AWS launches open-source robot AI toolkit

Amazon Web Services has launched its Physical AI Toolchain, an open-source toolkit for designing robots equipped with physical AI. The toolchain, developed in collaboration with NVIDIA, aims to help companies create intelligent machines capable of perceiving, reasoning, and acting in a few weeks rather than several years.

Toolchain accelerates robot development

The toolkit allows companies to generate synthetic training data through computers, reducing the need for costly real-world data collection. AI models are then trained and tested in simulations before being deployed on real hardware.

Robots learn and improve collectively

Once installed in robots, these models make decisions on-site without a continuous cloud connection. Data collected by the machines in the field is sent back to the cloud to retrain models, enabling the entire fleet to benefit from the learning of each robot.

AWS and NVIDIA provide key components

The toolchain's code is open-source, but the system relies on paying AWS services. Amazon's contribution includes SageMaker, its AI model training platform, EC2 computing servers equipped with graphics chips for simulations, IoT Greengrass for sending models to on-site machines, and Bedrock AgentCore for system coordination. NVIDIA provides specific robotics components such as Isaac Sim, a simulator that reproduces the real world, Isaac Lab for robot learning through trial and error, GR00T for training humanoids, and Cosmos for generating virtual worlds to expand training scenarios.

Startups increasingly use physical AI

A study conducted by Strand Partners and presented by AWS surveyed 4,420 startups in 20 countries. The report indicates that one out of seven startups already designs physical AI, with 17% of French startups developing physical AI compared to an average of 15% in the twenty countries studied. France is ranked among the markets "accelerating," tied with Singapore, while Germany leads with 24% of its startups engaged in the field.

Focus remains on specific AI applications

However, the study also highlights that only 12% of these designers are working on humanoid or fully autonomous robots. The primary focus remains on computer vision (55%) and embedded sensors (50%). The industry and automotive sectors (55%) are expected to be most affected, followed closely by logistics and trade (52%).

Mixed views on job impact

On employment, 58% of startups globally believe physical AI will create more jobs than it will eliminate. In France, only 50% share this view, while 19% fear the opposite, the highest share among the twenty countries studied.

Customer readiness remains a challenge

The main brake to the development of physical AI is not technological but customer-related, as only 25% of startups believe their market is ready to deploy physical AI. Only 14% think most companies in their country could take the step today, with many having never seen these machines in action. As a result, 70% of startups say that customers' slow decision-making is slowing their growth.

Financing and trade barriers hinder progress

Additional obstacles include 66% of startups having difficulty finding financing due to the high cost and long development time of equipment, and 48% citing tariffs or trade barriers as complicating access to components. Despite these challenges, 76% of startups expect physical AI to become common in their sector within five years.