Event

Post-hoc Epistemic Uncertainty Estimation Methods for VLA Policies

1 reports · 1 sources

Johannes Keller et al. propose post-hoc kernel-variance heads that estimate epistemic uncertainty from the embeddings of a single vision-language-action (VLA) policy; fitting requires neither regression targets nor gradient updates and does not alter the original policy. In experiments with SmolVLA on LIBERO-10, the method performs comparably to a two-policy ensemble on three tasks: the correlation between task-level uncertainty and success rate, demonstration acquisition in active learning, and referral to a stronger supervisory policy (as a proxy for human assistance); under a supervisory proxy trained on the evaluated perturbation family, the decoder head outperforms ensemble disagreement on within-task referral.