RLDS, the Reinforcement Learning Datasets format, is a data specification for storing sequential decision-making data as episodes made of ordered steps, each containing an observation, action, reward, and related fields. It is designed to preserve the full temporal structure of an episode, which makes it a research-grade, lossless standard. It became widely used in robot learning after being adopted for large cross-embodiment datasets.
RLDS, short for Reinforcement Learning Datasets, is a specification for how to store the kind of data that describes an agent making decisions over time. Its core data model is built around two nested levels. A dataset is a collection of episodes, and each episode is an ordered sequence of steps. Every step records the essentials of one moment in an attempt, typically the observation the agent saw, the action it took, the reward it received, and related fields such as whether the episode terminated. By making these contents and their meaning explicit, RLDS gives everyone a shared, unambiguous way to represent experience.
The defining virtue of RLDS is that it preserves the complete temporal structure of an episode without discarding information, which is why it is often described as a lossless, research-grade standard. It was created to address the fragmentation that came from every group storing sequential data in its own idiosyncratic way, and it builds on established dataset infrastructure to provide versioning, documentation, and distribution. RLDS became especially prominent in robot learning after it was adopted for large datasets that pool experience across many robots, where a common, faithful format is essential.
Key takeaways
RLDS stores sequential decision-making data as episodes composed of ordered steps.
Each step holds an observation, action, reward, and related fields, and the format preserves full temporal structure.
It is a lossless, research-grade standard that became widely used after adoption for large cross-embodiment robot datasets.
What RLDS provides
The nested structure RLDS uses to represent experience.
The nested structure RLDS uses to represent experience.
Level
What it contains
Dataset
A collection of episodes
Episode
An ordered sequence of steps making up one attempt
Step
A single moment, with observation, action, reward, and related fields
How it works
RLDS organizes data so that the ordering within each episode is preserved and the meaning of every field is defined, rather than left to convention. Because it specifies contents explicitly, tools can reliably read a dataset and re-align or transform it into whatever form a particular training method expects, without guessing at what each field means. Building on mature dataset infrastructure gives RLDS versioning, documentation, and easy sharing, so datasets recorded by one group can be used confidently by another. The emphasis throughout is on faithfully capturing the sequence of observations, actions, and rewards that define an agent's experience.
Why it matters
RLDS matters because sequential data is only useful if its order and meaning survive storage and sharing, and RLDS was built to guarantee exactly that. For anyone curating robot learning data, it represents the lossless end of the spectrum, prioritizing completeness and research fidelity, which is why it underpins influential large-scale datasets. Understanding RLDS helps clarify why standardizing data formats is such an important enabler of pooling experience across projects and robots, a prerequisite for the cross-embodiment ambitions of modern robot learning.
Frequently asked questions
What does RLDS stand for?
RLDS stands for Reinforcement Learning Datasets. It is a specification and set of tools for storing, sharing, and using sequential decision-making data, and it has been widely adopted in robot learning.
What is the structure of an RLDS dataset?
An RLDS dataset is a collection of episodes, each episode is an ordered sequence of steps, and each step contains the observation, action, reward, and related fields for one moment. This nesting preserves the temporal structure of every attempt.
Why is RLDS described as lossless?
Because it preserves the full sequence and the meaning of every field rather than compressing or discarding information, RLDS keeps an episode's complete temporal structure intact. This research-grade fidelity is what distinguishes it from more compact, deployment-oriented formats.