v1.0

Pretrained models for DeepMind Control Suite environments. Note that performance is roughly comparable to the original, except for ball_in_cup-catch (which has high variance).

Some models are trained for 2000 episodes as per previous versions of the paper, not 1000 episodes as per the final version (but all model checkpoints provided are from 1000 episodes). Does not include all preprocessing steps (5-bit quantisation and centering of observations).

cartpole-balance

cartpole-swingup

reacher-easy

finger-spin

cheetah-run

ball_in_cup-catch

walker-walk