This project provides a powerful integration of PyTorch and the Orocos Kinematics and Dynamics Library (KDL), all conveniently packaged within a Docker container. It is designed for researchers and developers working on robot kinematics, manipulation, and vision-based task training.

For experimental use only.

• About The Project
• Key Features
• Getting Started
  • Prerequisites
  • Installation and Setup
• Usage
  • Jupyter Lab Environment
  • Keyframe Extraction
  • Vision-Based Tasks
• Roadmap
• Contributing
• License
• Acknowledgments

PyTorch-KDL bridges the gap between classical robotics and modern deep learning frameworks. By wrapping the robust solvers of KDL within a PyTorch-compatible interface, this library allows for seamless gradient-based optimization of kinematic chains, making it ideal for a new class of robotic learning tasks.

This entire environment is containerized using Docker, ensuring reproducibility and eliminating complex setup procedures.

• PyTorch-Powered Kinematics: Leverage PyTorch's automatic differentiation with KDL's reliable kinematics solvers.
• Dockerized Environment: Get up and running in minutes with a pre-configured Docker container that includes all necessary dependencies.
• Jupyter Lab Integration: Interactively develop and test your robotics algorithms in a web-based Jupyter environment.
• Vision-Based Manipulation Tasks: Includes example implementations for vision-based pushing and grasping tasks.
• Keyframe Extraction: Tools to extract keyframes from demonstrations for imitation learning.

Follow these steps to build the Docker image and launch the container.

You must have Docker installed on your system.

• Install Docker Engine

1. Clone the repository:

   git clone [https://github.com/your_username/pytorch-kdl.git](https://github.com/your_username/pytorch-kdl.git)
   cd pytorch-kdl

2. Build the Docker image: The Dockerfile provided in the root of this project contains all the necessary build steps.

   docker build -t pytorch-kdl .

3. Run the Docker container: This command will start the container, map the necessary ports for Jupyter, and mount the current directory into the container's workspace.

   docker run -it -p 8888:8888 -v "$(pwd)":/workspace pytorch-kdl

After running the container, you will see a URL in your terminal that looks something like this: http://127.0.0.1:8888/lab?token=...

Copy and paste this URL into your web browser to access the Jupyter Lab interface. All your project files will be available in the /workspace directory.

The notebooks/ directory contains examples for using the keyframe extraction tools. These can be used to process robot demonstration data for behavioral cloning or other imitation learning techniques.

Explore the tasks/ directory to find the implementation of the vision-based pushing and grasping tasks. You can run these examples from the Jupyter notebooks to train and evaluate manipulation policies.

Example:

# Inside a Jupyter notebook
from tasks.pushing import PushingTask

# Initialize and run the pushing task
push_task = PushingTask()
push_task.run_training()