sohojoe / mujocounity

Reproducing MuJoCo benchmarks in a modern, commercial game /physics engine (Unity + PhysX).

MujocoUnity

Reproducing MuJoCo benchmarks in a modern, commercial game /physics engine (Unity + PhysX).

Presented March 19th, 2018 at the AI Summit - Game Developer Conference 2018 - http://schedule.gdconf.com/session/beyond-bots-making-machine-learning-accessible-and-useful/856147

---

IMPORTANT: Active development has moved to Marathon-envs

• This project is now a add-on for Unity ML-Agents Toolkit and is still being maintained and supported. See the new repro here - https://github.com/Unity-Technologies/marathon-envs

---

Legacy readme...

v0.2

supports Unity ml-agents (instructions)

Trained with ml-agents PPO:

unity_oai_hopper.xml

1.5m steps

unity_dm_walker.xml

3m steps

Trained with Baselines DDPG:

unity_oai_hopper.xml

300k steps 2m steps

unity_dm_walker.xml

1m steps 3m steps

Note: to reproduce you'll need to figure out how to patch OpenAI baselines with Unity.

Known Issues:

• oai_humanoid - is broken. Configurable joint needs updating to support multi-directional joints
• oai_half_cheetah - need to implement geom axis-angle
• dm_xxxx - need to implement class=
• phyx is not tuned properly
• Capsules are slightly too long - NOTE: this can cause collision issues whereby the leg may slightly poke the a foot and trigger collisions

From earlier version:

Ant - Random Ant - trained with ACKTR

mujocounity's Issues

Humanoid is broken

Configurable joint needs updating to support multi-directional joints

MujocoController.OnSensor is not working in ml-agents

issue is becuase ml-agents uses fixedUpdate so onSensor maybe cleared when the learning logic is called

Note: Analog uses update - need to find a solution that wont break this