Deployment · Advanced · 5-9 hours
NVIDIA Halos for Robotics
NVIDIA Halos is a full-stack functional-safety system for humanoids, industrial robots, and autonomous mobile robots; Halos OS is its operating-system layer, not a general robot-development OS.
Direct answer
What is NVIDIA Halos for Robotics?
The correct name is NVIDIA Halos, with Halos OS as one layer inside the platform. Halos OS runs on NVIDIA IGX Thor and supplies safety-oriented operating-system foundations for robotics; it does not replace Isaac, GR00T, ROS 2, a robot policy, or final system certification.
Definition and scope
NVIDIA Halos is a full-stack functional-safety system for humanoids, industrial robots, and autonomous mobile robots; Halos OS is its operating-system layer, not a general robot-development OS.
The robotics stack combines IGX Thor and its Functional Safety Island, Halos Core on Linux or Linux plus QNX, safety extensions and communications, application blueprints such as Outside-In Safety, and the Halos AI Systems Inspection Lab.
Naming correction: NVIDIA Halos is the overall safety platform. Halos OS is the operating-system and safety-software layer inside it.
System architecture
How the Halos robotics stack is organized
The product name covers multiple layers. Treating all of them as one operating system hides the actual safety architecture.
1. Platform safety
NVIDIA IGX Thor supplies accelerated AI compute plus a dedicated Functional Safety Island. NVIDIA describes the platform as IEC 61508 SIL 3 capable, with hardware monitoring, redundancy, isolation, in-system tests, and error-management mechanisms. Holoscan Sensor Bridge extends authenticated, low-latency sensor and actuator connectivity toward the safety domain.
2. Halos OS
Halos Core is the safety operating-system foundation. The Linux configuration includes the Linux runtime, Safety Extensions Package, Edge Safety Link, Functional Safety Island RTOS, and Safety MCU firmware. A Linux-plus-QNX configuration adds NVIDIA Hypervisor partitioning, with Linux handling AI/application workloads and QNX hosting safety-critical functions.
3. Applications and ecosystem assurance
Halos applications include safety blueprints such as Outside-In Safety. The Halos AI Systems Inspection Lab inspects how partners integrate preassessed Halos elements and produces evidence that can support, but does not replace, final certification by an independent certification body.
How the Outside-In Safety Blueprint works
The reference workflow adds infrastructure perception around the robot rather than relying only on sensors carried by the machine.
- Infrastructure cameras feed a sensor-input and perception pipeline that detects and tracks workers, robots, regions of interest, proximity events, and virtual tripwires.
- A Safety AI Monitor checks camera blockage, connectivity loss, image anomalies, and out-of-distribution conditions that can make AI outputs unreliable.
- A Safety Event Integrator fuses observations from multiple cameras, applies confidence thresholds, and rejects stale events.
- A finite-state Safety Decision Maker runs in the isolated safety domain and can command a safe stop, reduce speed, or change the robot's safety mode.
- Isaac Sim and RTX systems can generate synthetic camera streams and hardware-in-the-loop tests before deployment.
Product boundaries
What Halos replaces—and what it does not
Standards and certification path
- IEC 61508 provides the general functional-safety framework and Safety Integrity Level concepts used for industrial systems.
- ISO 13849 addresses safety-related parts of machinery control systems and is relevant to industrial robots and automation.
- ISO/IEC TR 5469 and the developing ISO/IEC TS 22440 address AI-related functional-safety considerations referenced by NVIDIA for safety agents.
- The Halos AI Systems Inspection Lab is accredited to ISO/IEC 17020 by ANAB. An inspection report supports downstream certification; it is not identical to certification of the final robot.
Critical distinction: an inspected or assessed platform element is not the same as a certified complete robot operating in a specific facility.
Deployment status as of June 2026
- NVIDIA announced Halos for Robotics on June 22, 2026.
- Halos Core for IGX is available in early access in Linux and Linux-plus-QNX configurations.
- The Outside-In Safety Blueprint is published on GitHub under Apache-2.0, while some deployment artifacts require NVIDIA NGC access and supported hardware.
- Agility Robotics is the inaugural humanoid partner; NVIDIA says Digit will be the first production robot shipping with Halos OS.
- The automated-trailer-loading reference concept has been inspected by TUV Rheinland, but that inspection should not be generalized into certification of every downstream deployment.
Limitations and open questions
- Most evidence currently comes from NVIDIA launch material, early-access documentation, partner statements, and reference concepts rather than long-running independent fleet data.
- The full safety case depends on the robot, sensors, actuators, facility, operating design domain, application logic, and human procedures—not only the compute module or OS.
- Outside-in cameras can reduce blind spots but create new dependencies on camera placement, lighting, networking, calibration, occlusion handling, privacy, and fail-safe behavior.
- AI monitors detect selected failure conditions; they cannot prove that every unsafe or out-of-distribution condition will be recognized.
- Early-access software, NDA documentation, hardware requirements, and certification costs can limit adoption for small robotics teams.
Primary sources
Claims on this page are grounded in NVIDIA product documentation, technical material, source code, and certification-program information. Product claims should still be separated from independent deployment evidence.
Learning path
First separate robot intelligence from robot assurance: learn Isaac and GR00T for training and autonomy, ROS 2 for middleware, then study Halos as the safety foundation used when moving systems toward production.
Recommended next topics
Related research
Key papers and benchmarks
ASIMOV Benchmark for Robot Semantic Safety
ASIMOV evaluates whether robot-brain foundation models understand unsafe physical situations and safety rules.
GR00T N1: An Open Foundation Model for Generalist Humanoid Robots
GR00T N1 uses a dual-system architecture for language reasoning and continuous humanoid control.
NVIDIA Isaac GR00T N1.7
GR00T N1.7 is an open vision-language-action model for generalized humanoid manipulation skills.
Research ecosystem
Organizations connected to this layer
Organization
NVIDIA
Robot foundation models, simulation, synthetic data, edge deployment, and functional safety
View profile →Organization
Agility Robotics
Production humanoid robots for logistics and industrial workflows
View profile →Organization
Boston Dynamics
Dynamic mobile robots and manipulation
View profile →Common questions
Frequently asked questions
Is NVIDIA Halos a robotics operating system?
NVIDIA Halos is the umbrella full-stack safety platform. Halos OS is its operating-system and safety-software layer for IGX Thor; it is not a replacement for ROS 2, Isaac, or a robot policy.
Is Halos OS open source?
The Outside-In Safety Blueprint is available on GitHub under Apache-2.0. Halos Core is early-access NVIDIA software, and some packages, documentation, and NGC artifacts require registration or supported hardware.
Does using Halos certify a robot as safe?
No. Halos provides assessed elements, inspection support, and certification evidence. The complete robot and its deployment still require system-level hazard analysis, validation, and any applicable independent certification or approval.
What is inside-out versus outside-in robot safety?
Inside-out safety uses sensors carried by the robot. Outside-in safety adds infrastructure cameras and safety agents that monitor shared spaces, blind spots, zones, and degraded perception conditions.