From Recovery to Locomotion: Learning Robust Humanoid Control via Curriculum and Policy Distillation

Stefanie Theodora Karolina Gschwind

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2025-87
May 16, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-87.pdf

Humanoid robots must operate robustly in dynamic, unstructured environments where falling is inevitable and manual resets are infeasible. While prior work has achieved impressive locomotion or recovery performance in isolation, few approaches address the challenge of unified, full-body control in real-world conditions. This thesis introduces an end-to-end framework for generalizable, robust control that enables humanoid robots to walk, fall, recover, and resume walking autonomously.

We first extend morphology-randomized training from quadrupeds to bipeds and humanoids, uncovering critical limitations in gait stability and failure recovery when generalizing across designs. To overcome these limitations, we propose a keyframe-based recovery curriculum that decomposes the complex getup task into human-inspired phases—transitioning through a kneeling intermediate pose to improve energy efficiency, stability, and success rate. We then integrate recovery and locomotion policies into a single unified controller using Dataset Aggregation (DAgger), enabling seamless transitions between walking and getting up without brittle switching logic.

Evaluated on realistic simulated environments featuring low obstacles, slippery patches, and external perturbations, our unified policy demonstrates superior robustness and consistency compared to baseline methods. This work provides a deployable control architecture for humanoid robots that bridges the gap between simulation and real-world demands, laying a foundation for lifelong, autonomous operation.

Advisor: Sergey Levine

"; ?>

BibTeX citation:

@mastersthesis{Gschwind:EECS-2025-87,
    Author = {Gschwind, Stefanie Theodora Karolina},
    Title = {From Recovery to Locomotion: Learning Robust Humanoid Control via Curriculum and Policy Distillation},
    School = {EECS Department, University of California, Berkeley},
    Year = {2025},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-87.html},
    Number = {UCB/EECS-2025-87},
    Abstract = {Humanoid robots must operate robustly in dynamic, unstructured environments where falling is inevitable and manual resets are infeasible. While prior work has achieved impressive locomotion or recovery performance in isolation, few approaches address the challenge of unified, full-body control in real-world conditions. This thesis introduces an end-to-end framework for generalizable, robust control that enables humanoid robots to walk, fall, recover, and resume walking autonomously.

We first extend morphology-randomized training from quadrupeds to bipeds and humanoids, uncovering critical limitations in gait stability and failure recovery when generalizing across designs. To overcome these limitations, we propose a keyframe-based recovery curriculum that decomposes the complex getup task into human-inspired phases—transitioning through a kneeling intermediate pose to improve energy efficiency, stability, and success rate. We then integrate recovery and locomotion policies into a single unified controller using Dataset Aggregation (DAgger), enabling seamless transitions between walking and getting up without brittle switching logic.

Evaluated on realistic simulated environments featuring low obstacles, slippery patches, and external perturbations, our unified policy demonstrates superior robustness and consistency compared to baseline methods. This work provides a deployable control architecture for humanoid robots that bridges the gap between simulation and real-world demands, laying a foundation for lifelong, autonomous operation.}
}

EndNote citation:

%0 Thesis
%A Gschwind, Stefanie Theodora Karolina
%T From Recovery to Locomotion: Learning Robust Humanoid Control via Curriculum and Policy Distillation
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 16
%@ UCB/EECS-2025-87
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-87.html
%F Gschwind:EECS-2025-87