NVIDIA’s Halos OS delivers built-in hardware safety for physical AI tasks in industrial settings.
Standard safety measures built for static enclosures fall short when self-guided robots share workspace with humans. Agility Robotics is now rolling out the Halos OS platform across its Digit humanoid lineup. The company harnesses NVIDIA’s IGX Thor processor to drive its human proximity detection mechanism. Specialists verify these rollouts using the Halos AI Systems Inspection Lab to secure regulatory approval.
Halos seamlessly adapts self-driving safety validation techniques for factory-floor purposes. NVIDIA invested roughly 18,000 years of combined engineering work and evaluated some 21 billion safety-focused transistors to craft the underlying software layer.
This design avoids repetitive engineering effort by reusing around seven million lines of proven code created for automotive contexts. Both the validation workflows and functional safety benchmarks transfer seamlessly between mobility and industrial use cases. TÜV SÜD along with TÜV Rheinland conducted separate third-party audits to certify hardware conformity for dual sectors.
NVIDIA shapes functional safety guidelines by chairing the IEC 61508 standards working group. The enterprise is actively involved in developing the ISO 25785-1 framework to define universal baseline criteria.
IGX Thor hardware for physical AI
The IGX Thor module delivers as much as 2,070 FP4 TFLOPs of processing power. Inside are 14 Neoverse ARM CPU cores paired with 128 GB of RAM delivering 273 GB/s throughput.
A dedicated Functional Safety Island is built into the chip, keeping safety-critical logic fully segregated from the main computing cores. The isolated island supports its own power supply, timing reference, and I/O pathways.
Over 22,000 discrete protection mechanisms continually monitor the system-on-chip for hardware errors. Numerous compute engines run identical tasks in lockstep to create architectural redundancy. Built-in diagnostic circuitry examines deep logic layers to catch slow-emerging defects.
The Safety Extension Package orchestrates all on-chip safety functions. It accumulates hardware error information and ships that telemetry directly to the Safety MCU. The software stack runs atop Halos Core Linux, moving data through an Edge Safety Link protocol. For applications demanding extreme workload separation, a secondary QNX virtual machine can be instantiated via the NV Hypervisor. That QNX layer quarantines safety packages from general AI processing.
The Holoscan Sensor Bridge governs external sensor connections over encrypted MACsec links. Deployments leverage this bridge to validate sensor streams at the network boundary. ConnectX RDMA hardware provides a low-latency video feed path to the core GPUs. The protocol works generically with any connected sensor or actuator device.
External perception network integration
Warehouse crews tackling automated trailer loading often face vehicle positioning breakdowns. Forklifts that rely solely on built-in sensors overshoot cargo walls, treating them as physical barriers.
Self-driving gear typically powers down completely once these false alarms activate. Sites get around it by routing overhead cameras into a unified processor pipeline. NVIDIA’s Metropolis platform pulls in those security feeds, calculating object movement paths and speed vectors. The pipeline converts spatial data into distinct triggers, watching for machines breaching geofenced perimeters or logging asset proximity events.
The Safety Event Integrator merges inputs from every camera to establish confidence scores. Slow or lagging video frames get rejected so choices reflect only real-time telemetry. Engineers program custom rules inside the integrator to dial in precise behavior thresholds. The Safety Decision Maker node translates those values into actual vehicle commands. It executes solely within the Functional Safety Island’s secure boundary.
When external cameras report an unoccupied loading zone, the node lifts the forklift’s software-based speed and movement limits. The truck accelerates to its peak work rate inside the trailer. The instant a colleague crosses the digital perimeter wire, a proximity alarm activates. That event notification traverses the mesh straight to the decision node. The system reinstates every original safety limit on the forklift the moment it detects human presence.
A Safety AI Monitor watches round-the-clock. It searches the entire video chain for link interruptions or image degradation. Condensation, abrupt light failure, and soiled lens detection trigger out-of-distribution notifications. As soon as those arrive, the decision node commands the vehicle into a fixed hardware-safe state. There it stays latched until visibility parameters fully normalize. The architecture ensures machine protection holds steady even when the facility departs from the original dataset assumptions.
Industrial deployment and certification strategies
NVIDIA’s AI Systems Inspection Lab holds ANAB accreditation for auditing machine safety integrations. It evaluates self-driving gear and industrial hardware setups equally.
Component vendors submit proprietary codebases to NVIDIA’s evaluation team. The lab cross-references outside applications against its library of pre-certified Halos modules. Auditors then issue thorough compliance reports confirming alignment with ISO 13849 criteria.
Agility Robotics is currently submitting its Digit firmware stack for a firmware-level cybersecurity assessment. The lab network spans 43 enrolled companies, counting Ouster, Peer Robotics, and Boston Dynamics among active contributors.
Plant managers kickstarting physical AI projects skip hands-on provisioning by pulling NVIDIA’s pre-built resources directly. Engineers invoke targeted agent skills to generate complete outside-in safety architectures automatically. The agent pulls all necessary packages and configures the Metropolis visual pipelines in one step.
That frees facility teams to channel all effort into custom task logic rather than re-addressing baseline platform compliance.
See also: Matter 1.6 standardizes ‘Joint Fabric’ network management
Want to learn more about the IoT from industry leaders? Check out IoT Tech Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including AI & Big Data Expo and the Cyber Security Expo. Click here for more information.
IoT News is powered by TechForgeMedia. Explore other upcoming enterprise technology events and webinars here.
