Gait-Conditioned Reinforcement Learning
with Multi-Phase Curriculum for Humanoid Locomotion

Tianhu Peng · Lingfan Bao · Chengxu Zhou

Humanoid Robotics Lab, Department of Computer Science, University College London
Team Website

A unified policy that performs standing, walking, and running with smooth transitions, via a gait-conditioned reward routing mechanism and a multi-phase curriculum. Real-world deployment on Unitree G1.

arXiv Video Cite Code: Coming Soon

Highlights

Gait-conditioned reward routing with per-mode terms and shared regulation.
Multi-phase curriculum: walk → walk+stand → +transitions → +run.
Masked critic/value loss & optional multi-head actor to avoid interference.
Real-robot evaluation on Unitree G1; smooth walk↔stand transitions.
Open-source training configs and logs for reproducibility.

Framework

A unified recurrent policy with gait encoding and a structured gait/velocity curriculum. Gait-aware losses reduce interference across modes and stabilize multi-gait learning.

Gait-Conditioned Reward

The gait mask routes gait-specific objectives (e.g., contact, push-off, stillness) while regulation terms remain shared, mitigating reward interference among run/walk/stand.

Gait mask and reward routing — Gait mask activates mode-specific rewards; regulation rewards are shared across gaits.

Simulation

Low-speed walk — Gait: **Walk (low speed)**

High-speed walk — Gait: **Walk (high speed)**

Low-speed run — Gait: **Run (low speed)**

High-speed run — Gait: **Run (high speed)**

Walk to Stand — Transition: **Walk → Stand**

Run to Walk — Transition: **Run → Walk**

Video

Citation

@inproceedings{peng2025gaitconditioned,
  title     = {Gait-Conditioned Reinforcement Learning with Multi-Phase Curriculum for Humanoid Locomotion},
  author    = {Peng, Tianhu and Bao, Lingfan and Zhou, Chengxu},
  booktitle = {IEEE-RAS International Conference on Humanoid Robots},
  year      = {2025},
  
}