Standard luggage for offline RL trajectories

What it does Minari defines a common format for offline reinforcement learning datasets and bundles popular reference sets such as D4RL. It wraps Gymnasium environments with a DataCollector to record episodes, then exposes observations, actions, rewards, terminations, and infos as structured episode data. Several training libraries already read the format natively.

The interesting bit The library treats recorded trajectories like Gymnasium treats live environments: you iterate over episodes instead of stepping through a simulator. The name even nods to the Japanese “Minarai,” meaning learning by observation—an apt metaphor for offline RL.

Key highlights

• Curated remote datasets (including D4RL) available out of the box
• DataCollector wrapper turns any Gymnasium env into a dataset generator
• Episode objects expose full trajectory metadata: observations, actions, rewards, terminations, truncations, and infos
• Adopted by TorchRL, d3rlpy, and AgileRL
• Maintained by the Farama Foundation alongside Gymnasium

Caveats

• The README never specifies the underlying serialization format or storage backend
• Dataset sizes, bandwidth requirements, and hosting guarantees are not discussed

Verdict A sensible starting point if you need to share or consume offline RL data across projects. If your pipeline is already online-only or deeply custom, you probably don’t need another abstraction.