Gecko Robotics announced on September 28 that it is collaborating with NVIDIA on the Open Agent Safety Platform. Gecko says it is using NVIDIA OpenShell to explore enforceable runtime boundaries for AI-powered robots, with developers defining permitted and prohibited actions.
The company describes an independent enforcement layer between an AI agent and hardware in its Komodo system. This is a company announcement about an engineering approach. It provides no independent certification or comprehensive performance results establishing safety across operating conditions, and it should not be read as a guarantee that autonomous equipment cannot act incorrectly.
NAOAT analysis: Specify the action that must be blocked
An industrial buyer should translate a broad claim about control into a concrete test. What command is prohibited, which component rejects it, and what happens to the machine after rejection? That sequence is easier to assess than an assurance that an AI model has been instructed to behave safely.
Permissions also need a defined scope. Reading an inspection record, changing a work order and commanding movement are different capabilities. A deployment review should identify the authority attached to each connection and establish who can change it. An approval to collect data should not silently become permission to control equipment.
The same review should examine failure conditions: loss of communications, stale instructions, conflicting commands and unavailable authorization services. Teams should document the expected behavior for each case and how an operator can regain control without depending on the agent’s cooperation.
Finally, evidence should connect the software boundary to the actual equipment configuration being purchased. A demonstration on one system does not automatically establish behavior on another. Buyers can ask for test records, version information and an explicit division of responsibilities between the robot supplier, the software provider and the operating team. That makes a promising control architecture reviewable in a specific workplace.