardware-neutral. The AWS page names SageMaker, IoT Greengrass and NVIDIA's physics engine, and gives no price, release date or code link.

Amazon Web Services launched an open-source Physical AI Toolchain for roboticists on Oct. 8, according to The Robot Report. The AWS page for the guidance it points to, Guidance for Physical AI for Robotics on AWS, lists no licence, links no repository, gives no price and carries no release date. The only date on it is the 2026 copyright line.
Five stages, two suppliers
The toolchain runs from synthetic data generation through model training, simulation and validation, and edge deployment, to continuous improvement, per The Robot Report. AWS supplies Amazon SageMaker for training and AWS IoT Greengrass for distributing models to edge devices. NVIDIA supplies Isaac Sim, Isaac Lab, Isaac GR00T and Cosmos.
| Supplier | Named parts | Role given |
|---|---|---|
| AWS | Amazon SageMaker | Model training |
| AWS | AWS IoT Greengrass | Distributing models to edge devices |
| NVIDIA | Isaac Sim, Isaac Lab, Isaac GR00T, Cosmos | No per-stage mapping in either source |
Neither source says which NVIDIA part handles which of the five stages, and we have not inferred it. The AWS page names only "NVIDIA's physics engine" for simulation-based training.
What the page does not say
The AWS page is a Solutions Library guidance. It says the sample code is provided under the AWS Customer Agreement or a separate written agreement, and that it is not meant for production accounts or critical data. It warns that deploying the content may incur AWS charges, such as for EC2 instances or S3 storage. That language belongs to a reference architecture, which is a different thing from the project The Robot Report describes as open source.
The two are possibly the same release, described from two angles. The reporting available does not settle it. A reader looking for the code should look for the repository before planning around the label.
Hardware-neutral, by design
Sri Elaprolu, Director of Frontier AI Science and Engineering at AWS, said "The toolchain is intentionally staying neutral to that final step." The final step is the robot. The toolchain stops at the edge device, which is what lets AWS skip the question of whose arms and legs run the model.
Elaprolu said it draws on lessons from Amazon's robotics operations, which include more than 1 million robots. The Robot Report also cites RLWRLD, a South Korean company that worked with AWS on a tactile sensing model for five-finger dexterity tasks, and Bedrock Robotics, which trains models for autonomous construction equipment on AWS. It notes Telexistence has deployed more than 300 humanoids in convenience stores. None of those is described as a user of this toolchain.
The RoboMaker gap
The toolchain is not a direct replacement for AWS RoboMaker, which AWS shut down in 2025, per The Robot Report. Elaprolu described RoboMaker as one component of AWS's robotics stack. That leaves roboticists who built on RoboMaker without a named migration path in either source.
What would change this read
A GitHub repository with a licence file would settle the open-source question in one commit. A published price for a training run would settle the cost question; today the only cost guidance is that AWS charges may apply. Until then the Oct. 8 launch is a description of intent from a trade publication and a reference-architecture page. For other robotics reporting, see Boston Dynamics' Atlas plan and Mecka AI's robot-training data round.
Sources
More in Hardware
- 01Ring's $250 Indoor Drone Flies Eight Minutes a Charge; Its Price Is Reported Two WaysThe Drone Cam ships January 27 pending FCC approval and the $1,000 Rover Cam on December 16, but CNBC and TechCrunch disagree on the drone's price.
- 02BYD's Charging Robot in Abu Dhabi Claims 10% to 70% in 5 Minutes, With No kW RatingEvery charge time in the coverage is BYD's own; no report gives the station's power, compatible cars or price.
- 03Boston Dynamics Puts Scooter Wheels on Atlas and Publishes No Speed FigureA single adapter swaps Atlas's feet for wheels in what the company calls a research testbed, trained with reinforcement learning in MuJoCo.
- 04AMD's 256-Core EPYC 9996 Has 16.4 ns L3 Latency; the 96-Core 9686F Has 10.7 nsChips and Cheese's measurements show Zen 6 and Zen 6c diverge inside the out-of-order engine, with the dense core trading cache speed for scale.