AI

When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 04:18:25 UTC.

Briefings

Vision-Language-Action (VLA) models enable robotic manipulation but suffer from stop-and-go execution due to expensive inference. This study analyzes action errors in long action chunks, finding they concentrate sharply around transitions between manipulation subskills. The authors introduce RACE, a framework that predicts transition timing via an auxiliary one-step denoising pass to condition action generation. Across simulations, RACE outperforms fine-tuning at the same chunk length, surpassing state-of-the-art efficient VLAs with 4x longer chunks. On a real robot, RACE reduces idle time by about 5x while maintaining a higher success rate than fine-tuning with the same chunk length.

Key points

• Action errors in long VLA chunks concentrate sharply around transitions between manipulation subskills, making transition timing critical for reliability.
• RACE framework predicts transition timing using an auxiliary one-step denoising pass to condition action generation, significantly reducing errors at subskill transitions.

Context

Vision-Language-Action models serve as unified policies for robotic manipulation, yet their expensive inference forces robots to pause between policy calls, resulting in stop-and-go execution that interrupts smooth motion and prolongs task completion. Extending the action chunk reduces policy calls and hence these pauses, but predicting farther into the future makes long-chunk execution unreliable. This study analyzes action errors within long chunks to understand where unreliability arises.

Limitations

Analysis focuses on action errors within long chunks, revealing concentration around subskill transitions. Simulation benchmarks show RACE outperforms fine-tuning at same chunk length, surpasses state-of-the-art efficient VLAs with 4x longer chunks, and on a real robot, RACE reduces idle time by about 5x compared to fine-tuning with the same chunk length. Code and a real-robot demo are available at the provided GitHub link.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 16/10/2026, 23:51:43 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeida6ol3al77kz442ksa7immw2ocl3xccnyie4nng3ifnecyu5v72y