NVIDIA used Automate in Chicago on June 22 to introduce Halos for Robotics, a safety system meant to give humanoids, autonomous mobile robots, and other physical AI machines a common architecture for working around people. Agility Robotics is the first public partner, with plans to use parts of the stack in the safety system for Digit, its humanoid robot used in logistics, manufacturing, and warehouse operations.
The announcement matters because the robotics market is moving from carefully fenced industrial automation toward machines that use AI models, distributed sensors, and edge compute in human-built spaces. In that environment, safety can no longer be treated as a last-mile checklist after a robot already works in the lab. NVIDIA is trying to make safety part of the same platform decision as the robot’s compute, operating software, perception pipeline, and certification path.
Halos for Robotics extends safety work NVIDIA previously built for autonomous vehicles into robots that move through factories, warehouses, hospitals, and eventually other less predictable spaces. The company says the system combines NVIDIA IGX Thor for industrial AI compute, Holoscan Sensor Bridge for real-time sensor connectivity, Halos OS for safety-related software, and the NVIDIA Halos AI Systems Inspection Lab for certification preparation.
Why Robot Safety Is Becoming a Platform Problem
Traditional industrial robot safety often assumes a controlled setting: fixed work cells, known paths, barriers, stop buttons, and human workers kept outside the danger zone. Humanoids and mobile robots change that premise. They are designed to share ordinary workspaces, react to changing layouts, use cameras and other sensors, and sometimes adjust behavior based on AI perception rather than a narrow preprogrammed route.
That makes the failure modes more complicated. A robot may need to slow down because a worker enters a loading area, stop because a camera feed is degraded, ignore a false obstacle inside a trailer, or fall back to onboard safety functions when external perception becomes unreliable. Those decisions involve hardware isolation, sensor trust, timing, software partitioning, AI model behavior, and documented safety logic.
NVIDIA’s pitch is that robot builders should not have to assemble those pieces from scratch for every deployment. In a technical blog, the company describes Halos as a three-layer system: platform safety at the hardware and sensor level, Halos OS for safety software, and ecosystem safety through inspection and certification support.
What Is Inside NVIDIA Halos for Robotics
At the hardware layer, IGX Thor is positioned as the main industrial compute platform. NVIDIA lists up to 2,070 FP4 TFLOPs of AI performance, 14 Arm Neoverse CPU cores, and 128GB of memory. More important for this story, it includes a dedicated functional safety island, safety monitoring mechanisms, built-in test features, watchdogs, and support for separating safety functions from the main AI compute domain.
Holoscan Sensor Bridge extends that safety chain to cameras and other devices over Ethernet. NVIDIA says the bridge can use ConnectX RDMA and GPU Direct for low-latency sensor streaming, scale to hundreds of sensors, and support secure data flow through MACsec, device authentication, watermarking, and an IEC 61508 SIL 2 safety protocol. In plain terms, Halos is not only about what the robot thinks. It is also about whether the robot can trust the sensor data it is acting on.
The software layer is Halos OS. Its Halos Core configuration supports Linux, or Linux plus QNX OS for Safety 8.0 when stronger partitioning is needed. In the QNX setup, NVIDIA uses a hypervisor to separate a Linux virtual machine for AI and application workloads from a QNX virtual machine for safety-critical functions. That split is meant to keep ordinary robot intelligence from interfering with the parts of the system responsible for safe behavior.
NVIDIA is also releasing the Halos Outside-In Safety Blueprint in early access as open source. The blueprint uses external cameras, AI perception, safety monitoring, event integration, and a finite-state safety decision maker to control robot behavior from the surrounding worksite, not only from sensors mounted on the robot itself.
The Warehouse Example Shows the Real Point
The most concrete example is automated trailer loading. In NVIDIA’s reference case, external facility cameras watch a loading area and a trailer entrance. A perception pipeline tracks people, forklifts, and regions of interest, while a safety monitor looks for conditions that can make the AI outputs unreliable, such as camera blockage, connectivity loss, image anomalies, or lighting changes that push the model outside its expected input distribution.
