This repository implements DrQ-v2 for pixel-based observation tasks on DMControl.

Installation

See the original Jaxrl repo for instructions.

Usage

cd examples
MUJOCO_GL=egl python train_pixels.py --env_name=quadruped-run --save_dir=./tmp/

To track experiments with Weights and Biases, append --track to the above command.

Tune the hyperparameters in examples/configs/drq_v2.py

Results

Verified performance in quadruped-run, quadruped-walk, acrobot-swingup, cheetah-run, and reacher-hard.

Wandb project opening soon....