RLDS (Reinforcement Learning Datasets)

A Google-proposed standard format and toolchain for organizing sequential-decision data as episodes made of steps.

RLDS is a data ecosystem Google Research released in 2021 for recording, sharing, and processing reinforcement learning, offline RL, and imitation learning data. Data has two levels: an episode carries metadata such as an episode_id and contains a sequence of steps; every step has two required flags, is_first and is_last, plus optional fields such as observation, action, reward, and discount. A companion tool called EnvLogger records the data, and it is published and read through TensorFlow Datasets (TFDS). This lets data from different labs be read with the same code: every sub-dataset in Open X-Embodiment is provided in RLDS episode format, and the training code for Octo and OpenVLA reads it directly. The PyTorch ecosystem has increasingly moved toward LeRobotDataset in recent years, and the RLDS GitHub repository was archived in November 2025.

ExampleUse tfds.load to load the BridgeData V2 subset inside Open X-Embodiment, iterate episode by episode over its steps, and pull out each step's image, language instruction, and action for training.

Also called

RLDS

Related

Open X-Embodiment · TensorFlow Datasets (TFDS) · TFRecord · LeRobotDataset · Hierarchical Data Format version 5 · Episode

Sources

RLDS: an Ecosystem to Generate, Share and Use Datasets in Reinforcement Learning
google-research/rlds (GitHub)
google-deepmind/open_x_embodiment (GitHub)

As of

2025-11