Humanoid robot control based on imitation learning has become a research hotspot for enhancing complex terrain locomotion capabilities due to its effectiveness in transferring human motion dexterity and adaptability. We proposes a novel imitation learning-based control framework called Human Decoupled Imitation Learning (HumanDIL). In it, a motion primitive reconstruction-based action redirection mechanism that extracts human "key skill set" and establishes a bipedal motion feature space, significantly improving data utilization and cross-terrain generalization in imitation learning. Then, a hierarchical control architecture where the lower body employs contact-aware end-to-end imitation learning for terrain-adaptive gait generation, while the upper body progressively enhances disturbance robustness through curriculum learning-guided impedance control. This framework strengthens posture coordination under dynamic disturbances through progressive complex terrain training. We evaluate HumanDIL on the HIT-Hu humanoid robot, validating its generalization capabilities and imitation performance across multiple tasks.
Overview
HumanDIL consists of two parts: 1) a motion retargeting algorithm to generate motion references, and 2) an imitation learning process that utilizes the retargeted motion data along with a decoupled reward function. HumanDIL combines end-to-end IL with curriculum learning in an upper-lower body decoupled control framework. Reward functions guide the upper and lower body separately, while curriculum learning progressively structures training from simple to complex tasks for coordinated whole-body motion.
 Wave |
 Pick. |
Walk
Run
Traversing slope
Climbing stairs
-
Download Isaac Sim from the website with vision >= 4.0.0, then follow the installation instructions.
-
Following the instructions to install Isaac Lab.
-
Once Isaac Lab is installed, install the external dependencies for this repo:
pip install -r requirements.txt -
Install HIT_omniverse with pip by running:
pip install -e .
This project is licensed under the MIT License. Note that the repository relies on third-party code, which is subject to their respective licenses.