🤖 Behavior cloning framework based on Demonstration-Aware Machine Learning (DAML) System

Author: Alejandro Alemany

The Demonstration-Aware Machine Learning (DAML) Framework is designed for human motion capture, processing, and robot trajectory learning. The system processes multi-view pose data, applies machine learning techniques, and integrates with robot simulation environments for training robotic control models.

This project includes:

• Pose extraction & filtering using MediaPipe
• Multi-view pose fusion for robust motion tracking
• Deep learning-based trajectory learning using PyTorch
• Robot control simulation in MuJoCo
• Scalable training pipeline with modular components

Ensure you have the following installed:

• Python 3.9
• PyTorch
• NumPy, OpenCV, Matplotlib
• MuJoCo (for simulation)
• MediaPipe (for pose extraction)

pip install -r requirements.txt

Download MuJoCo from mujoco.org Extract and install dependencies (mujoco-py for Python integration) Set up MUJOCO_PATH in your environment

Extracts poses from videos and saves processed data:

python scripts/process_demonstrations.py --input demonstrations/ --output processed_demonstrations/

Trains the DAML model on processed pose data:

python scripts/train_model.py

Runs the trained model inside a MuJoCo simulation:

python scripts/test_robot.py

Each module has a dedicated documentation file:

• 📌 DAML Framework
• 🎯 Pose Processing
• 🎨 Visualization
• 🏋️ Training Pipeline
• 🛠 Testing & Simulation
• 🔧 Utilities

🔗 References MuJoCo: https://mujoco.org/ PyTorch: https://pytorch.org/ MediaPipe: https://developers.google.com/mediapipe