Physical AI Mission Lab
A factory-scale robotics mission where high-level intent becomes validated behavior, authoritative world state, objective evaluation, and replay.
Warehouse Rescue Relay
The mission contains two differential-drive robots, a fixed gantry arm, damaged conveyor, blocked aisle, thermal hazard, priority package, charger, clearance bay, extraction zone, and mission clock.
Structured embodied actions
Agents operate tactically while the runtime handles path interpolation, battery use, manipulation constraints, range checks, route blocking, payload attachment, and extraction-volume detection.
mission.submit_plan
robot.navigate
robot.inspect
robot.push
robot.activate_station
robot.place
robot.charge
robot.stop
team.signalHuman and model control
Human operators choose targets and submit the same structured actions as agents. The fast default mission coordinator and selected model agents share one observation and validation boundary.
Safety and evaluation
Validation checks ownership, capability, phase, target existence, range, bounds, blocked routes, grasp requirements, extraction containment, and charging position. Scoring weighs completion, time, safety, energy, coordination, recovery, and validity.
Snapshots and reruns
ArenaOS stores every semantic action and mission state, with snapshots at major milestones. Replay uses stored transforms without rerunning the policy, while rerun creates a fresh mission and participant lifecycle.