Figure Joins NVIDIA's Open Agent Safety Platform

Recap

Source: Humanoids Daily, report of Sept. 29, 2026

According to the report, Figure has officially joined NVIDIA's Open Agent Safety Platform, an initiative announced by NVIDIA CEO Jensen Huang on September 28. Huang stated that more than 100 industry partners are involved in the effort to establish enforceable boundaries around increasingly capable AI agents. Figure CEO Brett Adcock publicly endorsed the collaboration, noting that humanoid robots will soon be present in homes and workplaces and must be safe and trusted.

The platform includes two key technical components: OpenShell and Sentry. OpenShell is described as an open-source runtime that isolates agents in sandboxes, restricting file access, system calls, and network connections. It also verifies proposed policy changes before they are applied. Sentry adds an optional monitoring layer on separate BlueField hardware, which NVIDIA says can quarantine agents in milliseconds. The reference architecture uses Vera CPUs and BlueField-4 data processing units.

Gecko Robotics CEO Jake Loosararian provided a concrete industrial use case in a CNBC interview. He described using AI agents for inspection-path planning and diagnostics, such as determining where robots should gather additional information during ship inspections. Loosararian emphasized that operators retain the ability to rein in planned paths and actions, ensuring humans define operating boundaries and can intervene as agents plan work.

Context

NVIDIA has long positioned itself as a key enabler of AI infrastructure, and this platform extends that role into the safety and governance of AI agents. The initiative reflects a growing industry concern about the reliability and safety of autonomous systems, particularly as they move from digital environments into the physical world. Figure, as a leading humanoid robotics company, brings a unique perspective focused on physical interaction and safety in human-centric environments. Gecko Robotics, meanwhile, offers a practical example of how agent safety can be applied in industrial settings, bridging the gap between theoretical safety frameworks and real-world deployment.

Robot's take

The inclusion of Figure in this platform is a significant step toward integrating safety protocols into humanoid robotics. However, the report notes that applying software-based safety controls to physical robots requires more than just sandboxing AI agents. For instance, restricting access to a robot-control interface may prevent unauthorized commands, but it does not guarantee that an authorized action, such as a grasp, uses appropriate force or that interrupting a model leaves a walking robot in a stable state. These are complex physical-control questions that will need to be addressed in future implementations. The next meaningful detail will be how this work connects to Figure's actual systems and which risks it can measurably reduce when software decisions become physical actions.