A robot that cannot tell when its own perception is unreliable is a hazard. We argue for a self-trust layer that reasons about contradictions across sensors — and acts on them.
Perception is only as good as its fallback
An autonomous robot rarely fails because one sensor broke. It fails because two sensors disagree — the camera reports a clear path while the LiDAR sees an obstacle, or the IMU records motion the controller never commanded — and the fusion layer quietly averages the confusion into a bad decision.
Current stacks treat these disagreements as calibration problems to tune away. But a disagreement is also information: it is often the earliest signal that perception itself is compromised, spoofed, or failing.
A Perception Reliability Score
Our proposed self-trust layer maintains a Perception Reliability Score: a continuously updated estimate of how much the robot should believe its own senses. The score is derived from contradiction patterns across camera, LiDAR, IMU and expected physical trajectories — not from any single sensor.
What makes this different from classical sensor fusion is what the score controls. It does not merely re-weight inputs; it changes behaviour. As the score degrades, the robot escalates through a policy ladder: continue with reduced confidence, slow down, restrict its motion envelope, or come to a controlled stop.
The hard questions
Three questions shape the research agenda. First, what formal notion of self-trust captures reliability across conflicting sensor streams without collapsing into noise? Second, how fast must the cross-sensor analysis run to stay useful for real-time control — we believe under ten milliseconds is the right budget. Third, how should the policy ladder be tuned so a degrading robot is safe without becoming useless?
If those questions have good answers, a robot that can doubt itself is safer than one that never does. The full problem statement, including the evaluation scenarios we plan to publish, is on the problem board.