Training
RLDS (Reinforcement Learning Datasets)
ActivePublished
Key
innovation
A standardized format and tooling for episodic (step) data in RL and learning from demonstrations.
Category
Training
Abstraction level
Building block
Operation level
Training
How it works
RLDS defines an episode-and-step structure and provides libraries to create, transform and load such datasets (preserving episode-boundary information). It integrates with TFDS, easing versioning, sharing and streaming of large datasets into training pipelines.
Problem solved
RL and demonstration data are inherently sequential (episodes of steps) and heterogeneous; a standardized, efficient way to store, share and reuse them was lacking.