TimeChamber is a massively parallel self-play framework designed for large-scale training and evaluation of reinforcement learning agents in 3D physics-based environments. It targets researchers and engineers working with multi-agent systems, offering significant speedups and efficient resource utilization, particularly on a single GPU.

How It Works

TimeChamber leverages NVIDIA's Isaac Gym for highly parallelized GPU-based simulation, enabling thousands of environments to run concurrently on a single GPU. This approach achieves high throughput (e.g., 80,000+ FPS on an RTX 3070Ti with 4,096 environments). For evaluation, it employs vectorized models and a prioritized fictitious self-play mechanism to efficiently calculate Elo ratings for multiple policies, mitigating training cycles and enhancing policy diversity.

Quick Start & Requirements

• Installation: Follow Isaac Gym installation instructions, then pip install -e .
• Prerequisites: NVIDIA GPU with CUDA, Isaac Gym.
• Resources: Requires Isaac Gym setup. Training can be resource-intensive.
• Docs: Isaac Gym

Highlighted Details

• Achieves high FPS with thousands of parallel environments on a single GPU.
• Supports parallel Elo rating calculation for policy evaluation.
• Includes three competitive multi-agent tasks: Ant Sumo, Ant Battle, and Humanoid Strike.
• Implements a prioritized fictitious self-play algorithm for stable training.

Maintenance & Community

The project is hosted on GitHub by InspirAI. Links to community channels or roadmaps are not explicitly provided in the README.

Licensing & Compatibility

The repository does not explicitly state a license. The use of Isaac Gym implies adherence to NVIDIA's terms. Compatibility with commercial or closed-source projects is not specified.

Limitations & Caveats

The framework is heavily dependent on Isaac Gym, which requires specific NVIDIA hardware and CUDA versions. The project appears to be research-oriented, and its long-term maintenance or stability for production use is not detailed.