Reproduce “Learning Accurate Whole-body Throwing with High-frequency Residual Policy and Pullback Tube Acceleration” 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
Yuntao Ma, Yang Liu, Kaixian Qu, Marco Hutter
Researchers at ETH Zürich and EPFL developed a control framework for legged mobile manipulators, enabling precise whole-body prehensile throwing by combining learning-based policies with a high-frequency, optimization-based robust acceleration module. The system improved throwing accuracy by 49.5% and achieved a mean landing error of 0.276 m for 6-meter targets, outperforming human participants in a comparative task.
Key insights
- The integration of learning-based (nominal and residual policies) and model-based (pullback tube acceleration) control strategies is effective for achieving high-precision, dynamic whole-body manipulation.
- High-frequency residual policies operating at 400 Hz can substantially improve end-effector velocity tracking accuracy, particularly at higher throwing speeds.
- Real-time, closed-loop optimization, specifically the Pullback Tube Acceleration, is vital for maintaining throwing accuracy by robustly compensating for uncertainties in object release timing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.