Nvidia is betting billions on physical AI, and a key part of that bet is Halos, a full-stack safety system designed to help developers deploy robots and self-driving cars without endangering people. Launched for autonomous vehicles in 2025, Halos expanded to robotics in June 2026, targeting applications from warehouse robots to humanoid machines. According to Nvidia's Amit Goel, safety is the next bottleneck as AI models and robot hardware become more capable.
The system pairs the IGX Thor computing module's independent safety processor with an operating system that monitors every hardware block and software library for failures. It also includes the Holoscan Sensor Bridge to identify corrupted sensor data, plus simulation tools and an inspection lab program. Adapting Halos from vehicles to robots required rethinking safety, since a robot vacuum and a warehouse forklift face very different risks. Nvidia built a programmable platform that lets developers define custom safety functions without breaking the underlying infrastructure.
Agility Robotics was first to embed Halos into its Digit 5 humanoid robot, moving safety sensors onboard so the robot can operate without external barriers or fixed work cells. Boston Dynamics is an early partner in the Halos safety accreditation program, while KION Group and LG are also incorporating the platform. Nvidia CEO Jensen Huang has said physical AI already drives nearly $10 billion in annual revenue, and the company sees safety as the key bottleneck to wider deployment.