Reproduce “Size Doesn't Matter: Material-State Reinforcement Learning for Excavator Transferable Soil 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
Lennart Werner, Pol Eyschen, Sean Costello, Pierluigi Micarelli, Andrei Cramariuc, Marco Hutter
A material-state-aware reinforcement-learning framework enables excavators to learn transferable soil-manipulation strokes in an MPM simulator and deploy them through a normalized end-effector interface. On an 11.5-ton excavator, the learned policy autonomously built a 42-meter embankment across 201 executions without stalls or operator intervention, while the same policy transferred without retraining to a 500-gram tabletop robot and adapted its behavior to soil hardness, force limits, and compaction state.
Key insights
- Material-state feedback enables the policy to distinguish and respond to altered or compacted soil, rather than treating equivalent surface geometries as identical.
- The learned behavior is expressed in normalized end-effector coordinates, allowing one policy checkpoint to transfer between an 11.5-ton excavator and a 500-gram robot after interface calibration rather than policy retraining.
- Force capability and terrain observations meaningfully shape the learned stroke: the policy changes penetration depth, attack angle, and scraping behavior as soil resistance and available force vary.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.