Source-backed decision guide · updated 2026-09-02

LeRobot vs RLDS vs MCAP for Robot Data

Choose a robot-data representation by workflow stage instead of treating LeRobot, RLDS, and MCAP as interchangeable file formats.

Direct answer

Use MCAP to preserve synchronized raw robotics logs, LeRobot when the target workflow is the Hugging Face robot-learning ecosystem, and RLDS when the training stack already consumes TensorFlow-style episodic datasets. They can form a pipeline: retain raw MCAP, then publish a versioned training view in LeRobot or RLDS.

Comparison

Choose by the artifact you need to preserve and the system that will consume it.

Option	Best for	Strength	Watch for
MCAP	Raw capture and replay	Preserves timestamped heterogeneous robot topics in one log container.	A log is not automatically a training-ready episode schema or label contract.
LeRobot	Hugging Face training and sharing	Fits LeRobot policies, dataset tooling, Hub distribution, and common video-plus-tabular training views.	Conversion must preserve timing, embodiment, feature semantics, and source revision.
RLDS	TensorFlow episodic pipelines	Represents steps inside episodes with explicit observation, action, reward, and terminal fields.	The surrounding TensorFlow data stack may be a poor fit for non-TF consumers.

MCAP

Best for: Raw capture and replay
Strength: Preserves timestamped heterogeneous robot topics in one log container.
Watch for: A log is not automatically a training-ready episode schema or label contract.

LeRobot

Best for: Hugging Face training and sharing
Strength: Fits LeRobot policies, dataset tooling, Hub distribution, and common video-plus-tabular training views.
Watch for: Conversion must preserve timing, embodiment, feature semantics, and source revision.

RLDS

Best for: TensorFlow episodic pipelines
Strength: Represents steps inside episodes with explicit observation, action, reward, and terminal fields.
Watch for: The surrounding TensorFlow data stack may be a poor fit for non-TF consumers.

Do not discard the acquisition truth

Training schemas are optimized views, while capture logs preserve what the robot emitted. Keeping the immutable raw source makes it possible to repair a conversion, add a missing signal, or reproduce a derived dataset later.

Version the conversion boundary

Record the source-log identity, converter version, feature mapping, clock alignment, dropped topics, filtering rules, and output revision. Without that lineage, two artifacts with the same dataset name may describe different training evidence.

Decision checklist

01Name the downstream policy or data loader before choosing the published training representation.
02Keep an immutable source revision and a machine-readable conversion manifest.
03Verify timestamp alignment, episode boundaries, action units, camera encoding, and missing-value policy.
04Run readiness checks on the derived artifact instead of assuming a successful conversion is usable.

Reviewed sources

Reviewed by OpenBot Catalog team

Hugging Face LeRobotUpstream robot-learning library and dataset conventions.Google Research RLDSUpstream episodic dataset specification and builders.MCAP specification and documentationTimestamped multimodal robotics log container.