How NVIDIA Halos Connects AI Development, Deployment and Safety Assurance

NVIDIA Halos is designed as an evidence-linked safety framework for physical AI, not as a single component that can make an autonomous vehicle or robot safe on its own. NVIDIA’s goal is to connect the safety-relevant work done in model development, simulation, hardware and software engineering, deployment, and later updates—creating a traceable basis for assessing the overall system over its lifecycle. 56

That distinction matters. A safety-assessed chip, a certified operating-system component or an inspected integration can support a safety case, but the company building and operating the vehicle or robot remains responsible for validating its end-to-end product, meeting regulatory requirements and operating it safely. Einride’s planned heavy-duty autonomous-trucking system illustrates that division of responsibility. 4

From cloud development to deployed machine

Halos brings together vehicle or robot architecture, chips, operating software, AI models, tools and services. For autonomous-vehicle development, NVIDIA describes a workflow that spans model training and testing, simulation with Omniverse and Cosmos, validation of the safety-relevant hardware and software stack, and deployment in the vehicle. 56

The practical premise is traceability: evidence created at one stage should inform the next. A model or system change can therefore be considered against prior validation and simulation results rather than being treated as an isolated software release. NVIDIA presents this as a way to make safety engineering more repeatable as systems, sensor configurations and software evolve. 6

NVIDIA describes Halos as covering platform safety, algorithmic safety and ecosystem safety. The first concerns the underlying compute and software foundation; the second addresses AI behavior; and the third includes integration, inspection and the broader supplier and certification ecosystem. 13

What the autonomous-vehicle stack includes

For AVs, the principal Halos elements include:

• NVIDIA DRIVE AGX Thor, the safety-engineered accelerated-compute platform.
• NVIDIA Hyperion, a full-stack vehicle platform and reference architecture for Level 4 autonomy.
• Halos OS, a unified safety software foundation built on Halos Core, the successor to NVIDIA DriveOS.
• NVIDIA Omniverse and Cosmos, used in NVIDIA’s cloud and simulation toolchain for development and validation. 610

Hyperion is not a finished robotaxi or autonomous truck. It is a reference architecture that integrates compute, sensors and software foundations, which vehicle developers adapt to their own applications and safety cases. 314

Einride’s Hyperion plan

Einride has said it will develop the next generation of its Einride Driver autonomous-trucking system on NVIDIA Hyperion. The collaboration includes adapting Hyperion’s compute, sensor, software and safety architecture for heavy-duty freight, while incorporating Halos and Cosmos for development and validation. 34

NVIDIA supplies the underlying compute platform, AI-development tools and safety framework; Einride retains responsibility for the end-to-end driving system, safety validation, regulatory approval and customer deployments. 4

What Halos adds for robotics

NVIDIA extended Halos to robotics and physical AI with a domain-specific stack. Its named foundations include NVIDIA IGX Thor for industrial edge AI and NVIDIA Holoscan Sensor Bridge for sensor I/O and processing, alongside the Halos software foundation and inspection pathway. 118

The overall safety principles are shared with AVs, but NVIDIA says the platforms, standards and supporting evidence remain specific to the robotics domain. That is an important limitation: a common framework does not mean that a factory robot, mobile robot and autonomous vehicle have interchangeable safety requirements. 6

The ecosystem: developers, sensors, silicon and integration

NVIDIA says its Halos AI Systems Inspection Lab has 43 ecosystem members. Named AV participants include AUMOVIO, Bosch, Gatik, Hesai, Lucid, MIRA, onsemi, PlusAI, Sony, Valeo and Wayve. In robotics, NVIDIA names embedded-software providers acontis and QNX, while Advantech and NexCOBOT build safety-designed IGX systems. 17

These organizations span autonomous-driving development, ADAS, sensing, silicon, systems integration, validation and assurance. Participation should not be read as an endorsement that every member’s final product is certified or deployed. 7

NVIDIA separately lists a Hyperion ecosystem of Aeva, AUMOVIO, Astemo, Arbe, Bosch, Hesai, Magna, OMNIVISION, Quanta, Sony and ZF. Astemo, AUMOVIO, Bosch, Magna, Quanta and ZF have announced Hyperion-based electronic control units, while Aeva, Arbe, Hesai, OMNIVISION and Sony are among partners qualifying sensor suites for the open platform. 20

On the semiconductor side, NVIDIA identifies onsemi in the AV inspection ecosystem, and names Infineon, NXP, STMicroelectronics and Texas Instruments as contributors of sensor, safety-microcontroller and other semiconductor technologies for robotics. 7 NXP also says it is a member of the Halos AI Systems Inspection Lab. 21

What independent assessment and accreditation do—and do not—show

NVIDIA cites more than 30 safety and cybersecurity certificates and assessment reports for Halos-related technologies. Its materials say TÜV SÜD certified NVIDIA’s automotive product-lifecycle and core hardware/software processes to Automotive Safety Integrity Level D (ASIL D); DriveOS 6.0 is certified as conformant with ISO 26262 ASIL D, and Thor-X was assessed as ASIL-D conformant. NVIDIA also cites ISO/SAE 21434 cybersecurity-process certification and a TÜV Rheinland independent UNECE safety assessment of NVIDIA DRIVE AV. 1217

Halos also includes the NVIDIA AI Systems Inspection Lab, an ANAB-accredited ISO/IEC 17020 inspection body. NVIDIA says the lab conducts impartial assessments and operates the Halos Certified program, helping partners prepare integrations for third-party certification by organizations including TÜV Rheinland, UL Solutions, TÜV SÜD, exida, SGS and CertX. 11118

The key caveat is scope. Accreditation and inspection can establish that an inspection body and its process meet defined requirements; a component certificate or assessment applies to the evaluated product and scope. Neither automatically certifies every vehicle or robot built with Halos as safe for every operating environment. Final-system developers still need application-specific validation, approvals and operational controls. 411

Why NVIDIA argues scalable safety infrastructure is necessary

NVIDIA points to projected deployment volumes as the reason to build reusable safety infrastructure. ABI Research projects an installed base of 49 million Level 3–5 autonomous vehicles by 2035, while Omdia estimates roughly 60 million industrial robots. 6

At that scale, each change in a model, sensor set, hardware configuration or software release can create a new assurance question. NVIDIA’s thesis is that systems need a common, evidence-linked approach across development and deployment so those changes can be evaluated consistently and audited more efficiently. 6

Halos is therefore best understood as an attempt to make safety assurance a continuous engineering discipline across the physical-AI lifecycle. Its promise is not that certification becomes automatic, but that developers have a structured foundation for building, documenting, inspecting and independently assessing increasingly complex autonomous systems. 156