This repository provides deep reinforcement learning (RL) libraries and examples for NVIDIA Jetson platforms, targeting robotics applications. It enables users to train AI agents using PyTorch for tasks like object manipulation and navigation, with the goal of transferring learned behaviors from simulation to real-world robots.

How It Works

The project implements discrete Deep Q-Learning (DQN) and continuous A3G algorithms in PyTorch. It offers a C++ API that interfaces with Python for efficient tensor memory transfer (ZeroCopy), allowing RL agents to be integrated into robotics systems and simulators like Gazebo. The approach focuses on "pixels-to-actions" learning, where agents process raw visual input to make decisions.

Quick Start & Requirements

• Install: Build from source using cmake ../ && make within the build directory after cloning and initializing submodules.
• Prerequisites: NVIDIA Jetson TX1/TX2, JetPack 3.2, PyTorch v0.3.0.
• Setup Time: PyTorch compilation can take 30-60 minutes on a Jetson.
• Resources: Includes Jupyter notebooks for verification and standalone Python scripts for Gym environments. C++ examples like catch and fruit are provided. Gazebo simulations for robotic arms and rovers are also included.
• Docs: intro-pytorch.ipynb, intro-DQN.ipynb

Highlighted Details

• Implements both discrete (DQN) and continuous (A3G) RL algorithms.
• Features a C++ API for integrating RL agents into robotics applications with zero-copy tensor transfer.
• Demonstrates "pixels-to-actions" learning using raw visual input.
• Includes examples for OpenAI Gym (CartPole, Lunar Lander) and Gazebo simulations (robotic arm, rover).
• Supports Torch7/Lua scripting as an alternative to Python.

Maintenance & Community

The repository appears to be a foundational project from dusty-nv, with no explicit mention of active community channels or recent updates in the provided README.

Licensing & Compatibility

The README does not explicitly state a license. Compatibility for commercial use or closed-source linking would require clarification.

Limitations & Caveats

This project is explicitly tied to older versions: PyTorch v0.3 and JetPack 3.2. This significant version dependency may pose challenges for integration with modern hardware and software stacks. The README also notes that newer examples can be found in other repositories.