D4RL PyBullet: Offline Reinforcement Learning Datasets for PyBullet Environments

4 PyBullet physics environments (HalfCheetah, Hopper, Walker2D, and Ant) providing offline reinforcement learning trajectories. The data consists of state-action-reward sequences generated by behavioral policies ranging from random to expert performance levels.

Use Cases

• Train offline reinforcement learning agents using the observations and actions arrays to learn policies without environment interaction.
• Benchmark reward modeling algorithms by analyzing the rewards scalar associated with specific state transitions.
• Evaluate temporal difference learning methods using the next_observations and terminals fields to compute target values.

Strengths

• Contains standard transition tuples including observations, actions, rewards, and terminal flags.
• Features multiple data quality tiers such as 'random', 'medium', 'medium-replay', and 'expert' trajectories.
• Supports four continuous control tasks: HalfCheetah, Hopper, Walker2D, and Ant.