TraceFlow improves robot task success by guiding frozen policies
TraceFlow: Guiding Frozen Flow-Matching Robot Policies with Success and Failure Traces
Robotics
Summary
Robots sometimes follow fixed action plans that can't learn from their past mistakes during operation. This paper introduces TraceFlow, a method that helps a robot adjust its actions in real time by comparing current behavior to both successful and failed past attempts, even when only the outcome (success or failure) is known. The authors show that TraceFlow can nearly double success rates in a real robot packing task and improve performance in some simulated robot tasks without changing the robot’s original learned policy. This helps robots perform better by using their own past failures as useful feedback alongside successes.
What this means in practice
- •For industrial robotics teams: Enhance real-time robot task success on assembly or packing lines by integrating TraceFlow guidance using stored successful and failed action sequences.
- •For robotics simulation developers: Improve evaluation and training of robot manipulation tasks by adding feedback from failure traces along with successes in simulated environments.
Authors
Jiaxuan Zhang, Ruizhe Liu, Yu Zhang, Yanchao Yang
Abstract
A vision-language-action (VLA) policy with a flow-matching action expert generates each action chunk (a short command sequence) by integrating a learned velocity field; once its weights are fixed, the success or failure of an earlier rollout cannot change the chunk generated now. Concurrent test-time methods give a frozen policy such an input from retrieved successes, a learned critic, a verifier, or a dynamics model, but none uses the robot's own failed rollouts as negative evidence with nothing but a terminal outcome bit. We introduce TraceFlow, a progress-aligned guidance field that turns the action densities of retrieved successful and failed rollouts into a bounded correction to a frozen flow-matching action expert, using one terminal outcome bit per rollout and no other label. Its TraceBank stores traces, time-ordered state-action records with a terminal label, starts from the target-task training traces, and later admits the deployed robot's own rollouts. On an ordered real-robot packing task the base completes 21 of 50 trials in order, TraceFlow 39, and one stacking round without any weight update 47, with wrong-sequence episodes falling from 20 to 0. In simulation the gain is selective: with per-suite selected settings, TraceFlow raises RoboMemArena Sequence from 78.92\% to 91.50\% task success and Transferring from 54.41\% to 62.00\% at stacking round 2, leaves the 26-task aggregate unchanged, lowers Counting and Occlusion by 1.12 and 1.42 points, and changes LIBERO-Plus (Long) by +1.27 points (p = 0.0733). Stacking gains are finite, every branch peaking before round ten, and the bank's success-to-failure ratio predicts no retrieval allocation.