Recoleta
GitHub
Source note

ChunkFlow: Towards Continuity-Consistent Chunked Policy Learning

Authors
Zhao Yang, Yinan Shi, Mingyuan Yao, Wenyao Xue, Yawei Jueluo, Longjun Liu
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionGeneralist Robot PolicyRobot Data ScalingDexterous ManipulationSim2real
Open arXivMarkdown source

Summary

ChunkFlow addresses temporal jitter in vision-language-action policies that generate overlapping action chunks. It combines seam-aware training with deterministic overlap blending to improve continuity, long-horizon success, and inference efficiency without adding a policy forward pass at deployment.

Problem

  • Consecutive action chunks can disagree in their overlapping steps because they use shifted observations and executed histories, creating boundary jitter and unstable robot motion.
  • Inference-only smoothing methods reweight conflicting actions but do not correct the policy errors that cause seam mismatch, allowing residual errors to accumulate over long rollouts.

Approach

  • Partition each chunk into frozen, editable seam, and future zones, then linearly blend overlapping actions so control shifts deterministically from the previous chunk to the current one.
  • Train raw chunk predictions with a boundary loss plus first-order total-variation and second-order curvature penalties.
  • Corrupt action histories and use scheduled sampling so the policy is exposed to execution-history errors during training.
  • Fine-tune with advantage-weighted actor-critic updates using post-blending histories while retaining the seam and continuity regularizers.

Results

  • On CALVIN ABC-D, ChunkFlow reports an average episode length of 4.30, with MSD-Δa of 0.075, MSD-Δ²a of 0.154, seam jump of 0.209, HF ratio of 0.431, and TV-L1 of 0.001.
  • On LIBERO-Long, it achieves 93.4% success versus 92.6% for PI0.5, 83.7% for PI0.5-RTC, 83.8% for CLIP-RT, and 53.7% for OpenVLA.
  • On LIBERO-Long, it reports the lowest listed motion and artifact metrics: MSD-Δa 0.042, MSD-Δ²a 0.197, MSD-Δ³a 0.235, seam jump 0.082, HF ratio 0.135, and TV-L1 0.011.
  • Its average reasoning latency on LIBERO is 4.43 ms, compared with 18.47 ms for PI0.5-RTC and 9.04 ms for PI0.5.
  • In two real-robot tasks, the paper reports 9/10 successful trials; the excerpt does not provide task-level baselines or statistical uncertainty.

Link