Feasibility-Guided Planning Over Multi-Specialized Locomotion Policies
Ying-Sheng Luo, Lu-Ching Wang, Hanjaya Mandala, Yu-Lun Chou, Guilherme Henrique Galelli Christmann, Yu-Chung Chen, Yung-Shun Chan, Chun-Yi Lee, Wei-Chao Chen
AI summary
Problem
Coordinating multiple specialized locomotion policies for navigation remains complex, with existing classical planners lacking adaptability and hierarchical RL approaches sacrificing interpretability while requiring full retraining for new skills.
Approach
Each locomotion policy is paired with a Feasibility-Net that predicts directional traversability and terrain familiarity. These policy-specific feasibility tensors are fused and fed into classical graph search algorithms to generate optimal, interpretable paths.
Key results
- Jointly trained per-policy feasibility estimation framework
- Unified RL and supervised training paradigm eliminating separate data pipelines
- Directional feasibility tensor representation enabling interpretable graph-based planning
- Validated efficient, reliable navigation across diverse terrains in simulation and real-world tests
Why it matters
Enables scalable, interpretable, and adaptable navigation for legged robots by allowing classical planners to leverage diverse learned locomotion skills without costly retraining.
Abstract
Planning over unstructured terrain presents a significant challenge in the field of legged robotics. Although recent works in reinforcement learning have yielded various locomotion strategies, planning over multiple experts remains a complex issue. Existing approaches encounter several con- straints: traditional planners are unable to integrate skill- specific policies, whereas hierarchical learning frameworks often lose interpretability and require retraining whenever new policies are added. In this paper, we propose a feasibility-guided planning framework that successfully incorporates multiple terrain-specific policies. Each policy is paired with a Feasibility- Net, which learned to predict feasibility tensors based on the local elevation maps and task vectors. This integration allows classical planning algorithms to derive optimal paths. Through both simulated and real-world experiments, we demonstrate that our method efficiently generates reliable plans across diverse and challenging terrains, while consistently aligning with the capabilities of the underlying policies.