NVIDIA Halos
Last reviewed
Sources
5 citations
Review status
Source-backed
Revision
v1 · 1,719 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
5 citations
Review status
Source-backed
Revision
v1 · 1,719 words
Add missing citations, update stale details, or suggest a clearer explanation.
NVIDIA Halos is a full-stack, comprehensive safety system developed by NVIDIA that unifies AI compute and safety across silicon, systems, software, and tools and services. It was first introduced for autonomous vehicles in 2025 and was extended to robotics and physical AI on June 22, 2026, when NVIDIA announced what it called the industry's first full-stack safety system for physical AI. Halos brings together NVIDIA's hardware and software safety solutions with its AI safety research, and it is built on more than 18,600 engineering years of autonomous vehicle safety development. Its stated goal is to give developers and system builders a unified foundation for designing, validating, deploying, and certifying safe AI-driven machines, from cloud to vehicle and, more recently, to industrial and humanoid robots.[1][2][3]
NVIDIA developed Halos to consolidate the many safety-focused technologies it had built for autonomous driving into a single named system. The company first unveiled its AI Systems Inspection Lab at the CES trade show in January 2025, and in March 2025 it formally launched NVIDIA Halos as a comprehensive safety system for autonomous vehicles (AVs). The AV version of Halos was described as unifying safety elements across vehicle architecture, AI models, chips, software, tools, and services to support the safe development and deployment of autonomous vehicles from the data center to the car.[4][5]
NVIDIA framed Halos around three focus areas that carried forward into the robotics version:
The AV system relied on three classes of computers: NVIDIA DGX systems for model training, NVIDIA Omniverse and NVIDIA Cosmos for simulation, and NVIDIA DRIVE AGX for in-vehicle deployment. NVIDIA's DriveOS 6.0 operating system conforms with the ISO 26262 automotive functional safety standard at Automotive Safety Integrity Level D (ASIL D), the highest integrity level defined by that standard, and the DRIVE AGX Hyperion platform connects the SoC, DriveOS, and sensors in a reference electronic control unit architecture. When Halos for AVs launched, NVIDIA cited more than 15,000 engineering years invested in vehicle safety and more than 10,000 hours of contributions to international standards committees; by the time of the robotics expansion, the cited engineering investment had grown to more than 18,600 engineering years.[4][3]
On June 22, 2026, NVIDIA announced Halos for Robotics, which it described as the industry's first full-stack, comprehensive safety system for robotics and physical AI that unifies AI compute and safety. The expansion extends NVIDIA's autonomous vehicle safety foundation to robotic systems that operate in dynamic industrial environments and increasingly work alongside people. Deepu Talla, vice president of robotics and edge AI at NVIDIA, said: "With NVIDIA Halos for Robotics, developers and system builders can harness NVIDIA's proven autonomous vehicle safety foundation to develop safer robots faster."[1][2]
Halos for Robotics connects the layers needed to build, validate, and deploy robotic systems, spanning AI compute, system software, sensor data, safety applications, and inspection. The system is organized into three layers.[1][3]
| Layer | Purpose | Key components |
|---|---|---|
| Compute and connectivity | Industrial-grade AI compute with built-in safety and real-time sensor connectivity | NVIDIA IGX Thor; NVIDIA Holoscan Sensor Bridge |
| Software stack | Safety-related operating functions and external safety supervision | NVIDIA Halos OS; Halos Core; NVIDIA Halos Outside-In Safety Blueprint |
| Validation and certification | Independent assessment of functional and AI safety to prepare for third-party certification | NVIDIA Halos AI Systems Inspection Lab |
The foundation of Halos for Robotics is NVIDIA IGX Thor, an industrial-grade AI compute platform built on NVIDIA's Thor system-on-chip (related to NVIDIA's Jetson Thor robotics compute line and the NVIDIA Blackwell architecture). NVIDIA states that IGX Thor delivers up to 2,070 FP4 TFLOPS of AI performance, 14 Arm Neoverse CPU cores, and 128 GB of memory at 273 GB/s of bandwidth. For functional safety, IGX Thor includes an IEC 61508 SIL 3 capable Functional Safety Island (FSI), a dedicated safety processor with up to 12,000 DMIPS and its own I/O, power, and clocks that is physically isolated from the main compute domain. NVIDIA describes over 22,000 safety mechanisms providing diagnostic coverage across the SoC, along with in-system test (logic and memory built-in self-test) for latent fault coverage.[3]
The NVIDIA Holoscan Sensor Bridge provides real-time sensor connectivity for robotics and safety workloads. It uses ConnectX RDMA and GPUDirect to enable low-latency sensor streaming, supports an end-to-end IEC 61508 SIL 2 safety protocol with watermarking and camera testing, and offers MACsec for device authentication and encrypted data flow.[3]
