Reproduce “Learning Tool-Aware Adaptive Compliant Control for Autonomous Regolith Excavation” 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
Andrej Orsula, Matthieu Geist, Miguel Olivares-Mendez, Carol Martinez
Researchers at the University of Luxembourg developed a system for autonomous regolith excavation that learns tool-aware adaptive compliant control, leveraging high-fidelity simulation and model-based reinforcement learning. The approach demonstrates robust zero-shot generalization across diverse tool geometries and significantly improves excavation safety and efficiency, with a visual agent excavating 0.27L of material with the least dust and jerk.
Key insights
- Learned adaptive compliant control is crucial for safe and efficient excavation, producing markedly smoother interactions and significantly less dust generation compared to rigid control strategies.
- Training with procedural diversity in robot tool geometries is essential for achieving robust zero-shot generalization to unseen tools, enabling the agent to learn abstract excavation principles.
- Augmenting the adaptive agent with visual feedback (depth maps) enhances performance, leading to higher excavated volumes, reduced dust, and smoother motions due to improved situational awareness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.