DMControl: One Million Transitions for Reinforcement Learning Tasks

DMControl 1M contains trajectories for reinforcement learning tasks, stored as separate .npz files. Each file contains one million transitions. The dataset was created by ml-jku and last updated on October 27, 2024.

Use Cases

• Training reinforcement learning agents based on the provided one million transitions per task.
• Benchmarking imitation learning or offline RL algorithms based on the stored trajectory data.
• Analyzing agent behavior in DMControl tasks based on the recorded state-action transitions.

Strengths

• Contains one million transitions per task, providing a substantial amount of data.
• Trajectories are stored in a structured .npz format per task for organized access.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown beyond the per-file transition count, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

ml-jku

Collection Method

Likely generated from simulations in DMControl environments.

Freshness

Last updated 2024-10-27 15:54:32; freshness should be verified.