Recoleta
GitHub
Source note

Reflective VLA: In-Context Action Consequences Make VLAs Generalize

Authors
Qing Lian, Kent Yu, Lei Zhang
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot Foundation ModelIn Context LearningAction ConsequencesRobot Generalization
Open arXivMarkdown source

Summary

Reflective VLA adds past observation-action-consequence triplets to a vision-language-action policy so the robot can infer camera, calibration, and actuation effects during deployment. The paper reports higher success under LIBERO distribution shifts without test-time fine-tuning, while keeping strong standard benchmark performance.

Problem

  • Reactive VLAs predict the next action from the current instruction and observation, so one frame may not identify camera-to-robot geometry, robot calibration error, or actuation bias.
  • These hidden embodiment factors matter because the same command can produce different visible motion in a new setup, which hurts deployment generalization.
  • Plain temporal history can track task progress, but it may not bind an executed action to the scene change that followed it.

Approach

  • The model keeps a rolling context of triplets (O_i, A_i, O'_i), where O_i is the pre-action observation, A_i is the executed action chunk, and O'_i is the observation after the full chunk horizon C.
  • Previous action chunks are embedded as 8 tokens in the VLM token space, alongside third-person images, wrist images, proprioception, language, and time tokens.
  • A continuous flow-matching action expert attends to the full shared VLM prompt and the current observation before predicting the next action chunk.
  • A block-causal mask trains all K sampled frames in one forward pass while blocking access to each target's own future action and consequence.
  • At inference, the system stores the policy's own executed chunks and observed outcomes, then reuses cached VLM keys and values for past triplets.

Results

  • On standard LIBERO, Reflective VLA reaches 97.6% average success, compared with 96.9% for the matched reactive pi_0.5 baseline, a +0.7 point gain.
  • On SimplerEnv-Bridge, it reaches 78.2% average success, compared with 72.9% for the matched reactive baseline, a +5.3 point gain.
  • On LIBERO-Plus, it reaches 87.7% average success, compared with 82.3% for the matched reactive baseline, a +5.4 point gain. The largest listed gain is on the Robot shift: 72.9% vs 50.0%, or +22.9 points.
  • On LIBERO-Plus-Hard, it reaches 68.8% average success, compared with 64.6% for the matched reactive baseline, a +4.2 point gain. Camera† is 76.3% vs 74.0%, and Robot Calibration† is 61.3% vs 55.2%.
  • The context ablation on Camera, Camera†, and Robot Calibration† gives 77.8% average for full (O,A,O') context, compared with 73.1% reactive, 72.3% for O only, and 73.8% for (O,A).
  • The context-length ablation with full triplets improves average success from 73.1% at K=1 to 75.3% at K=2, 76.7% at K=4, and 77.8% at K=8.

Link