FastTD3: A High-Performance C++ Implementation of Twin Delayed DDPG (TD3)

This release introduces FastTD3, a lightweight and optimized C++ implementation of the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, designed for speed and modularity. Built with LibTorch and structured for research and production readiness, FastTD3 enables fast experimentation and integration with simulation environments such as MuJoCo.

Features

• Pure C++17 implementation using LibTorch (PyTorch C++ frontend)
• Modular design: easy to customize actor, critic, and replay buffer
• Support for deterministic continuous control tasks
• Efficient training pipeline with batch updates and noise injection
• Ready to integrate with simulators like MuJoCo or custom environments

What's Included

• TD3 agent implementation (TD3Agent)
• Actor-Critic neural network architecture
• Replay buffer with uniform sampling
• Configurable training loop
• CMake build system with presets
• Example usage with dummy environments

📈 Performance

FastTD3 is designed for speed and scalability, suitable for both academic research and real-time control applications.