ChunkFlow Integrates Smoothing Constraints Directly Into Training

stepjamUK · x · 2026-07-16

This paper on ChunkFlow tackles the "seam" issue in action chunking for real-time controlled VLAs. To meet latency requirements, actions must be predicted in chunks, but adjacent chunks often conflict at boundaries due to slight shifts in observation and history. The common practice of smooth fusion during inference looks fluid visually but fails to correct the divergent predictions themselves, potentially accumulating errors under noisy execution histories.

The author argues that embedding these constraints into training is more effective. ChunkFlow divides each chunk into frozen, editable, and future zones, applying deterministic fusion at boundaries while adding seam and continuity losses to directly constrain the policy. An AWAC phase later adapts based on the fused execution history without discarding these structural constraints.

Experiments show it achieves 93.4% on LIBERO with an average latency of 4.43ms. The author admits this is only 0.8 points higher than π0.5, which is near noise on this benchmark and not the core finding. More notably, the boundary discrepancy metric dropped to 0.082, significantly lower than all baseline methods.

Original post →

More from Embodied

Embodied channel →