Single policy estimates uncertainty to guide robot learning actions
Post-hoc Epistemic Uncertainty for Vision-Language-Action Policies
RoboticsMachine Learning
Summary
Robots that follow vision and language instructions sometimes need help when they're unsure about a task. Instead of training many different robot programs to spot when they don’t know something, the authors developed a way for just one program to tell how unsure it is. Their method uses math to measure uncertainty after training, without changing the robot’s main program. This helps robots decide when to ask for help or more examples, making learning more efficient and safer.
What this means in practice
- •For robotics engineers: Implement uncertainty estimates within a single vision-language robot policy to guide when robots should request human help or additional demonstrations during learning.
- •For automation system developers: Use faster uncertainty scoring from post-hoc kernel-variance heads to reduce computational costs when deploying adaptive robot assistants in real-world environments.
- •For industrial automation operators: Improve decision-making on when to escalate complex robot tasks to human supervisors using efficient uncertainty indicators derived from a single policy.$Commercial implications: Enables industrial robotic systems to reliably defer to human oversight on uncertain tasks, improving safety and reducing failure risks.
Authors
Johannes Keller, Albrecht M. Bloße, Adrian Lindenmeyer, Thomas Neumuth, Daniel Schneider
Abstract
Vision-language-action (VLA) policies need to recognize gaps in their training knowledge to request useful demonstrations or assistance in new situations, especially when failures are costly. Disagreement within an ensemble of independently trained policies can guide these decisions, but requires training and running multiple policies. We propose post-hoc kernel-variance heads that estimate epistemic uncertainty from the embeddings of a single policy. Each head computes the conditional variance of a Gaussian process conditioned on embeddings from the training demonstrations. Fitting requires no regression targets or gradient updates and leaves the policy unchanged. In experiments with SmolVLA on LIBERO-10, the kernel variance heads perform competitively with two-policy ensemble scores in task-level correlation with task success, in demonstration acquisition during active learning, and in referral to a stronger supervisor policy as a proxy for human assistance. For SmolVLA, the head on the vision-language backbone reduces uncertainty-scoring time by up to approximately a factor of 50 compared with ensemble scores, while also substantially reducing setup costs. We also show that the established evaluation protocol on LIBERO-10 primarily rewards recognition of coverage differences between tasks, while execution variability weakens assessment of within-task rankings. Thus, controlled perturbations provide a complementary evaluation: modality-specific heads identify unfamiliar camera, instruction, or state inputs, while fused heads can detect unfamiliar combinations of individually familiar inputs. With a supervisor proxy trained on the evaluated perturbation families, the decoder head improves within-task referral over ensemble disagreement. These findings show that useful uncertainty estimates for adaptive robot learning can be obtained from a single policy.