The Halo Protocol

At a GXO warehouse near Memphis, an Agility Robotics Digit unit lifts a tote off a conveyor and turns to place it on a pallet — and less than a meter away, a human coworker crosses its path without breaking stride. Neither of them looks twice. That unremarkable moment is the entire point of a piece of engineering NVIDIA unveiled on June 23, 2026: Halos for Robotics, the industry's first full-stack safety system built specifically for machines that are about to share physical space with millions of people who never trained for it.

For a decade, industrial robots have solved the human-proximity problem by avoidance: cages, light curtains, and floor markings that keep flesh and steel apart. Humanoid robots break that solution by design. Their entire commercial pitch is that they can work in spaces built for people — walking the same aisles, reaching the same shelves, riding the same freight elevators. Figure AI is now producing a humanoid roughly once per hour at its BotQ factory. Amazon has crossed a million deployed robots company-wide. Unitree is targeting 10,000 to 20,000 units of its R1 humanoid this year alone. The capability curve has bent upward fast. The certification curve, until this summer, had barely moved.

Why Cages Don't Work Anymore

Traditional industrial-robot safety is, at its core, a geometry problem: keep the robot's working envelope and the human's walking envelope from ever overlapping, and enforce that separation with a hard barrier. It's a solvable, well-understood engineering problem, which is exactly why robots have spent forty years fenced off from the people they work near. A humanoid picking totes beside a person, riding a service elevator, or walking a hospital corridor cannot be fenced. Its entire economic value depends on operating in unmodified human space, at human pace, without a safety cage eating the floor space it was bought to save.

That shift turns robot safety from a geometry problem into a perception and judgment problem — the robot has to continuously, correctly infer where every nearby human is, where they are about to be, and what its own body will do in the next fraction of a second, then act on that inference fast enough to prevent contact it doesn't intend. No humanoid robot has yet completed independent third-party certification proving it can do that reliably under the same rigor applied to elevators, aircraft autopilots, or industrial robot arms. Halos for Robotics is NVIDIA's attempt to build the infrastructure that certification will eventually run on, before deployment scale outruns it.

What's Actually Inside the Stack

Halos for Robotics isn't a single product so much as a certification-grade architecture assembled from pieces NVIDIA already had lying around from a different industry. The company spent years building Halos as a safety layer for self-driving cars — a domain that already forced it to solve the problem of trusting an AI system's split-second decisions enough to put a regulator's stamp on them. Halos for Robotics repackages that infrastructure for a target that walks on two legs instead of rolling on four wheels.

The stack has three broad layers. At the base sits NVIDIA IGX Thor, an industrial-grade onboard computer that pairs AI processing with hardware-level safety functions, so the chip doing the robot's thinking is architecturally separated from the circuitry that can force a safe stop. Above that runs Halos Core, the software layer that governs safety-related behavior — the rules for how a robot interprets uncertainty, how conservatively it moves near an unpredictable human, and when it must yield control to a hard-coded safe state rather than trust its own judgment. And wrapping the whole system is what NVIDIA calls the Outside-In Safety Blueprint: an open-source framework that uses cameras mounted on the facility itself — not just the robot's own sensors — to watch the wider environment and dynamically tighten or loosen a robot's safety constraints depending on who and what is nearby. A robot alone in an aisle can move at full working speed; the moment an external camera flags a person entering that aisle, Outside-In can throttle it down before the robot's own onboard sensors even register the change.

Agility Robotics Goes First

Agility Robotics is the first company to build Halos for Robotics into a commercial machine, integrating NVIDIA IGX Thor and Halos Core into the proprietary human-detection system already running on its Digit humanoid. Digit is built for exactly the environments where the stakes of getting proximity wrong are highest: logistics floors, manufacturing lines, and warehouses, with customers that already include Amazon, GXO, Schaeffler, and Toyota Motor Manufacturing Canada. Agility and NVIDIA are now running Digit's safety-related software, AI components, and cybersecurity protections through a dedicated testing lab aimed at meeting IEC 61508, ISO 13849, and ISO/IEC TR 5469 — three overlapping standards covering functional safety of electronic control systems, safety-rated machinery controls, and, notably, the newest of the three, AI specifically operating inside a safety-related system.

That third standard is the one worth sitting with. ISO/IEC TR 5469 exists because the older safety frameworks were written for deterministic machinery — systems that do the same calculable thing every time given the same input. A neural network making a judgment call about a moving human is not that kind of system, and regulators have spent the last several years building the vocabulary to certify it anyway. Halos for Robotics is, in effect, NVIDIA's bet that it can get there before a serious real-world incident forces the industry to build that vocabulary the hard way.

The Business Case Hiding Inside the Safety Case

It would be a mistake to read Halos purely as altruism. Every serious analyst tracking humanoid robotics has converged on the same conclusion: the industry's transition from pilot programs to platform-scale deployment is gated less by whether a robot can do the job and more by whether an insurer, a factory safety officer, or a hospital risk-management team will sign off on it working next to staff. Morgan Stanley has sharply raised its shipment forecasts for Chinese humanoid manufacturers this year specifically because early commercialization is now underway at scale — but scale multiplies exposure to a single bad headline. One widely reported injury involving an uncertified humanoid could freeze enterprise buying decisions across the entire sector for a year or more, the way a handful of early self-driving car incidents slowed robotaxi rollouts in multiple U.S. cities.

NVIDIA has already lived through that exact dynamic once, in autonomous vehicles, which is precisely why Halos for Robotics is built on the AV safety architecture rather than starting from scratch. Selling the chips and software that make humanoid robots capable is a much bigger and more durable business if the chips and software that make them provably safe come from the same vendor, packaged together, so that a customer buying compute for a warehouse robot is simultaneously buying the certification pathway that lets it operate there legally and profitably.

What Happens Next

The next few months put this infrastructure under real observation. The World Artificial Intelligence Conference in Shanghai this July, the World Robot Conference in August, and the second World Humanoid Robot Games later that month will all put dozens of competing humanoid platforms in close proximity to crowds of untrained bystanders — arguably the largest uncontrolled proximity test the technology has faced. None of those events require Halos-equivalent certification to participate, which means the safety gap this article describes will be visible in real time to anyone watching a robot brush past a spectator on a conference floor.

The more consequential test is quieter: whether Agility's certification effort with NVIDIA actually clears IEC 61508 and its companion standards on the timeline implied by the announcement, and whether competitors — Figure, Boston Dynamics, Unitree, China's AgiBot — adopt an equivalent framework or race ahead without one. Functional safety certification is slow by design; it exists specifically to resist the pressure to ship fast. Whether that friction becomes the industry's shared floor or gets treated as one vendor's optional add-on will do more to determine how fast humanoid robots actually enter daily human life than any benchmark measuring how well one can fold a shirt or climb a stair.

Robots are, for the first time, about to become common enough that most people will eventually stand within arm's reach of one without choosing to. The question Halos for Robotics is really asking isn't whether the machines are capable. It's whether anyone can prove, to a regulator's satisfaction, that they know the difference between a person and an obstacle quickly enough for the distinction to matter.

Sources