Introspection for Reliable Localisation

Why it matters

Visual place recognition systems output a best match for every query, even when that match is wrong. Introspection gives the robot a sense of its own reliability: whether a match should be accepted, which technique or map to use, and when to fall back to a safer behaviour.

Approach

• Calibrated operating points. Dhyey Manish Rajani’s work on uncertainty estimation selects matching thresholds that transfer reliably to new environments, so that a target precision set in the lab still holds in deployment.
• Multi-modal sensor introspection. Margaux Edwards is developing methods that monitor several sensors at once to predict, and recover from, place recognition failures.
• Predicting performance before it matters. We predict whether sequence matching will help for a given query, choose map densities that perform well locally, and select the most complementary technique from an ensemble.

Introspection is a core part of my ARC DECRA Fellowship on adaptive and efficient robot positioning.