Every robot move,
proven safe before it happens.
An NVIDIA Nemotron planner proposes. A deterministic governance gate decides. Every decision leaves a signed, hash-chained receipt. The AI can be brilliant or wrong; the robot only ever does what the rules allow.
How it works
Plans the mission and replans after a denial. Fast Nemotron models handle per-step calls. Every output is only a proposal.
Geofence, zone speed, people and vehicles, racking, hazard zones, perception confidence, operator escalation. No model in the loop.
Only approved moves run. Every denial sends STOP, so the robot never continues toward an old target.
Every decision is chained and signed with the rules version it was judged under. Anyone can verify it later.
Untrusted by default
Every outside input can add caution. None can remove it.
Nemotron proposes; the gate judges every move. Plans that cut through racking are re-routed along the aisles.
Nemotron 3.5 Super VL reads the robot camera. It is cross-checked with the sensors; a sighting they can't confirm can only slow the robot.
Tavily hazard guidance passes a firewall and a two-source rule. Zones can only tighten; the gate never reads the text.
Lessons come only from the gate's own records and change what the planner proposes, never what the gate allows.
Live system
What is running behind this page