ChunkFlow: Towards Continuity-Consistent Chunked Policy Learning

2026-07-14

AI Summary

This work addresses the temporal inconsistency and degraded task success in vision-language-action models caused by boundary jitter during chunked inference. To mitigate these issues, the authors propose ChunkFlow, a framework that jointly optimizes chunking strategies during both training and inference. ChunkFlow partitions each chunk into frozen, editable, and future regions and incorporates seam-aware chunking, deterministic overlapping fusion, and first- and second-order continuity losses. It further integrates historical perturbation, scheduled sampling, and AWAC fine-tuning to enhance robustness. Theoretical analysis demonstrates that overlapping pre-fusion error decays with increasing overlap size. Experiments show that ChunkFlow significantly improves task success rates and temporal stability under low-latency inference on CALVIN, LIBERO, and real-world robotic platforms.

Abstract

Vision-language action (VLA) models increasingly adopt chunked action heads to satisfy real-time constraints; however, this introduces boundary jitter: overlapping regions between consecutive chunks often yield inconsistent predictions, degrading temporal coherence and the task success rate. Existing methods, such as inference-time blending, merely reweight mismatched proposals without correcting underlying errors, leading to residual accumulation under biased or noisy histories. We propose ChunkFlow, a seam-aware training-and-execution framework for chunked policies that aligns chunk structure with boundary execution. It partitions each chunk into frozen, editable, and future zones, applies deterministic overlap blending at execution, and trains raw predictions with seam and first- and second-order continuity losses. History corruption and scheduled sampling improve robustness to executed-history errors, while an AWAC fine-tuning stage adapts the policy without removing these structural regularizers. Under mild smoothness assumptions, pre-blending seam discrepancies provably decay with increasing overlap. Experiments on CALVIN, LIBERO, and real robots show an improved success-stability trade-off with low-latency inference. Project page: https://cytoderm-ai.github.io/chunkflow.

Problem

• chunked policy
• boundary jitter
• temporal coherence
• vision-language action models
• action prediction inconsistency

Innovation

• chunked policy learning
• boundary jitter
• continuity-consistent training
• seam-aware blending
• visual-language action models