Physical AI — robots operating autonomously alongside people in factories, warehouses, hospitals and homes — is expanding faster than many expected. Traditional safety approaches designed for structured, caged environments struggle as robots move into less structured spaces. NVIDIA today announced NVIDIA Halos for Robotics, a platform that integrates AI compute and functional safety into a single, full‑stack solution.
Extending proven AV safety to robotics
NVIDIA builds Halos for Robotics on years of functional safety work in autonomous vehicles (AVs). The company reports more than 18,000 engineering years of vehicle safety effort, over 21 billion safety transistors assessed, and more than 7 million lines of safety‑assessed code; it also cites over 22,000 platform safety monitors, 330+ AV safety research papers and 30+ certificates and assessment reports.
NVIDIA’s argument is that the processes, tools and foundational standards used in AV safety (ISO 26262 → IEC 61508, ISO 13849) can be extended rather than rebuilt for robotics. Third‑party assessments by TÜV SÜD and TÜV Rheinland are cited as confirming cross‑domain compliance. NVIDIA also participates in standards work — including convenorship of IEC 61508 and involvement in ISO/IEC TS 22440 and other committees — to shape robotics safety requirements.
Architecture and core components
Halos for Robotics follows a three‑layer safety model similar to Halos for AV: platform safety (hardware), Halos OS (software) and ecosystem safety.
- Platform hardware: NVIDIA IGX Thor and NVIDIA Holoscan Sensor Bridge (HSB)
- Software: Halos OS, including Halos Core and Halos applications/Blueprints
- Ecosystem: Halos AI Systems Inspection Lab and partner network
Key IGX Thor specifications cited: up to 2,070 FP4 TFLOPs of AI performance, 14 Neoverse ARM CPU cores and 128 GB memory at 273 GB/s bandwidth. IGX Thor is presented as an industrial AI compute module with integrated functional safety hardware rather than a general‑purpose compute board.
Documented hardware safety features include an IEC 61508 SIL 3 capable Safety Island (FSI) with up to 12K DMIPs and isolated I/O/power/clocks, more than 22,000 diagnostic mechanisms across the SoC, systematic development to IEC 61508 SC 3 for IPs used in safety contexts, diversity and redundancy options (GPU/CPU, GPU/PVA, CCPLEX CPU/FSI CPU), In‑System Test (IST) for latent fault coverage, and features supporting Freedom from Interference (FFI) and Dependent Failure Initiator (DFI) such as SMMU, GPU watchdogs, hardware GPU context switch, NOC firewalls and clock/voltage/thermal monitors.
These hardware safety mechanisms are controlled by the Safety Extension Package (SEP) service within Halos Core, which collects and routes hardware errors to the FSI and Safety MCU (SMCU), and includes FSI/SMCU reference firmware, an Error Propagation Layer (EPL) and the Edge Safety Link protocol. An application note on using IGX and Halos OS for safety is available under NDA.
Holoscan Sensor Bridge (HSB)
HSB connects sensors and actuators to IGX over Ethernet to extend safety to the sensor edge. Capabilities highlighted are low latency (ConnectX RDMA and RTX GPU Direct for real‑time streaming), scalability to hundreds of sensors and hundreds of Gbit/s, a multimodal, domain‑agnostic protocol for any sensor or actuator, and built‑in security and safety features (MACsec, IEC 61508 SIL 2 end‑to‑end safety protocol, watermarking and camera testing support included in Halos Core).
HSB chip and sensor partners named include Texas Instruments, STMicroelectronics, NXP Semiconductors and Lattice Semiconductor; IGX ODM partners include Advantech, Nexcobot, Inventec and Connect Tech. Safety MCU and sensor partners include Infineon, NXP Semiconductors and Texas Instruments.
Halos OS and Halos Core configurations
Halos OS provides certified building blocks between hardware and applications. At present the available elements are Halos Core (the safety OS) and Halos Applications (safety blueprints such as Outside‑In Safety). Middleware and Infra tools for robotics are available but not yet certified for safety applications.
Halos Core is the next generation of NVIDIA DriveOS certified to automotive safety standards and runs on IGX Thor. Two configurations are available: Halos Core Linux, and Halos Core Linux plus QNX with an NV Hypervisor. The Linux‑only option provides a safe software foundation (Linux runtime, SEP, Edge Safety Link, FSI RTOS and Safety MCU RTOS firmware). The Linux+QNX option partitions IGX into isolated VMs — a Linux VM for AI/app workloads, and a QNX VM for safety‑critical functions — enabling stronger software partitioning for higher safety integrity cases.
