핵심 개념
The author proposes a framework combining model-based control, search, and learning to design efficient control policies for agile locomotion on stepping stones.
초록
Legged robots face challenges in agile locomotion on stepping stones. The proposed framework combines NMPC with MCTS to find feasible plans quickly. Diffusion models handle multi-modality in the dataset effectively. The policy learned through supervised learning can generate reactive contact plans even in dynamic environments.
통계
Median execution time: 8.35s
Number of iterations: 930 (median)
Average number of NMPC simulations: 5.8
인용구
"The main goal is to propose an efficient framework based on a combination of nonlinear MPC (NMPC), Monte Carlo tree search (MCTS), and supervised learning."
"We demonstrate automatic online surface selection for dynamic quadrupedal locomotion through a learned feedback policy."