ENCP: Episode-Normalized Conformal Prediction for Vision-and-Language Navigation
A paper proposes Episode-Normalized Conformal Prediction (ENCP) for uncertainty estimation in Vision-and-Language Navigation (VLN). ENCP rescales a nonconformity score by the policy's residual confidence and calibrates one maximum score per episode. Under exchangeable calibration and test episodes, ENCP covers the ground truth at every step with probability at least 1 - α. Across four VLN policies and three nonconformity scores on R2R and REVERIE datasets, ENCP meets all reported empirical step-coverage targets on the seen-to-unseen evaluation.
The paper introduces ENCP, a conformal prediction method tailored for VLN that provides episode-level coverage guarantees while allowing dependence among steps. It is evaluated on R2R and REVERIE datasets with four VLN policies and three nonconformity scores, meeting empirical step-coverage targets.
ENCP addresses the failure of standard conformal prediction for dependent, variable-length VLN episodes by normalizing nonconformity scores with residual confidence and calibrating per episode. This construction ensures step-wise coverage under exchangeability, a key technical contribution for sequential decision-making under uncertainty.
Reliable uncertainty estimation in VLN can improve safety and trust in embodied AI agents, potentially enabling deployment in real-world navigation tasks where ambiguous predictions must be flagged.
The method could reduce risk in autonomous navigation systems by providing calibrated uncertainty, which is valuable for robotics and AI assistant products requiring safe operation.
Future work may extend ENCP to other sequential decision-making domains and test its robustness under distribution shift beyond seen-to-unseen evaluation.