Reproduce “FastVLA: Task-Conditioned and Temporally Adaptive Visual Token Compression for Efficient Vision-Language-Action Models” 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
Jiuyi Xu, Chengyuan Qian, Zhengzhong Tu, Yang Sui, Yangming Shi
FastVLA adaptively prunes visual tokens for robot policies, selecting task-relevant tokens and adjusting the retained budget at each action step according to recent attention concentration. In LIBERO’s suites, OpenVLA-OFT’s LLM-backbone FLOPs fall from 3.98 to 1.65 trillion and action latency from 118.52 to 81.17 milliseconds, while success decreases from 97.0% to 96.5%. Real-robot evaluation covers two tasks on one robot.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.