D4RL Atari: Offline Reinforcement Learning Trajectories for Atari 2600

D4RL Atari provides logged experience trajectories for Atari 2600 games, serving as a wrapper for datasets originally released by Google Research. Updated in August 2024 by takuseno, it enables offline reinforcement learning research by providing fixed datasets of agent interactions.

Use Cases

• Training offline reinforcement learning models using logged state-action-reward trajectories
• Benchmarking batch RL algorithms against standard Atari 2600 pixel observations
• Evaluating policy optimization techniques using fixed transition data

Strengths

• Wraps official Google Research Atari datasets
• MIT licensed for open research
• Updated as recently as August 2024

Limitations

• Specific record counts and file sizes are not explicitly defined in the wrapper metadata
• Dependent on external Google-hosted data sources

Provenance

Source

Google Research

Collection Method

logged experience from trained DQN agents

Freshness

Last updated August 2024.