Reproduce “Asynchronous Fast-Slow Vision-Language-Action Policies for Whole-Body Robotic Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Teqiang Zou, Hongliang Zeng, Yuxuan Nong, Yifan Li, Kehui Liu, Haotian Yang, Xinyang Ling, Xin Li and 1 more
DuoCore-FS lets a slow vision-language model write task guidance to a buffer while a fast controller combines it with current images and body state to generate action chunks. On Astribot’s S1 popcorn task, it completed all four stages in 18 of 20 trials, versus 17 of 20 for π₀, at 32.3 Hz versus 12.5 Hz. Tests cover one robot and limited kiosk tasks.
Key insights
- Decoupling high-level semantic reasoning from low-level action generation through an asynchronous fast-slow architecture effectively overcomes VLM inference bottlenecks while maintaining VLM-driven intelligence.
- A specialized geometry-aware action tokenizer is essential for efficiently learning and representing high-dimensional, structured whole-body actions, enabling stable policy performance.
- End-to-end co-training with cross-timescale sampling effectively aligns the slow and fast systems, ensuring robust performance under asynchronous deployment conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.