.png){kind=icon}
.png){kind=icon}
Return
Robust bipedal walking control through a privileged learning-based, hierarchical imitation learning method
.png){kind=icon}
PRE
AI
DOI：10.1007/s00521-025-11808-z.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Reinforcement learning-based controllers enable robots to autonomously learn walking behaviors but often suffer from inefficiencies due to sparse rewards. A common solution involves guiding robots with reference trajectories to imitate predefined motions, accelerating learning but limiting adaptability and independence in unpredictable environments. Additionally, deploying these controllers in real-world applications is complicated by the need to integrate reference motion inputs. This paper introduces a privileged learning-based, hierarchical imitation learning method to address these limitations. Initially, a teacher policy is trained with privileged access to reference trajectories to learn optimal state transitions through reinforcement learning. Subsequently, a student policy is trained to imitate the teacher, relying solely on its internal state without requiring reference trajectories or external sensors. Simulation results and robustness analyses demonstrate that reference-driven policies restrict joint flexibility and underperform on uneven terrains or low-friction surfaces, often becoming unstable under external forces. In contrast, the proposed reference-free method dynamically adapts to uncertain environments, maintaining stable walking even when subjected to random disturbances. This approach reduces reliance on predefined trajectories, sensors, and external perception models, providing a more robust and adaptable solution for autonomous robot locomotion.
Keywords:
Robust walking control
Privileged learning
Hierarchical imitation learning
Reference-free
Uncertain environments
Journal
IF:
4.5
Papers:
830
Citations:
3.2W