Reproduce “A1: A Fully Transparent Open-Source, Adaptive and Efficient Truncated Vision-Language-Action Model” on OpenResearch

About this paper

Kaidong Zhang, Jian Zhang, Rongtao Xu, Yu Sun, Shuoshuo Xue, Youpeng Wen, Xiaoyu Guo, Minghao Guo and 15 more

The A1 framework introduces a fully open-source, adaptive, and efficient Vision-Language-Action model designed for low-cost, high-throughput robot manipulation. It achieves up to a 72.3% reduction in inference time for flow-matching based action heads and up to a 76.6% reduction in VLM backbone computation, while consistently maintaining or improving success rates across various simulation and real-world robotic benchmarks.

Key insights

• Flow-matching trajectories often converge to a correct action mode within a limited number of denoising steps, indicating diminishing returns from extensive iterations.
• Intermediate hidden states within the VLM backbone frequently contain sufficient features for initial action prediction, suggesting that processing the VLM to its full depth is often unnecessary.
• Computational resources should be adaptively expended only when predicted actions show significant changes, optimizing efficiency by reducing redundant computations.

Summary by alphaXiv.