Figure AI, the humanoid robotics company led by CEO Brett Adcock, has joined NVIDIA's newly announced Open Agent Safety Platform, signaling that the industry is moving beyond hardware development to tackle the harder problem: ensuring robots can be trusted around people. NVIDIA CEO Jensen Huang announced the platform on September 28, revealing that over 100 industry partners are involved in the effort to create enforceable boundaries around increasingly capable AI agents.

Why Does Robot Safety Matter Right Now?

The timing of Figure's participation is significant. As humanoid robots transition from research labs and controlled factory settings into homes and workplaces, the stakes for safety have shifted. Unlike a robot operating alone in a warehouse, a humanoid working alongside people or in a home environment must be predictable, controllable, and trustworthy even when things go wrong. Adcock framed the collaboration directly around this reality, stating that humanoid robots will soon operate in these environments and "have to be safe and trusted".

The challenge isn't just about preventing a robot from breaking something. It's about ensuring that when an AI system makes a poor decision, the robot's physical actions are constrained by systems that operate independently of the AI itself. This is where NVIDIA's platform comes in.

How Does NVIDIA's Safety Platform Actually Work?

• OpenShell Runtime: The platform includes an open-source runtime that applies controls outside the AI agent itself, isolating each agent in a sandbox with restrictions on file access, system calls, and network connections. This prevents unauthorized actions even if the AI model attempts them.
• Policy Verification: Before new access permissions are applied, the system flags them for review, ensuring that policy changes don't accidentally grant dangerous capabilities without human oversight.
• Sentry Hardware Layer: An optional monitoring and enforcement layer runs on separate BlueField hardware, designed to keep safety monitoring beyond the reach of the agent and its host software. NVIDIA says it can quarantine agents in milliseconds if needed.

The key architectural principle is straightforward but powerful: the system responsible for choosing an action should not be the only system responsible for policing it. In practical terms, this means an AI agent can propose an action, but a separate, independent system verifies whether that action is safe before the robot executes it.

What Does This Mean for Humanoid Robots Specifically?

Figure appears to be the only dedicated humanoid manufacturer shown in NVIDIA's published ecosystem graphic, though the platform includes other robotics companies like Gecko Robotics and Skild AI. However, Figure has not yet disclosed how it will integrate the safety platform into its own systems or which specific robots will use it.

The application is more concrete for industrial inspection work. Gecko Robotics CEO Jake Loosararian described using AI agents for inspection-path planning and diagnostics, with agents deciding where robots should gather additional information while remaining within parameters established by people. In his account, operators can intervene to rein in planned paths and actions, giving "humans in control" a practical meaning.

For humanoids operating in homes and workplaces, the principle applies but the implementation is more complex. Restricting access to a robot-control interface could prevent an unauthorized command, but it wouldn't by itself ensure that an authorized grasp uses appropriate force or that interrupting a model leaves a walking robot in a stable state. Those are additional physical-control questions that any real-world implementation would have to address.

What Happens Next?

Adcock's public endorsement confirms that Figure considers NVIDIA's initiative relevant to its own work, though his announcement does not explain what the company will contribute or which systems will use it. NVIDIA CEO Jensen Huang framed the effort as the beginning of a shared foundation for trustworthy agent systems, arguing that confidence in how AI is deployed will help unlock its economic potential.

The immediate news is Figure's participation in a broader effort to enforce limits around AI agents. The next meaningful detail will be how that work connects to Figure's actual systems and which specific risks it can measurably reduce when software decisions become physical actions. As humanoid robots move from announcements to actual deployment in real-world environments, the ability to demonstrate safety won't be optional; it will be essential to adoption.