Skip to content
Changda Tian

Capability-Aware Planning Learning-Based Locomotion Control

Capability-Aware Locomotion for Hexapod Robots

A hierarchical learning system that teaches the Qingzhui hexapod what it can traverse — and plans long-range paths it can actually execute.

Period
2020 – 2023
Platform
Qingzhui hexapod · CoppeliaSim
Methods
Reinforcement Learning · Capability Abstraction · Path Planning
With
Yue Gao (SJTU)
Paper
Layered learning framework of the capability system: terrain encoding, low-level capability maximizing, high-level capability abstraction, and capability-based global path planning.

Overview

A legged robot’s ability to cross terrain depends on its structure, its topology, and its locomotion controller. Existing planners ignore this: they produce paths without asking whether the robot behind them can execute the motion. This project makes the robot’s traverse capability — its terrain- and controller-dependent success rate — a learned, queryable model at the heart of both control and planning.

Research problem

Two coupled questions drive the work:

  1. Given the current terrain, which control strategy and body topology maximize the robot’s chance of getting through?
  2. Given a learned picture of that capability, how should a global planner choose long-range paths?

System

The system has three learned components on top of a classical control stack:

  • Capability maximizing — a reinforcement-learning agent picks the best motion-control strategy and topology for the current terrain and foothold state. The chosen controller generates torso and feet trajectories, executed through inverse kinematics and joint PD control.
  • Capability abstraction — a supervised network predicts traverse capability from encoded terrain maps and foothold states, trained on large numbers of simulated locomotion trials produced by the low-level layer.
  • Capability-based path planning — classic planners are augmented with the capability model so that global guidance paths conform to what the robot can do.

The first version of this pipeline, CapPlanner, was published at ROBIO 2022 and was a finalist for the Best Paper in Biomimetics award. The extended system adds capability maximizing and was published at IEEE RCAR 2024.

CapPlanner two-layer framework: the top-level controller learns the traverse ability of the bottom-level-controller-plus-robot system and issues commands that conform to it.

Results

We trained the framework in simulation and ran long-range locomotion experiments both in simulation and on the physical Qingzhui hexapod. Across terrains of varying complexity, capability-aware planning substantially improved global locomotion success compared to capability-blind baselines — the robot chooses routes it can survive, and switches topology when the terrain demands it.

Animated clip of the hexapod robot walking
Qingzhui traversing terrain with the learned locomotion stack.

Publications

  • CapPlanner: Adaptable to Various Topology and Locomotion Capability for Hexapod Robots. Changda Tian, Yue Gao. IEEE ROBIO 2022, pp. 519–524 — Best Paper in Biomimetics finalist. DOI
  • Learning Capability to Enhance Locomotion Control and Planning for Legged Robots. Yue Gao, Changda Tian, Yang Zhang. IEEE RCAR 2024, pp. 25–30. DOI