Source note
AnchorRefine: Synergy-Manipulation Based on Trajectory Anchor and Residual Refinement for Vision-Language-Action Models
Summary
AnchorRefine splits robot action prediction into a coarse trajectory anchor and a small correction step. The paper targets precision failures in vision-language-action policies and reports consistent gains on LIBERO, CALVIN, and real-robot tests.
Problem
- Standard VLA policies predict full action chunks in one space, so large arm motions dominate training and small correction signals get weak supervision.
- This hurts precision-critical manipulation such as final pose alignment, contact timing, and gripper closing, where small errors can flip success to failure.
- The paper points to gripper control as a major failure source: over 88% of LIBERO failures in their analysis involve gripper-related errors.
Approach
- AnchorRefine uses a two-stage action model: an anchor planner predicts a coarse action trajectory, then a refinement module predicts the residual error between that anchor and the ground-truth action.
- For arm motion, the final action is
anchor + residual, so the second module only learns the small correction that remains after the coarse plan. - Training is sequential: first train the anchor planner on the original action target, then freeze it and train the residual branch on anchor-relative residuals.
- For gripper control, the paper adds a decision-aware refinement head that corrects the anchor gripper prediction based on how close it is to the open/close decision boundary.
- The method is designed to plug into both regression-based and diffusion-based VLA backbones, shown here with GR-1 and X-VLA.
Results
- On LIBERO-LONG, AnchorRefine (GR-1) improves success rate from 74.5% to 82.3%, a +7.8 point gain.
- On LIBERO-LONG, AnchorRefine (X-VLA) improves success rate from 95.8% to 97.4%, a +1.6 point gain.
- On CALVIN ABC→D, AnchorRefine (GR-1) improves 5-task chain success from 52.0% to 55.3% (+3.3), 3-task success from 68.5% to 71.9% (+3.4), and average sequence length from 3.51 to 3.64 (+0.13).
- On CALVIN ABC→D, AnchorRefine (X-VLA) improves 5-task chain success from 74.6% to 76.5% (+1.9), 4-task success from 81.3% to 83.7% (+2.4), 3-task success from 86.9% to 89.1% (+2.2), and average sequence length from 4.31 to 4.40 (+0.09).
- The abstract claims gains of up to 7.8% in simulation success rate and 18% in real-world success rate.
- Among the listed LIBERO results, AnchorRefine (X-VLA) reaches 97.4%, higher than the reported 95.8% X-VLA baseline and 95.2% AtomicVLA result in the table.