A new way for robot policies to recognize gaps in their knowledge
Post-hoc kernel-variance heads estimate uncertainty from a single policy.
Vision-language-action (VLA) policies need to recognize gaps in their training knowledge. This helps them request assistance in new situations when failures are costly.
Post-hoc kernel-variance heads estimate epistemic uncertainty from the embeddings of a single policy. These heads compute the conditional variance of a Gaussian process conditioned on embeddings from training demonstrations.
Fitting these heads requires no regression targets or gradient updates. The process leaves the policy unchanged.
In experiments with SmolVLA on LIBERO-10, these heads perform competitively with two-policy ensemble scores. They work in task-level correlation with task success and during demonstration acquisition in active learning.
The head on the vision-language backbone reduces uncertainty-scoring time by up to a factor of 50 compared with ensemble scores. This also substantially reduces setup costs.
The LIBERO-10 evaluation protocol primarily rewards recognition of coverage differences between tasks. Execution variability weakens assessment of within-task rankings.
Key facts
- Post-hoc kernel-variance heads estimate epistemic uncertainty from a single policy.
- The head on the vision-language backbone reduces uncertainty-scoring time by up to a factor of 50.
- SmolVLA with these heads performs competitively with two-policy ensemble scores on LIBERO-10.