Procgen Benchmark Expert Trajectories from PPO Agent

Procgen Benchmark provides expert trajectories generated by a PPO reinforcement learning agent trained on 16 procedurally-generated gym environments. The dataset, created by EpicPinkPenguin, was last updated on March 8, 2026. Environments were created on 'easy' distribution mode with unlimited levels.

Use Cases

• Train imitation learning models using the expert state-action trajectories from environments like 'bigfish' and 'chaser'.
• Benchmark offline reinforcement learning algorithms on pre-collected PPO agent data across 16 distinct Procgen environments.
• Analyze agent behavior and policy performance using the recorded trajectories for each game level.
• Perform behavioral cloning by learning from the expert actions provided in the trajectory data.

Strengths

• Contains data from 16 distinct procedurally-generated environments.
• Trajectories generated by a trained PPO agent, providing expert demonstrations.

Limitations

• Specific row counts, column names, and dataset size are unknown.
• Limited to the 'easy' distribution mode of the Procgen environments.
• Potential bias from using a single PPO agent's policy for trajectory generation.

Provenance

Source

EpicPinkPenguin on Hugging Face.

Collection Method

Expert trajectories generated by a PPO reinforcement learning agent trained on each Procgen environment.

Freshness

Last updated March 8, 2026.