Aggregating millions of transitions collected from 8 distinct unsupervised exploration algorithms across multiple DeepMind Control Suite environments. It provides a benchmark for offline reinforcement learning by offering trajectories that lack task-specific rewards or expert demonstrations, focusing instead on diverse state-space coverage.
Use Cases
- Train offline RL agents using the observation and action columns to evaluate performance on non-expert data
- Compare the effectiveness of different exploration strategies by analyzing the reward distributions across the 8 included algorithms
- Develop reward-free pre-training models using the diverse state-space coverage provided in the exploratory trajectories
Strengths
- Includes trajectories from 8 exploration algorithms including ICM, RND, and APT
- Covers multiple DeepMind Control Suite environments such as Walker, Cheetah, and Jaco
- Contains 1 million transitions per environment-agent pair
- Provides observation, action, reward, and next_observation fields for every transition