Feasibility-Guided Planning over Multi-Specialized Locomotion Policies

TL;DR

This work tackles planning over heterogeneous, unstructured terrain by coordinating multiple specialized locomotion policies. It introduces Feasibility-Net, a policy-specific predictor trained jointly with each policy to output a directional, terrain-aware feasibility tensor from local heightmaps and task commands, augmented with a VAE for out-of-distribution detection. Elevation maps are transformed into multi-directional feasibility tensors which are fused across policies using a max operation and then planned via graph search, with explicit policy selection at each step. The framework demonstrates strong simulation results across varied terrains, robust adaptation as policies improve, and real-world validation with a quadruped, achieving high success rates and interpretable policy switching, while supporting seamless addition of new skills.

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 constraints: 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.

Figures (8)

• Figure 1: Feasibility-guided planning enables optimal path selection & policy switching over mixed terrain via policy-specific feasibility representations.
• Figure 2: Overview of the feasibility-aware planning framework. Locomotion policies and Feasibility-Net models are jointly trained using shared environment rollouts, where the feasibility models learn to predict velocity tracking performance and terrain distributions simultaneously with policy optimization.
• Figure 3: Sliding window methodology for elevation map to feasibility tensor transformation. Local heightmap patches are extracted at each spatial location and processed through multiple directions to generate policy-specific feasibility predictions.
• Figure 4: Multi-policy feasibility tensor fusion and planning framework. Individual policy-specific feasibility representations are combined through maximum fusion to create unified cost functions, enabling graph search algorithms to discover optimal paths with transparent policy selection.
• Figure 5: Feasibility-guided planning adaptation across training progression on stepped terrain. As locomotion capabilities develop, the planner progressively selects shorter and more direct routes, demonstrating dynamic adaptation to policy skill evolution.