Nvidia Expands Halos Safety Stack for Robotaxis and Humanoid Robots
Nvidia expands Halos safety tools from robotaxis to humanoids, as its physical AI business nears $10 billion in annual revenue and deployments grow globally.
Summary
Nvidia has invested billions in physical AI despite AI data centers driving its multitrillion dollar valuation. Halos launched in 2025 for self-driving and autonomous vehicles, then expanded as Halos for Robotics in June 2026 for warehouse mobile robots, factory humanoids and surgical robots. CEO Jensen Huang said on the All-In Podcast in March 2026 that physical AI already generated nearly $10 billion annually, after naming it Nvidia’s second most important growth category in 2025.
The stack combines the Nvidia IGX Thor module, whose independent processor handles safety workloads, with an operating system that continuously monitors hardware and software while isolating safety-critical computing. Nvidia Holoscan Sensor Bridge connects sensors to processing and flags corrupted data through components including microcontrollers and field programmable gate arrays. Simulations test robots virtually, while an inspection lab reviews safety artifacts. Unlike largely consistent automotive standards, robotics needs programmable safeguards spanning hallway vacuums, heavy forklifts and robots that may nearly stop before blind corners when traveling six miles per hour.
Agility Robotics first integrated Halos into Digit 5, moving sensors inside the humanoid and removing dependence on external work cell equipment. Boston Dynamics is pursuing Halos accreditation across Spot, Stretch and Atlas, which could enter parent Hyundai’s automotive factories by 2028. KION Group is applying it to autonomous forklifts and mobile robots, while LG is developing a humanoid based on Nvidia’s Isaac GROOT foundation model. Nvidia also backs Agility, Figure AI and robotics model developers, and partnered with Unitree on an open humanoid reference design. Deployment is accelerating across the United States, China and Europe, but autonomous general-purpose humanoids remain unproven.
Positives
- Nearly $10 billion in annual physical AI revenue establishes robotics and autonomous systems as a substantial Nvidia business.
- Digit 5 incorporates its safety sensors internally, allowing Agility Robotics to change tasks without rebuilding external work cell infrastructure.
- IGX Thor separates safety processing, while Halos continuously monitors hardware and software and isolates critical workloads.
- Boston Dynamics is extending safety accreditation across Spot, Stretch and Atlas ahead of potential Hyundai factory deployments by 2028.
- Programmable safeguards let developers address different risks across vacuums, forklifts, humanoids, surgical robots and autonomous mobile machines.
Risks & concerns
- Robotics lacks the relatively consistent functional safety definitions used across autonomous vehicle companies and countries.
- Custom safety functions increase flexibility but can reduce Nvidia’s control over the underlying stack.
- Robots traveling six miles per hour may nearly halt before blind corners because factories, warehouses and hospitals remain unpredictable.
- Autonomous, general-purpose humanoids capable of performing practically any human task remain unproven.
- Broader operation near people raises the consequences of corrupted sensor data, hardware failures and software interference.