NVIDIA Halos Targets Safety Challenge as Physical AI Moves Into the Real World
Physical AI is moving beyond research labs and into roads, factories, warehouses and other environments where machines must make decisions around people. As autonomous vehicles and robots become more capable, NVIDIA is arguing that safety needs to be designed into the entire system rather than checked only before deployment.
NVIDIA says ABI Research projects an installed base of 49 million Level 3–5 autonomous vehicles by 2035, while Omdia estimates that about 60 million industrial robots will be deployed between 2026 and 2035. Those projections highlight the scale of the safety challenge facing companies developing machines that can sense, decide and act in the physical world.
The company’s NVIDIA Halos platform is designed to address that challenge by connecting hardware, software, AI behavior, simulation, validation and deployment into a broader safety framework.
Why Physical AI Safety Is Different
Traditional safety engineering becomes more complicated when machines operate in environments that constantly change.
A factory floor can contain moving workers, robots, vehicles and equipment. Roads can change from one second to the next because of pedestrians, weather, construction or unexpected driver behavior.
NVIDIA says autonomous systems therefore need to recognize changing conditions, adapt their behavior and reach a safe state when something unexpected happens.
AI introduces another layer of complexity. Testing an AI-powered machine requires more than checking whether conventional hardware and software functions correctly. Developers also need to evaluate AI behavior and establish safeguards that operate during design, deployment and validation.
NVIDIA points to emerging standards, including ISO/IEC TS 22440, as part of the developing framework for addressing AI-specific safety risks.
NVIDIA Halos Extends From AVs to Robotics
NVIDIA originally introduced Halos as a comprehensive safety system for autonomous vehicles, covering vehicle architecture, AI models, chips, software, tools and services. The company launched its robotics-focused Halos system in June 2026, extending the approach to humanoid and industrial robots.
For autonomous vehicles, the Halos architecture includes NVIDIA DRIVE AGX Thor hardware, NVIDIA Hyperion, Halos OS, Halos Core and middleware, along with AI models and simulation and validation tools.
NVIDIA says Halos connects development and simulation in the cloud with deployment inside the vehicle, allowing safety evidence to remain traceable throughout the vehicle lifecycle. Its AV safety architecture includes design-time, deployment-time and validation-time safeguards.
For robotics, NVIDIA uses IGX Thor, Halos Core, Holoscan Sensor Bridge and simulation technologies including Isaac Lab and Omniverse.
The system is also designed to support an “outside-in” approach, where external cameras and AI systems can provide additional awareness of the environment around robots. NVIDIA describes this as a way to complement sensors mounted directly on a robot.
Simulation Becomes Critical for AI Safety
One of the biggest challenges is the sheer number of situations autonomous machines may encounter.
Testing every possible combination in the physical world can be impractical. NVIDIA says simulation, synthetic data and scenario reconstruction can therefore complement real-world testing.
Its autonomous-driving safety architecture uses NVIDIA Omniverse and Cosmos for simulation and validation, while robotics developers can use Isaac Lab and Omniverse-based environments to test robot behavior across different conditions and edge cases.
This becomes increasingly important as machines receive software and AI-model updates. A system that changes after deployment may require additional validation depending on the nature and impact of the change.
Companies Are Building Around the Halos Ecosystem
NVIDIA says its Halos ecosystem includes automakers, mobility companies, robotics developers, software providers, sensor manufacturers, semiconductor companies and safety organizations.
For autonomous vehicles, companies including Geely, Isuzu, Nissan and Einride are developing Level 4-ready vehicles using NVIDIA Hyperion, according to NVIDIA. Uber, Grab and Lyft are also among mobility providers using the platform for robotaxi development.
In robotics, NVIDIA says Agility is integrating IGX Thor and Halos Core into the safety system for its Digit humanoid robot. Other companies involved across the ecosystem include QNX, Advantech, NexCOBOT, Infineon, NXP, STMicroelectronics and Texas Instruments.
Independent Assessment Is Part of the Process
NVIDIA says safety cannot rely solely on the technology vendor’s own assessment.
For automotive applications, TÜV SÜD has certified NVIDIA’s Automotive Product Lifecycle software process and DriveOS 6.0 to ISO 26262 ASIL D, while NVIDIA says TÜV Rheinland has performed an independent UNECE safety assessment of NVIDIA DRIVE AV.
For robotics, NVIDIA says TÜV Rheinland is inspecting IGX Thor, Halos OS and Holoscan Sensor Bridge for functional-safety certification readiness.
NVIDIA also says the Halos AI Systems Inspection Lab has been accredited by the ANSI National Accreditation Board under ISO/IEC 17020. The lab is intended to inspect scoped Halos integrations and help developers prepare for final system-level certification by independent third-party organizations.
What Happens Next for Physical AI Safety?
The expansion of autonomous vehicles and robotics means safety is becoming an ongoing engineering requirement rather than a final certification step.
As AI models become more deeply involved in physical decisions, developers will need to demonstrate not only that machines can perform tasks, but also that they can respond safely when conditions change or systems encounter situations outside their normal operating assumptions.
NVIDIA’s Halos approach is one attempt to build that process across the full physical-AI lifecycle. Whether the framework becomes broadly adopted will depend on how developers, regulators, certification organizations and manufacturers apply these technologies to real-world systems.
For physical AI, the next phase of growth may therefore depend as much on proving safe behavior as on improving what autonomous machines can do.
