Reproduce “Large Scale Robotic Material Handling: Learning, Planning, and Control” 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
Filippo A. Spinelli, Yifan Zhai, Fang Nan, Pascal Egli, Julian Nubert, Thilo Bleumer, Lukas Miller, Ferdinand Hofmann and 1 more
Researchers from ETH Zurich and Liebherr-Hydraulikbagger GmbH developed a comprehensive framework for autonomous bulk material handling using a 40-ton hydraulic manipulator, integrating reinforcement learning for both optimal pile attack point selection and robust, oscillation-aware motion control. This system, validated in real-world scenarios, achieved grasping efficiencies and throwing precision comparable to expert human operators in challenging industrial tasks, demonstrating robustness against sensor noise and dynamic environmental conditions.
Key insights
- Reinforcement learning policies, particularly for attack point planning, exhibit superior performance and robustness to sensor noise compared to traditional heuristics, even when heuristics operate with ideal information.
- RL controllers effectively manage complex, nonlinear hydraulic dynamics and underactuated gripper oscillations, maintaining high safety constraint satisfaction (e.g., tube constraints) during real-world operation.
- Approximating granular material release with discrete loads in simulation enables RL policies to learn dynamic throwing behaviors that improve real-world material placement accuracy and pile compactness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.