Reproduce “General Flow as Foundation Affordance for Scalable Robot Learning” 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
Chengbo Yuan, Chuan Wen, Tong Zhang, Yang Gao
Researchers at Tsinghua University and affiliated institutes developed "General Flow," a language-conditioned 3D flow prediction model trained on large-scale human RGBD video datasets, to enable scalable robot learning. The framework achieves an 81% average success rate in zero-shot human-to-robot skill transfer across 18 diverse real-world manipulation tasks.
Key insights
- Training on diverse, large-scale human video data endows the General Flow model with emergent properties like robustness to hand occlusion and label noise.
- The model exhibits language-driven semantic control, generating distinct actions for identical visual inputs based solely on language instructions.
- ScaleFlow dynamically adapts predicted trajectory lengths based on the spatial context of the object, showing "scale spatial adaptation."
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.