UR-VC: Unsupervised Robotic Value Correction for Time-Derived Progress Proxies
UR-VC (Unsupervised Robotic Value Correction) is an unsupervised, training-free offline method for correcting time-derived progress labels in robot learning. It retrieves similar states from other episodes and aggregates their time labels to obtain corrected progress estimates. The paper is published on arXiv, number 2607.12892v1.
UR-VC proposes an unsupervised robotic value correction method that uses aggregation of time labels from similar states across episodes to correct time-derived progress proxies, without requiring human labels or reward annotations.
The method leverages the regularity of similar states appearing across episodes in demonstration data, retrieving and aggregating time labels to correct progress estimates, avoiding the misleading nature of monotonically increasing time labels in contact-rich manipulation.
This work provides a low-cost, unsupervised alternative for obtaining dense progress signals in robot learning, potentially reducing reliance on human annotation.
This method may reduce data annotation costs for robot learning systems and accelerate the application of robot skill learning in industrial scenarios.
Future work could verify whether the method outperforms original time labels in various robot manipulation tasks and whether it can be extended to more complex tasks.