Search papers, labs, and topics across Lattice.
This paper introduces Unsupervised Robotic Value Correction (UR-VC), an innovative method for refining time-derived progress labels in robotic manipulation tasks, which are often inaccurate due to the monotonic nature of time as a proxy for progress. By leveraging the recurrence of similar states across different episodes, UR-VC aggregates time-derived labels to provide a more accurate estimate of task progress without the need for manual annotations or additional models. The method demonstrates improved task success in real robotic applications, particularly in complex manipulation scenarios like cloth flattening and folding, highlighting its effectiveness in capturing nuanced progress dynamics.
Time-derived progress labels can mislead robotic learning, but UR-VC corrects these inaccuracies to enhance task success in real-world manipulation.
Modern robot learning systems increasingly rely on dense progress or value signals to evaluate intermediate states, guide policy learning, and detect task completion, making the quality of these signals critical. Since such dense labels are rarely available at scale, normalized time within a demonstration is often used as a scalable substitute: later frames are treated as higher progress. However, this time-derived label is only a noisy proxy for physical task progress. In contact-rich manipulation, a robot may make progress and then lose it through slips, failed grasps, or partial undoing, while the time-derived label continues to increase monotonically. We introduce Unsupervised Robotic Value Correction (UR-VC), an offline, training-free method for correcting time-derived progress labels. UR-VC exploits a simple regularity in demonstration data: similar states often recur across different episodes, but at different timestamps. Instead of trusting the timestamp from a single trajectory, UR-VC retrieves similar states from other episodes and aggregates their time-derived labels to obtain a corrected progress estimate. UR-VC requires no manual progress labels, reward annotations, or additional value model. We evaluate UR-VC on real bimanual cloth flatten-and-fold data, a long-horizon deformable-object manipulation task with visible intermediate progress. The corrected labels capture local regressions and non-uniform progress that normalized time cannot represent, while preserving the overall task trend. We further use the corrected signal to construct advantage labels for VLA training, following recent advantage-conditioned policy learning. UR-VC shows a positive trend in real-robot task success under matched data, model, and training settings.
