My research investigates how motion priors can support locomotion when commands, terrain, and task requirements change. These three studies examine different locations for adaptation: the parameters of a biologically inspired controller, a high-level policy built on animal skills, and terrain-aware references generated from human motion. Together, they explore how to retain useful movement structure while responding to the environment.
Unitree Go1 · IROS 2024
2024 · EPFL BioRob
Online adaptation of CPG parameters
Central pattern generators provide a compact, biologically inspired structure for rhythmic locomotion. Their parameters still need to match the robot’s commands and operating conditions.
At EPFL’s BioRob, I investigated online Bayesian optimization of these parameters, enabling a quadruped to adapt to changes in target velocity, ground friction, slope, and payload. Hardware experiments on the Unitree Go1 examine adaptation directly on the robot. Here, the motion structure remains a CPG, while its parameters change through real-world interaction.
Animal motion capture provides reusable locomotion skills, but imitation on flat ground does not by itself address uneven terrain. This work first trains a low-level policy to imitate dog motion, establishing reusable movement skills.
A high-level, goal-conditioned policy then learns residual corrections that adapt those skills to terrain and navigation requirements. Experiments on ANYmal demonstrate stair traversal, uneven-ground locomotion, and local navigation. Adaptation operates through the residual policy while retaining the movement structure supplied by the animal prior.
Human motion provides coordinated whole-body skills, but fixed reference trajectories cannot accommodate every terrain arrangement. This framework uses a diffusion model to generate terrain-aware motion references and a learned tracker to execute them on a humanoid.
The tracker is trained with reinforcement learning, then fine-tuned in closed loop with the motion generator frozen. The motion generator conditions its references on terrain; the framework does not use a separate residual adaptation policy. Onboard experiments on the Unitree G1 demonstrate coordinated traversal of boxes, hurdles, stairs, and mixed terrain.