marathon-envs  by Unity-Technologies

Continuous control environments for Unity ML-Agents Toolkit

Created 7 years ago
311 stars

Top 86.5% on SourcePulse

GitHubView on GitHub
Project Summary

MarathonEnvs provides a suite of high-dimensional continuous control environments built with Unity's PhysX engine, targeting researchers in video games and robotics. It offers benchmarks for locomotion and AI control, enabling the application of advanced robotics research to game development and providing a Unity-based alternative to established suites like DeepMind Control Suite or OpenAI Gym's MuJoCo environments.

How It Works

The environments leverage Unity's native PhysX physics simulator for realistic, high-fidelity simulations. They are designed to be compatible with the Unity ML-Agents Toolkit and any OpenAI Gym-compatible reinforcement learning algorithm. The suite includes diverse environments inspired by research papers like DReCon for controller-based agents and DeepMimic for learning from motion capture, alongside procedurally generated terrains for generalization studies and classical locomotion tasks.

Quick Start & Requirements

  • Install: pip3 install mlagents==0.14.1
  • Prerequisites: Unity 2018.4 (or compatible version specified by release), ML-Agents Toolkit.
  • Setup: Clone the repository and build/install the correct runtime into the envs\ folder.
  • Docs: Getting Started, ML-Agents Guides, Web Demo

Highlighted Details

  • Diverse environment categories: Controller (DReCon), Style Transfer (DeepMimic), Procedural Terrains, Classical Locomotion, and Sparse Reward variants.
  • Supports training via Unity ML-Agents or standard Gym-compatible RL libraries.
  • Environments are inspired by and reference key research papers in character control and reinforcement learning.
  • Offers pre-trained models and baselines in some community projects.

Maintenance & Community

  • Maintained by Unity community contributors; not officially supported by Unity Technologies.
  • Community channels: Discord, Twitter: @iAmVidyaGamer.

Licensing & Compatibility

  • License: Not explicitly stated in the README. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

  • The project is community-maintained and not officially supported by Unity Technologies.
  • Specific environment versions are tied to particular Unity editor versions (e.g., v3.0.0 requires Unity 2020.1 beta.12), potentially leading to compatibility challenges with newer Unity releases.
Health Check
Last Commit

5 years ago

Responsiveness

Inactive

Pull Requests (30d)
0
Issues (30d)
0
Star History
5 stars in the last 30 days

Explore Similar Projects

Starred by Jiayi Pan Jiayi Pan(Author of SWE-Gym; MTS at xAI) and Jianwei Yang Jianwei Yang(Research Scientist at Meta Superintelligence Lab).

allenact by allenai

0%
369
Open-source framework for embodied AI research
Created 5 years ago
Updated 3 weeks ago
Feedback? Help us improve.