Recoleta
GitHub
Source note

AttenA+: Rectifying Action Inequality in Robotic Foundation Models

Authors
Daojie Peng, Fulong Ma, Jiahang Cao, Qiang Zhang, Xupeng Xie, Jian Guo, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot Foundation ModelsWorld Action ModelsManipulation PolicyLoss Weighting
Open arXivMarkdown source

Summary

AttenA+ improves robot foundation model training by giving higher loss weight to slow, precision-heavy action steps. It reports small but consistent success-rate gains on Libero and RoboTwin 2.0 without changing model backbones.

Problem

  • VLA and WAM models usually give every action timestep the same loss weight, although grasping, aligning, and placing are more error-sensitive than fast approach motions.
  • This matters because last-centimeter errors can fail long-horizon manipulation tasks, even when most of the trajectory is correct.

Approach

  • Compute instantaneous velocity magnitude from the ground-truth action. On Libero, the method uses continuous motion dimensions and omits the binary gripper state.
  • Convert velocity into a training weight: low velocity gets a larger weight, and high velocity gets a lower weight.
  • Test four velocity-to-weight mappings: inverse, inverse squared, exponential decay, and logarithmic.
  • Apply the weights to the existing training loss, such as weighted L1 loss for OpenVLA-OFT and weighted flow-matching loss for pi-style generative policies.
  • Clip and optionally normalize weights so near-static steps do not control the gradient and the average loss scale stays close to the baseline.

Results

  • On Libero, AttenA+OFT reaches 98.60% average success rate versus 97.10% for OpenVLA-OFT, a +1.50 point gain. Average error rate drops from 2.90% to 1.40%.
  • Libero task scores for AttenA+OFT are 99.0% ± 0.16 on Spatial, 100.0% ± 0.00 on Object, 98.8% ± 0.28 on Goal, and 96.6% ± 0.30 on the 10-task long-horizon split.
  • Against OpenVLA-OFT on Libero, gains are +1.4 points on Spatial, +1.6 on Object, +0.9 on Goal, and +2.1 on long-horizon tasks.
  • With the generative pi-0.5 backbone on Libero, AttenA+pi-0.5 reaches 97.95% average success rate versus 96.85%, a +1.10 point gain. Average error rate drops from 3.15% to 2.05%.
  • On RoboTwin 2.0, AttenA+WAM reaches 92.46% average success rate versus 91.80% for Fast-WAM and 92.20% for LingBot-VA. It reports 93.06% on Clean and 91.86% on Randomized.
  • The paper also claims real-world validation on a Franka manipulator and cross-task generalization, but the excerpt gives no concrete real-world success-rate numbers.

Link