Gecko Robotics and Nvidia Build Guardrails for Robot Agents

Gecko Robotics CEO Jake Loosararian said on Bloomberg Tech that his company is working with Nvidia on an open agent safety platform that sets boundaries around AI agents operating robots in military and industrial settings, arguing humans must stay in control as autonomous systems move into the physical world.

The safety debate around artificial intelligence has spent three years inside a browser window. Gecko Robotics CEO Jake Loosararian used a Bloomberg Tech appearance on Sept. 28 to argue that the argument is about to move outdoors, onto factory floors, into refineries and naval yards, and into places where a mistaken output is not a bad paragraph but a damaged asset or an injured worker.

Loosararian told Ed Ludlow that the next phase of AI safety is about keeping humans in control as autonomous systems move into the physical world. He described work with Nvidia on an open agent safety platform designed to set boundaries around AI agents that operate robots, covering military as well as industrial settings. His core claim was that safety and speed of deployment are not opposing forces — that a system can be fenced in and still be shipped quickly.

What an agent safety platform actually has to do

The phrase needs unpacking. An AI agent, in current usage, is a model given the authority to take actions rather than just produce text: call a tool, query a database, move a machine. In software, the blast radius of a bad agent action is usually a wrong record or an unwanted email. In physical infrastructure, an agent that can command a robot is an agent that can command torque, pressure, position and motion.

A safety layer for that class of system has to answer questions that chatbot guardrails never confront. Which actions is this agent permitted to initiate at all? Within what envelope of speed, force and location? What conditions force it to stop and hand back to a person? Who is the person, and how fast can they intervene? What is the immutable record of what the agent decided and why?

That is what "setting boundaries" means in the context Loosararian described: not filtering the language a model produces, but constraining the set of commands it is allowed to issue to hardware, and preserving a human decision point above it. The word that matters in the description of the joint effort is open. An open platform implies published interfaces and shared definitions rather than a proprietary black box — which is the only structure under which a defense customer, an industrial operator and a chip vendor can audit the same behavior.

Why Gecko and Nvidia are the pairing

Gecko Robotics, which is privately held, built its business on robots that climb and inspect industrial assets — boilers, tanks, hulls, pressure vessels — and on the software layer that turns those inspections into a structural picture of what is degrading and where. That gives the company something most AI safety discussions lack: a long record of machines operating in environments where the consequence of error is physical and the customer is a plant manager or a naval program office, not a developer.

Nvidia supplies the compute and the robotics software stack underneath a large share of the industry's physical-AI work. A safety framework that runs on the platform most robot developers already target has a structural distribution advantage over one that does not. Nvidia shares traded at 230.18, up 2.27% from a previous close of 225.07, as of 17:43 GMT on Sept. 28, 2026, with a session range of 226.50 to 233.21 — a gain on a day when the broad market slipped. The Bloomberg Technology interview did not attach commercial terms to the collaboration.

The military clause is the hard part

Extending the platform to military settings raises the stakes on every design decision. Defense procurement has its own doctrine on human control of autonomous systems, and vendors selling into that market must demonstrate, not assert, where the human sits in the decision chain. A published, inspectable boundary layer is easier to defend in a program review than a vendor's internal policy document.

It also creates an alignment of interest that is rarer than it sounds. Industrial buyers want the same thing defense buyers want — provable limits, auditable logs, a stop that works — so a framework built to satisfy the harder customer tends to satisfy the easier one. That is the commercial logic behind covering both in a single platform rather than shipping two.

Loosararian's assertion that safety and rapid deployment are compatible is worth reading as a sales argument as much as a philosophical one. In regulated industrial environments, the thing that actually delays a deployment is not the model's capability but the operator's inability to prove what the system will and will not do. A credible boundary layer, on that reading, is an accelerant: it converts an open-ended risk question into a bounded one that a safety officer can sign.

Where this sits in the standards race

There is no settled standard for constraining agents that drive machinery. Functional safety standards exist for industrial automation, and they were written for deterministic control systems — logic that does the same thing every time. A learned policy that reasons about a situation does not fit that template cleanly. The gap between the two is precisely the territory the Gecko–Nvidia effort is aiming at, and it is territory several constituencies want to define: chip and software vendors, robot makers, industrial insurers, defense program offices and regulators.

Whoever publishes the interface that others build against tends to set the terms. That is the strategic prize in an open platform, and it is why the announcement matters beyond either company's immediate revenue.

What to watch from here

Three things will tell you whether this is a framework or an announcement. First, publication: does a specification, reference implementation or test suite appear that outside parties can inspect and build against? Second, adoption by companies that compete with Gecko — an open platform used only by its authors is a product, not a standard. Third, whether procurement language starts citing it; the moment an industrial or defense buyer writes a boundary-layer requirement into a contract, the framework has real economic weight.

The market backdrop for the conversation was mildly negative. The S&P 500 tracker (SPY) was at $767.27, down 0.53%; the Nasdaq 100 tracker (QQQ) at $738.08, down 0.86%; and the Dow tracker (DIA) at $515.43, down 0.40%, all as of 17:43 GMT. Physical AI is a capital-intensive bet on a multi-year horizon, and none of these sessions decide it. What decides it is whether operators of heavy assets trust an agent enough to let it touch the machine — which is exactly the trust the platform is designed to manufacture.

Key facts

• Who spoke: Gecko Robotics CEO Jake Loosararian, on Bloomberg Tech with Ed Ludlow, Sept. 28, 2026
• The project: An open agent safety platform built with Nvidia to set boundaries around AI agents operating robots
• Scope: Military and industrial settings
• NVDA price: 230.18, +2.27% (prev close 225.07), as of 17:43 GMT Sept. 28, 2026

Frequently asked questions

What is an AI agent safety platform?

It is a control layer that sits between an AI agent and the hardware it can command, defining which actions the agent may take, under what limits, and when it must stop and return control to a person. Unlike chatbot content filters, it constrains physical commands rather than language, and it keeps an auditable record of decisions.

What did Jake Loosararian say on Bloomberg Tech?

The Gecko Robotics CEO said the next phase of AI safety is about keeping humans in control as autonomous systems move into the physical world. He described work with Nvidia on an open agent safety platform that sets boundaries around AI agents operating robots in military and industrial settings, and argued safety and rapid deployment are compatible.

Is Gecko Robotics publicly traded?

No. Gecko Robotics is a privately held company, so there is no listed share price or ticker for investors to trade. Its business centers on robots that inspect industrial assets such as boilers, tanks and hulls, plus the software that converts those inspections into structural assessments for operators.

Why does Nvidia matter to this effort?

Nvidia supplies much of the compute and robotics software that physical-AI developers already build on. A safety framework that runs natively on that stack reaches developers without requiring them to change platforms, which gives it a distribution advantage over a standard proposed by a party without that installed base.

What does 'open' mean for this kind of platform?

An open platform generally implies published interfaces, shared definitions and implementations that outside parties can inspect, test and build against, rather than a proprietary system whose behavior only the vendor can verify. That inspectability is what lets a defense program office or an industrial safety officer independently confirm the limits.

How did Nvidia shares trade on the day of the interview?

Nvidia was quoted at 230.18, up 2.27% from a previous close of 225.07, with a session range of 226.50 to 233.21, as of 17:43 GMT on Sept. 28, 2026. The broader market was weaker that session: the S&P 500 tracker fell 0.53% and the Nasdaq 100 tracker fell 0.86%.

Sources

• Gecko CEO: Physical AI Needs Humans in Control — Bloomberg Technology