NVIDIA unveils Halos, a functional safety system for AI-powered robots

NVIDIA has announced Halos, a comprehensive functional safety stack built specifically for physical AI: autonomous robots that work side by side with people on factory floors, in warehouses, hospitals and homes.

Why old safety models don't work

For decades, industrial robots worked behind metal cages and fences. Their safety rested on isolation: if a person entered the danger zone, the machine stopped. Sensors were simple, the environment was predictable, and the robot's behavior was fully deterministic.

Physical AI changes the rules. Humanoid robots, mobile manipulators and the new generation of autonomous delivery robots operate in unstructured, constantly changing environments next to people. Their behavior is driven by neural networks: complex, opaque, probabilistic systems. Traditional safety verification methods, built on deterministic algorithms, simply do not apply to such systems. Moreover, neural network decisions are hard to verify with an exhaustive test suite: there is no explicit logic to walk through and cover.

Analysts forecast that by 2030, more than a million autonomous robots will be working in industrial and service sectors. Without a reliable safety architecture for physical AI, that growth will stay blocked, both by regulators and by the market. NVIDIA Halos is meant to close exactly this gap.

What Halos is made of

Halos is not a standalone module but a safety architecture that runs through the robot's entire compute stack. The system spans several layers:

• Perception safety: real-time monitoring of data from cameras, lidars and other sensors, detecting faults and anomalies before they reach the models
• Planning safety: checking the AI planner's decisions against formal safety constraints before execution
• Control safety: hardware and software mechanisms for emergency stops and failsafe transitions
• Independent system monitor: a separate component that watches the others and is deliberately independent of the main AI stack
• Certification traceability: automatic documentation for compliance with the ISO 26262 and IEC 61508 standards

Halos is designed as a native part of the NVIDIA Isaac platform. Integration with Isaac Sim makes it possible to test safety in virtual scenarios with thousands of hours of simulated human interaction, long before the first physical prototype reaches a real production site.

Who needs it right now

The first audience is companies already bringing humanoid and mobile robots to commercial markets. In warehouse logistics, production lines where humans and machines work together, and medical facilities, regulatory requirements are getting stricter and the cost of a mistake is higher.

"Physical AI is arriving faster than most expected.

AI safety is the key," NVIDIA Developer Blog.

For robot makers, Halos changes the equation: instead of building functional safety from scratch in each product, they get a ready, verified architecture. That shortens time to market and reduces certification risk. This matters especially in medicine and public spaces, where safety is not a technical parameter but a legal and ethical requirement.

What it means

NVIDIA is steadily building a vertical stack for the era of physical AI: from specialized chips (Jetson, Thor) and simulators to training frameworks and now functional safety systems. Halos closes the last major gap in that chain.

Until now, robot makers handled safety engineering on their own, without standardized tools. For small teams, compliance with ISO 26262 meant years of expertise and a dedicated staff of functional safety specialists. NVIDIA's platform approach could change that and noticeably speed up the commercialization of physical AI in industry and beyond.