Humanoid Imitation Learning

This repository implements imitation learning for a humanoid character using motion capture data and PyBullet physics simulation.

Setup

# Clone the repository
git clone https://github.com/your-username/humanoid_imitation.git
cd humanoid_imitation

# Install dependencies
pip install -r requirements.txt

Running the Pre-trained Model

1. Download the pre-trained model

Download the model folder from Google Drive and place it in the root directory.

2. Run the model

python test_model.py --model_dir model --model_name best_model --render

Add --slowmo 1.5 to slow down the animation for better visualization:

python test_model.py --model_dir model --model_name best_model --render --slowmo 1.5

To record a video of the simulation:

python test_model.py --model_dir model --model_name best_model --render --record --video_name enhanced_walk_demo

Training Your Own Model

./run_improved_training.sh

Monitor Training Progress

# View training metrics for the latest run
tensorboard --logdir=logs/$(ls -t logs | head -1)

Access http://localhost:6006 in your browser to view training metrics.

Behaviour Cloning

This repository includes a behavior cloning implementation to learn from expert demonstrations using a neural network policy.

Prerequisites

Ensure you've installed all the required dependencies:

pip install -r requirements.txt

Running Behavior Cloning

Training Mode

To train a policy using behavior cloning:

python3 behaviour_cloning/run_imitation_learning.py train \
    --env jvrc_walk \
    --expert-model <path_to_expert_model>/actor.pt \
    --logdir <output_directory> \
    --n-itr 500 \
    --num-procs 12

Additional training options:

• --env: Environment type (jvrc_walk or jvrc_step)
• --expert-model: Path to the expert policy model to imitate
• --logdir: Directory to save weights and logs
• --n-itr: Number of iterations for training (default: 500)
• --num-procs: Number of parallel processes for training (default: 12)
• --recurrent: Use LSTM network instead of feedforward
• --yaml: Path to config file for the environment

Evaluation Mode

To evaluate a trained behavior cloning policy:

python3 behaviour_cloning/run_imitation_learning.py eval \
    --path <path_to_trained_model>/actor.pt \
    --out-dir model_evaluation \
    --ep-len 10

Evaluation options:

• --path: Path to the trained model (either directory containing actor.pt or direct path)
• --out-dir: Directory to save evaluation videos
• --ep-len: Episode length in seconds (default: 10)

Example

Run the provided pre-trained model:

python3 behaviour_cloning/run_imitation_learning.py eval --path model/actor.pt --out-dir model_evaluation

Alternative Evaluation Script

The repository also includes an alternative evaluation script that runs multiple episodes and reports average performance:

python3 behaviour_cloning/evaluate_imitation.py --model_path model/actor.pt

Options:

• --model_path: Path to the trained policy model file (required)
• --num_episodes: Number of episodes to evaluate (default: 5)
• --no_render: Disable rendering to run without visualization

Example with no rendering (useful for headless servers or performance testing):

python3 behaviour_cloning/evaluate_imitation.py --model_path model/actor.pt --no_render

References:

1. https://github.com/rohanpsingh/LearningHumanoidWalking
2. https://github.com/xbpeng/DeepMimic.git