Nvidia's bet on physical AI: Halos, its safety system for robotaxis and humanoid robots

Nvidia is best known for the chips that power AI data centers and have driven its market capitalization into the trillions, but the company is also betting heavily on what it calls "physical AI": robotics and autonomous vehicles. Part of that bet involves building a full-stack safety system that other companies can use as a foundation to reduce the risk of their machines harming the people working near them.

That system is called Halos. It launched in 2025 as a set of hardware and software tools designed to help developers implement safety guardrails in autonomous cars and other vehicles. In June 2026, Nvidia extended the architecture with "Nvidia Halos for Robotics," aiming to enable safe deployments of autonomous mobile robots in warehouses, humanoid robots walking through factories and even surgical robots. Amit Goel, who leads Nvidia's robotics and edge computing ecosystem, explained that as AI models and robotic hardware grow more capable, safety is shaping up to be the next major bottleneck, which is why the company launched Halos for Robotics: to unlock the potential of these systems.

The business context behind this expansion matters. In March 2026, Nvidia CEO Jensen Huang said on the All-In podcast that the physical AI business was already generating nearly $10 billion in annual revenue for the company. Huang began pointing to physical AI as Nvidia's second-largest growth category back in 2025. In parallel, the company has invested directly in several robotics firms, including humanoid robot startups Agility Robotics and Figure AI, as well as companies developing foundation models for robotics. Nvidia has also partnered with China's Unitree to offer an "open reference design" for a humanoid robot aimed at researchers.

On the technical side, Nvidia's safety system starts with the hardware: the IGX Thor compute module, built for robotics and industrial applications and equipped with a separate processor dedicated exclusively to safety-related workloads. According to Goel, in the context of physical AI the functional system and the safety system need to run on the same silicon. On the software side, the Halos operating system constantly monitors "every hardware block" and "every software library" to catch faults quickly, while isolating safety-critical workloads to prevent interference. Added to that is the Nvidia Holoscan Sensor Bridge, which connects sensor data to safety processing in a way that makes corrupted data easy to identify, and which can be integrated into individual hardware components such as a microcontroller or an FPGA. The package is rounded out by simulation environments for testing robots virtually and an inspection lab program that lets robotics companies get rapid feedback on any safety "artifact" detected during development.

Adapting Halos from autonomous vehicles to robotics was no trivial matter, because it meant accommodating very different definitions of safety. According to Goel, functional safety in autonomous driving stays fairly uniform across manufacturers and countries, but a vacuum robot cleaning a hallway has radically different safety requirements than a robotic forklift handling heavy loads at a loading dock. That forced Nvidia to rethink much of the groundwork in order to build a programmable platform that gives developers the "hooks" they need to define custom safety functions without losing control over the underlying infrastructure: flexibility, Goel warned, can come at the cost of losing control over the technology stack. That customization is also necessary because robots operate in complex, unstructured environments: a robot moving down a hallway cannot see its own blind spots or what might appear around a corner, so in factories, warehouses or hospitals it tends to slow almost to a stop before turning, since things can come from any direction.

When it comes to real-world deployments, Agility Robotics was the first to build Halos into its Digit 5 humanoid robot, designed to work safely near people without requiring isolated workstations or physical barriers. Nvidia's system allowed Agility to integrate the relevant sensors and safety hardware inside the robot itself, something that previously depended on external sensors placed around the robot's work cell. As Goel puts it, the robot is now "literally unchained": safety travels with it wherever it goes, so it can operate in different areas of a factory or warehouse without having to set up new infrastructure every time it is assigned a different task.

Massachusetts-based robotics company Boston Dynamics is also taking part as an early partner in the Halos safety accreditation program, focused on building a common safety platform for its Spot robot dog, its wheeled robot with a robotic arm, Stretch, and its humanoid, Atlas, which could begin deploying in the car factories of its parent company, Hyundai, in 2028. Other companies adopting Halos include Germany's KION group, which deploys autonomous forklifts and other mobile robots, and South Korea's LG, which is developing its own humanoid robot based on Nvidia's humanoid foundation model, Isaac GROOT.

The article frames all of this within a broader shift: the United States, China and Europe are driving a growing deployment of robots that goes well beyond humanoids, testing these machines in an ever-wider range of environments. Goel sums up the paradigm change by noting that robots used to be confined to a single station doing a single task, whereas in the new generation of robotics, with generalized bodies and skills, it is essential that a single robot be able to perform multiple tasks across multiple environments—and that the only way to achieve that in real production is by improving safety. The article itself leaves the underlying question open: it remains to be seen whether humanoid robots will ever become general-purpose autonomous workers capable of performing virtually any human task.