This PR adds support for pixel observations in DMControl tasks, and replaces the previous replay buffer with a faster implementation, resulting in approx 1.5x faster training times for state observations and 8x for pixel observations. This PR also adds other minor quality-of-life improvements such as the ability to only install dependencies for environments that one wishes to use (rather than all 4 environments).
This PR adds support for pixel observations in DMControl tasks, and replaces the previous replay buffer with a faster implementation, resulting in approx 1.5x faster training times for state observations and 8x for pixel observations. This PR also adds other minor quality-of-life improvements such as the ability to only install dependencies for environments that one wishes to use (rather than all 4 environments).