Autonomy hits an edge case
The local stack reaches a state it was not built to resolve.
AI Supervisor for robots
Give every robot a supervisor.
Fathom handles the moments when autonomy runs out. Send the scene, robot state, and allowed skills. Get one bounded decision your onboard stack can accept or reject.
Exceptions, not motor controlBounded skillsHuman escalation
blocked-door-017needs judgment
01
Robot reports an impasse
Frame, state, task, and recent attempts.
blocked02
Fathom checks the scene
Reasons within the skills you approved.
inspect03
Supervisor chooses
Retry, reroute, hold, or ask a person.
reroute04
Robot validates locally
Execution and safety remain onboard.
acceptedThe problem
Robots stop. People take the call.
A blocked door or dropped item can turn a robot into a support ticket. Human attention becomes the ceiling on fleet scale.
A person studies the scene
They understand the problem and choose the next safe skill.
Fathom takes the first call
A larger model resolves routine exceptions and escalates the rest.
The Supervisor
Judgment above the control loop.
Fathom does not drive joints. It sees the same evidence a remote operator sees and chooses from a small set of approved actions.
Understand, decide, escalate.
- SeeFrames, state, history, and current task.
- ChooseOne approved skill with bounded parameters.
- LearnLink the decision to the later outcome.
Control, validate, stay safe.
- ReflexesFast reactions and local planning.
- AuthorityWatchdogs, safety gates, and actuators.
- FallbackReject stale or invalid requests.
The moat
Every exception becomes evidence.
The model is replaceable. The hard asset is a growing record of what robots saw, what the Supervisor chose, what local safety accepted, and what happened next.
InputFrozen evidence
The exact scene and robot context.
ContractApproved skills
The only choices the model may return.
OutcomeResolved or escalated
A business result, not a vague score.
ImprovePromote the better revision
Keep the same Supervisor name.
Start with one stuck case