Reproduce “DynamicWAM: Dual-Path Motion Conditioning for World-Action Models in Dynamic 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
Yunfan Lou, Hewen Gao, Xiyu Zhu, Zhuoran Qiao, Xuan Han, Yifan Yang, Yifan Ye, Boxian Yao and 1 more
DynamicWAM conditions a world–action model on recent motion to improve grasping moving objects. It combines rendered optical-flow histories, which preserve spatial motion patterns, with kinematic tokens encoding displacement, duration, velocity, and acceleration. In DOMINO, both paths raise success rate from 22.7% to 38.2% and manipulation score from 38.3 to 53.2; this ablation supports complementarity within the tested setup.
Key insights
- Both dense spatial motion cues (history-flow) and quantitative kinematic information (kinematic tokens) provide distinct yet complementary benefits for dynamic manipulation, with their combination yielding superior performance.
- Motion magnitude and timing information, often lost in normalized visual flow representations, are critical for tasks requiring precise contact with moving targets, and kinematic tokens effectively capture this.
- Distilling a compact video expert and integrating it with asynchronous execution (RTC) enables practical, real-time robot operation in dynamic environments by effectively managing computational latency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.