NVIDIA Halos OS is the robotics safety software stack that runs on IGX Thor. Its core element, Halos Core, provides safety-related operating functions and is offered in two configurations. The Linux configuration provides a Linux runtime for application and compute workloads together with a Safety Extensions Package for hardware error collection and dispatch, the Edge Safety Link safety communication protocol, and real-time operating system firmware for the Functional Safety Island and the safety microcontroller. The Linux and QNX configuration adds an NVIDIA hypervisor layer that partitions IGX into isolated virtual machines: a Linux VM for AI and application workloads and a QNX VM for safety-critical functions.[3]
The NVIDIA Halos Outside-In Safety Blueprint uses external infrastructure cameras and AI agents to monitor a robot's surroundings and control its behavior, adding a layer of supervision independent of the robot's own onboard perception. According to NVIDIA, the blueprint is built from several cooperating components: a Sensor Input Processing Pipeline that ingests camera streams and converts them into actionable events; a Safety AI Monitor that continuously checks the perception pipeline for conditions that could compromise detection accuracy, such as out-of-distribution inputs, camera blockage, connectivity drops, and image anomalies; a Safety Event Integrator that fuses events from multiple camera perspectives and requires a high confidence threshold; and a Safety Decision Maker that runs a finite state machine on the IGX Functional Safety Island to output safety control signals.[3]
The NVIDIA Halos AI Systems Inspection Lab is an inspection program accredited by the ANSI National Accreditation Board (ANAB) as an ISO/IEC 17020 inspection body. NVIDIA describes it as the first worldwide program accredited for AI and functional safety across both autonomous vehicles and robotics. Partners submit products to be inspected against preassessed Halos stack elements, such as the IGX system-on-module, Halos Core, and Halos applications, and receive certificates that help them pursue final system certification through third-party certification bodies. The original AV inspection lab, announced at CES 2025, was described as the first program accredited by ANAB for an inspection plan integrating functional safety, cybersecurity, AI safety, and regulations into a unified framework.[3][5]
Halos is designed to help partners meet established functional safety and AI safety standards. The robotics version emphasizes industrial and machinery standards, while the automotive version centers on ISO 26262.[1][3]
| Standard | Domain | Description |
|---|---|---|
| IEC 61508 | General functional safety | The foundational international standard for the functional safety of electrical, electronic, and programmable electronic safety-related systems, defining Safety Integrity Levels (SIL 1 to SIL 4). |
| ISO 13849 | Machinery safety | Specifies safety requirements and guidance for the design of safety-related parts of machinery control systems. |
| ISO/IEC TR 5469 | AI functional safety | A technical report addressing the use of artificial intelligence within functional safety systems and the properties AI must exhibit to be used safely. |
| ISO 26262 | Automotive functional safety | The automotive functional safety standard, defining Automotive Safety Integrity Levels (ASIL A to ASIL D); NVIDIA's DriveOS conforms with the most stringent level, ASIL D. |
Agility Robotics is the first adopter of Halos for Robotics, integrating Halos components into its Digit humanoid robot for warehouse, logistics, and manufacturing applications. NVIDIA notes that Digit serves customers including Amazon, GXO, Schaeffler, and Toyota Motor Manufacturing Canada.[1][2]
NVIDIA reports an ecosystem of more than 40 organizations participating in Halos, including more than 43 members of the AI Systems Inspection Lab. The partners span several categories.[1][3]
| Category | Partners |
|---|---|
| Software and operating systems | Acontis; FreeRTOS; QNX |
| Embedded systems | Advantech; NexCobot |
| Semiconductors and sensors | Infineon; NXP Semiconductors; STMicroelectronics; Texas Instruments |
| Industrial applications | FORT Robotics; Inventec; KION Group; Lyte AI; Neurealm |
| Certification bodies | TUV Rheinland; TUV SUD; UL Solutions; exida; SGS; CertX |
| New Inspection Lab members | Agility Robotics; Lyte; Ouster; Peer Robotics |
NVIDIA has also cited robot makers such as Boston Dynamics among the broader set of companies engaging with the Inspection Lab program.[3]
Halos Core for NVIDIA IGX is available in early access for registered Linux developers, with a Linux plus QNX OS for Safety 8.0 configuration for safety-critical deployments. The NVIDIA Halos Outside-In Safety Blueprint is available in early access as open-source software on GitHub.[1][3]
The robotics expansion of Halos arrives as humanoid robots and other autonomous machines move out of caged, controlled cells and into workplaces where they must share space with people, vehicles, and equipment. By packaging safety-assessed silicon, a safety operating system, an external monitoring blueprint, and an accredited inspection program into one branded stack, NVIDIA positions Halos as a way to shorten the path from prototype to certifiable, deployable physical AI. Industry coverage described it as a notable attempt to bring automotive-grade functional safety discipline to the fast-moving humanoid and industrial robotics sector.[1][3]