Reproduce “DynaForge: Planning-Guided Residual Learning for Dynamic Manipulation Demonstration Generation” 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

Yiyang Jin, Yu Zheng, Xiao He, Hesheng Wang

DynaForge generates demonstrations for robots manipulating moving objects by combining collision-aware planning, object-centric inverse kinematics, and learned residual action corrections near contact. It retains planning for stable motion while updating targets from object pose. Across nine simulation tasks, residual corrections raise mean demonstration-generation rate from 41.30% to 78.37%. With 800 demonstrations per task, downstream success is 49.11% versus 7.07% for DOMINO. Results remain bounded by authored task structure and simulator fidelity, with weak precision insertion and modest physical transfer.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.