If the system confirms that a forklift is inside a trailer and no worker is in the loading area, the safety decision maker can temporarily mute some onboard safety constraints so the forklift does not crawl or stop unnecessarily in a tight trailer space. If a worker enters the area, or if the camera conditions degrade, the system restores the safer operating mode.
That example explains why NVIDIA is framing Halos as both a safety and productivity platform. The goal is not simply to stop robots whenever uncertainty appears. It is to give robot deployments enough validated context to keep working when conditions are safe, and to fall back when perception quality or human proximity makes that confidence disappear.
Agility Puts Halos Into a Real Humanoid Deployment Path
Agility Robotics gives the announcement a practical anchor. The company is integrating NVIDIA IGX Thor and Halos Core into Digit’s proprietary safe human detection system, and it will participate in the NVIDIA Halos AI Systems Inspection Lab. NVIDIA says Digit deployments involve customers including Amazon, GXO, Schaeffler, and Toyota Motor Manufacturing Canada.
Digit is a useful test case because it is not a robot confined to one repetitive arm motion. It is built to move through facilities designed for people, carry totes, and handle material movement where fixed automation can be expensive or awkward. NVIDIA’s earlier Agility case study describes Digit using simulation-first training in Isaac Sim and Isaac Lab, with warehouse and manufacturing deployments at GXO and Schaeffler.
Halos does not make Digit universally safe by itself. It gives Agility a preassessed safety foundation, compute layer, and inspection path that can be adapted into its own system. For customers evaluating humanoids, that distinction matters. The question is not whether a robot can carry a bin in a demo. It is whether the robot’s perception, control, software isolation, degraded-sensor behavior, cybersecurity protections, and certification documentation are mature enough for a live worksite.
Certification May Be the Hardest Part
The Halos AI Systems Inspection Lab may be as important as the hardware. NVIDIA says the lab is accredited by the ANSI National Accreditation Board as an ISO/IEC 17020 inspection body for functional and AI safety for physical AI. Its role is to inspect how partners integrate Halos elements, then provide reports that can be used with certification bodies such as TUV Rheinland, TUV SUD, UL Solutions, exida, SGS, and CertX.
The relevant standards include IEC 61508 for functional safety, ISO 13849 for safety-related control systems, and ISO/IEC TR 5469 for AI-related functional safety considerations. These are not small details. Robots that work around people need a safety case that can survive procurement reviews, insurance questions, workplace rules, and regulator attention. A credible certification path can shorten the distance between a promising pilot and a system a manufacturer or warehouse operator is willing to scale.
NVIDIA says Halos Core for IGX is available in early access for registered developers in Linux and Linux plus QNX configurations. The Outside-In Safety Blueprint is also available in early access on GitHub. That means this is still partly a developer and partner ecosystem story, not a finished plug-and-play safety label for every robot.
What To Watch Next
The next test is adoption beyond NVIDIA’s first wave of partners. Halos will become more consequential if robot makers, sensor suppliers, system integrators, and certification bodies treat it as a common safety foundation rather than another vendor-specific stack. NVIDIA listed partners across software, embedded systems, sensors, silicon, industrial applications, and assessment agencies, including QNX, FreeRTOS, Advantech, Infineon, NXP, Texas Instruments, FORT Robotics, KION Group, and others.
The other test is whether Halos can handle the messy edge cases that make physical AI different from cloud AI. A warehouse camera can be blocked. A worker can enter the wrong zone. A robot can move from a mapped aisle into a temporary layout. An AI perception model can become less reliable because lighting, clutter, or sensor placement changed. Safety systems have to recognize those conditions quickly enough to act, while keeping operations useful enough that companies do not disable safeguards in pursuit of throughput.
NVIDIA is betting that robot safety will become a platform layer, much like accelerated compute became a platform layer for AI training and inference. If that bet is right, the companies that define the safety stack for physical AI may shape not only which robots get deployed, but which deployments businesses and regulators are willing to trust.
