On June 22, 2026, NVIDIA unveiled Halos for Robotics — the first full-stack safety system purpose-built for physical AI. Agility Robotics, maker of the bipedal Digit humanoid, became the first company to integrate the platform into its fifth-generation robot, targeted for deployment in late 2026.
Key takeaways
- NVIDIA Halos for Robotics is built on the IGX Thor compute module with an isolated Functional Safety Island (FSI) rated to IEC 61508 SIL 3.
- Agility Robotics is the first adopter, integrating Halos into fifth-generation Digit — designed for "cooperative safety" without physical barriers.
- The "Outside-In" architecture uses external facility cameras to dynamically manage a robot's speed constraints in real time.
- NVIDIA launched the Halos AI Systems Inspection Lab, an ANAB-accredited ISO/IEC 17020 body that pre-assesses systems before final TÜV or UL certification.
- Agility Robotics is valued at approximately $2 billion and is preparing a new funding round for further production scaling.
From autonomous vehicles to the factory floor
NVIDIA did not build this from scratch. Halos OS inherits over 18,000 engineering years of automotive safety development and more than 7 million lines of safety-assessed code from the company's autonomous vehicle programs. The company transferred that foundation to an entirely different environment — not highways, but production lines.
The Halos for Robotics stack spans three layers. At the hardware level sits the NVIDIA IGX Thor compute module, an industrial-grade unit containing an isolated FSI certifiable to SIL 3 under IEC 61508. The Holoscan Sensor Bridge handles low-latency sensor connectivity, extending functional safety all the way to the sensor edge. On top of that runs NVIDIA Halos OS, with Halos Core managing safety-critical operating functions and hardware error dispatch.
The third layer is ecosystem-facing. NVIDIA opened the Halos AI Systems Inspection Lab — an accredited inspection body under ISO/IEC 17020. Manufacturers can have their systems pre-assessed before seeking final certification from agencies such as TÜV Rheinland or UL Solutions. In practice, that shortens time to market and lowers the risk of failing a formal certification audit.
Digit steps outside the workcell
Agility Robotics has historically deployed Digit behind physical barriers. The robot operated in designated zones, and interaction with human workers was carefully limited. That is changing.
The fifth-generation Digit — targeted for Q4 2026 deployment — is designed for "cooperative safety": a robot that detects human presence and dynamically adjusts its speed and path without requiring a fence. The new model's payload capacity is 50 lbs (about 22.7 kg), deliberately aligned with OSHA's recommended manual-lifting limit for human workers.
Onboard safe human detection runs on the NVIDIA IGX Thor module. Detection happens directly on the robot, without sending data to the cloud. That matters in factory environments, where network latency could mean an accident before a response arrives.
The Outside-In architecture
Halos also introduces an open-source tool called the Outside-In Safety Blueprint. Traditional robotic safety relies entirely on onboard sensors — an "inside-out" approach with real limitations. In a cluttered loading dock, a robot may crawl because its sensors mistake trailer walls for obstacles.
Outside-In reverses that logic. External cameras installed throughout the facility feed into a four-stage pipeline. The Sensor Input Processing Pipeline converts video streams into object-tracking events. The Safety AI Monitor checks for environmental degradation — detecting, for example, a burnt-out light that distorts the image. The Safety Event Integrator validates and fuses data, discarding stale timestamps. Finally, the Safety Decision Maker, running on the isolated FSI, instructs the robot.
The practical effect: if external cameras confirm no humans are in the area, the robot can work at full speed. When a worker steps into the zone, safety guardrails activate automatically.
Why this matters
Humanoid deployments in 2026 still face a hard conversation: how do you convince a safety inspector, a corporate safety officer, and an insurer's lawyers that a robot is safe? The absence of a shared certification standard has been one of the biggest operational barriers across the industry.
Halos does not solve this with a single announcement — but it clearly lowers the entry threshold. An accredited pre-certification lab means manufacturers do not start negotiations with TÜV from a blank page. The open-source Outside-In Safety Blueprint gives the entire ecosystem a shared reference point for safety architecture discussions.
The urgency intensified after the Figure AI lawsuit, in which a former safety head described skull-fracture risks from structural flaws. NVIDIA is clearly positioning itself as a safety infrastructure provider for the whole industry — not just for Agility.
If this standard takes hold, it could accelerate humanoid deployment across more facilities and reduce the insurance-related risks that block deal after deal.
What's next
Fifth-generation Digit with Halos integration is planned for Q4 2026 deployment, as announced by Agility Robotics CEO Peggy Johnson at the Abundance Summit 2026.
Agility Robotics is preparing a new funding round by end of 2026 to scale production of cooperatively safe units.
The Outside-In Safety Blueprint is open source — the industry can submit implementations and propose changes independently of NVIDIA partnerships.
