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

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