A dataset compiler that converts heterogeneous robot trajectories (RLDS, LeRobot, Open X-Embodiment) into a streamable storage format with strict validation and sliceable metadata.
- Multi-format ingestion: Load from RLDS/TFDS, LeRobotDataset v3, Open X-Embodiment
- Canonical representation: RLDS-aligned Episode/Step structure
- Strict validation: Structural, semantic, and statistical checks
- LeRobotDataset v3 output: Parquet + MP4 + relational metadata for Hub streaming
- RLDS/TFDS export: Secondary format for ecosystem compatibility
- Sliceable metadata: Query and filter by task, robot, episode properties
- Cross-embodiment transforms: Camera selection, action mapping, normalization
# Core installation
pip install embodied-datakit
# With TensorFlow Datasets support (for OXE)
pip install embodied-datakit[tfds]
# With video encoding
pip install embodied-datakit[video]
# Full installation
pip install embodied-datakit[all]# Probe a dataset
edk ingest oxe://berkeley_autolab_ur5
# Compile to LeRobotDataset v3
edk compile oxe://berkeley_autolab_ur5 --split "train[:100]" -o ./compiled/ur5
# Inspect the result
edk inspect ./compiled/ur5
# Export to RLDS format
edk export-rlds ./compiled/ur5 -o ./rlds/ur5See docs/ for detailed documentation:
- Upstream Formats - RLDS, LeRobot v3, OXE requirements
- Compiler Contract - Transformation guarantees
- CLI Reference - Command-line interface
- Validation Policy - Validation rules and severity
Proprietary – Research/Portfolio Use Only
This software is provided under a restrictive proprietary license:
- ❌ No commercial or professional use
- ❌ No redistribution or derivative works
- ✅ Non-commercial research with required citation
- ✅ Personal evaluation
If you use this software in academic research, you must cite it:
@software{bharadwaj2026embodieddatakit,
author = {Bharadwaj, Varun},
title = {EmbodiedDataKit},
year = {2026},
url = {https://github.com/waycodes/embodied-agent}
}See LICENSE and CITATION.cff for full terms.