Code to train a virtual character to juggle a soccer ball with reinforcement learning. Install my fork of RAISIM first and put the this project inside raisimGymTorch/raisimGymTorch/env/envs/.
vec_ppo.py allows you to train soccer juggling from scratch. It should take around 30000 iterations to obtain a good policy. Depends on your hardware setup, this could take between 12-30 hours.
Run test_policy.py to visualize a pretrain policy.
ActorCrititNetMann in model.py implements our layer-wise Mixture of Experts, which is mathematically equivalent to the weight-blended Mixture of Experts introduced in the paper Mode-Adaptive Neural Networks for Quadruped Motion Control. Our layer-wise interpretation allows for up to 10x speed up compared to the original implementation. This implementation is inspired by Motion VAE.