Nvidia is expanding its bet on what it calls “physical AI” by transferring the Halos safety system from autonomous vehicles to a broader range of robots, including mobile robots in warehouses, humanoid robots inside factories, and surgical robots. The company is betting that safety has become the main bottleneck as AI models and robotic hardware capabilities improve.
Nvidia launched the Halos system in 2025 to help autonomous-vehicle developers build controls that limit the risks of autonomous systems. In June 2026, it announced Halos for Robotics, an expansion that addresses the differences in risk between a floor-cleaning robot, an autonomous forklift transporting heavy loads, and a robot operating in a medical or industrial environment.
A Platform Combining Hardware and Software
The platform starts with the Nvidia IGX Thor computing unit designed for robotic and industrial applications, which includes an independent processor for safety-related workloads. According to Amit Goel, head of Nvidia’s robotics and edge-computing ecosystem, the functional system and the safety system operate on the same chip, with sensitive tasks separated to reduce the possibility of interference between them.
At the software level, the Halos operating system monitors hardware components and software libraries to detect failures quickly, while the Nvidia Holoscan Sensor Bridge connects sensor data to safety processing and helps detect corrupted data. The package also includes tools for simulating robots in virtual environments and software for a testing laboratory that allows partners to receive feedback on safety issues during development.
What Changes in Practice?
Safety in robotics differs from safety in self-driving cars because a robot may move between factories, warehouses, and hospitals with varying layouts and risks. Nvidia therefore designed the platform to allow developers to define customized safety functions while attempting to preserve the underlying architecture that enforces oversight of the system.
This becomes more important when a robot operates in an unstructured environment or approaches a corner that blocks its view. In such cases, the robot may have to slow down or stop before the turn because it cannot know what will appear ahead. These cases show that safety does not depend on the AI model alone, but on sensors, computing, software isolation, and virtual and field testing together.
Early Applications and Partnerships
Agility Robotics was the first company to integrate Nvidia Halos into its Digit 5 humanoid robot. Nvidia says that integrating safety sensors and components inside the robot reduces reliance on external sensors placed around the work area, allowing the robot to carry its safety mechanisms with it between different tasks and locations.
Boston Dynamics is also participating early in the safety-certification program, with interest in applying the platform to the legged Spot robot, the wheeled Stretch robot, and the humanoid Atlas robot. Other participants include the German KION Group, which uses autonomous forklifts and mobile robots, and South Korean company LG, which is developing a humanoid robot based on Nvidia’s Isaac GROOT foundation model. Nvidia also has investments in Agility Robotics and Figure AI, as well as a partnership with China’s Unitree to provide an open reference design for a humanoid robot intended for researchers.
Limitations and Open Questions
The expansion of Halos does not prove that humanoid robots have become independent workers capable of carrying out any task. The platform’s success will depend on developers’ ability to define suitable safety functions for each environment and to verify system performance outside test conditions. However, the transition of robots from fixed stations to multiple tasks in different locations makes a programmable safety layer a key factor in their commercial deployment.