Reproduce “MotionVLA: Injecting Geometric Motion into Vision-Language-Action Model” 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
Shanglin Yuan, Weiheng Zhao, Xianda Guo, Wei Sui, Li Yu, Wenyu Liu, Xinggang Wang
Researchers at Huazhong University of Science and Technology introduce MotionVLA, a vision-language-action (VLA) model that injects motion-consistent historical evidence via a novel motion-history interface. This approach enhances long-horizon manipulation capabilities, achieving a 12 percentage point increase in success rate on RoboTwin2.0 and generating robot executions with path efficiency closer to expert demonstrations compared to prior methods.
Key insights
- Effective VLA memory should represent the underlying motion and temporal continuity connecting observations, rather than just accumulating fragmented individual frames or raw geometric data.
- A selective retrieval mechanism (Decouple-then-Recouple) for historical motion evidence, guided by current visual context, is critical for stable control and preventing the integration of unaligned or noisy information.
- Grounding motion tokens in control-relevant dynamics via auxiliary supervision is essential for their utility, ensuring they encode actionable information rather than merely static appearance summaries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.