Ecosystem partners at this layer include Blackberry on QNX, Acontis for EtherCAT/FSOE, and FreeRTOS (with AWS offering a Safety Certification Bundle as part of Halos OS). Both Halos Core configurations are available now for early access to registered developers.
Reference blueprint: Outside‑In Safety
NVIDIA provides an Outside‑In Safety Blueprint as a reference for building functional safety agents that augment onboard robot perception with external infrastructure cameras and AI. The blueprint runs on IGX and is available as open source. It includes documentation to help address AI functional safety standards such as ISO/IEC TR 5469 and the forthcoming ISO/IEC TS 22440.
Blueprint components described:
- Sensor Input Processing Pipeline (SIPP): uses the NVIDIA Metropolis VSS (video search and summarization) reference perception stack to ingest camera streams, detect and track objects, and convert those tracks into discrete events (ROI entry/exit, proximity, tripwire crossings).
- Safety AI Monitor (SAIM): monitors perception pipeline integrity and detects out‑of‑distribution inputs, camera blockage, connectivity drops and image anomalies; on detection it raises alerts that prevent downstream decisions from relying on degraded AI outputs, causing fallback to onboard safety until conditions recover.
- Safety Event Integrator (SEI): fuses events from multiple cameras, applies staleness checks and confidence thresholds, and forwards validated events to the Safety Decision Maker.
- Safety Decision Maker (SDM): runs as a finite state machine on IGX’s isolated Functional Safety Island and issues actions such as safe‑stop signals, speed adjustments or temporary mute/unmute of onboard constraints based on integrated events.
Developers can use NVIDIA RTX Pro with NVIDIA Isaac Sim to generate synthetic camera streams for testing and hardware‑in‑the‑loop validation.
Use case: automated trailer loading (ATL)
The ATL reference shows how outside‑in safety can improve throughput while preserving safety. Inside‑out perception often struggles inside trailers, limiting autonomous forklifts. Using the VSS Blueprint to monitor the loading area, the SDM can temporarily mute onboard safety when the forklift is inside the trailer with no workers present, restoring full efficiency; when a worker enters the region of interest the system immediately reactivates safety. If SAIM detects degraded camera conditions (e.g., lighting failure or obscuration), it triggers an out‑of‑distribution event and the SDM returns the system to standard safety mode.
NVIDIA presents this approach as delivering higher throughput and more reliable safety coverage than relying on onboard sensing alone.
Inspection Lab and certification pathway
NVIDIA established the Halos AI Systems Inspection Lab, an ANAB‑accredited ISO/IEC 17020 Inspection Body, to provide a structured pathway from design to certificate for Halos‑based systems. Partners request the Lab to inspect Halos integration; NVIDIA experts assess the system against preassessed Halos elements (IGX SoM, Halos Core, Halos Applications) and issue an Inspection Certificate and Report. Partners then take that inspection documentation to third‑party certification agencies such as TÜV Rheinland, TÜV SÜD, SGS, exida, CERTX or UL Solutions for final system certification. Because Halos elements are preassessed, partners can focus certification effort on their application logic, shortening time and lowering cost to certification.
The article names Agility (maker of the Digit humanoid) as integrating NVIDIA IGX Thor and Halos OS into its human detection system and joining the Halos AI Systems Inspection Lab. The Inspection Lab membership exceeds 43 companies and includes Boston Dynamics, KION Group, Infineon, Texas Instruments, NXP Semiconductors, Lattice Semiconductor, Synapticon, Reynolds & Moore, SecEdge and FORT Robotics; recent additions include Agility, Lyte AI, Neurealm, Ouster and Peer Robotics.
How to get started
Developers can register for NVIDIA Halos Core early access and access the Halos Outside‑In Safety Blueprint on GitHub (NVIDIA/halos‑outside‑in‑safety). Two agent skills, warehouse‑deploy and halos‑deploy, are provided to automate prerequisites, NGC downloads, configuration and the NVIDIA VSS Blueprint deployment so teams can adapt the reference blueprint to their own use cases.
Summary
NVIDIA Halos for Robotics packages hardware (IGX Thor, Holoscan Sensor Bridge), a certified safety OS (Halos Core), reference safety blueprints and an ANAB‑accredited inspection pathway to help companies apply AV‑grade functional safety to robots, humanoids and AMRs while aiming to reduce time and cost to certification.
