cadene/droid_1.0.1: Large-Scale Franka Manipulation Dataset

The cadene/droid_1.0.1 dataset is a large-scale robot manipulation collection comprising 95,600 episodes and 27.6 million frames recorded from Franka robotic arms at 15 frames per second. Created using the LeRobot v2.1 framework, each episode pairs Parquet-serialized state-action trajectories with three synchronized video streams, yielding 286,800 total video files. The data is partitioned into 95 chunks of 1,000 episodes each to facilitate incremental loading and distributed training. All episodes are assigned to a single training split, and the dataset does not include explicit task labels or semantic annotations in the published metadata, leaving task segmentation and goal inference to downstream processing pipelines.

The dataset is released under the Apache-2.0 license, one of the most permissive open-source licenses available. Apache-2.0 grants you the right to use, reproduce, modify, and distribute the data in both research and commercial settings without paying royalties, obtaining additional permissions, or disclosing proprietary modifications. You may train foundation models, deploy teleoperation systems, or integrate learned policies into commercial robots and SaaS products. The only requirement is to include a copy of the Apache-2.0 license notice in any redistributed or derivative works. This makes cadene/droid_1.0.1 suitable for startups, enterprise R&D teams, and academic labs alike, with no restrictions on revenue generation or proprietary use.

Robotics teams building vision-language-action models, imitation learning policies, or world models for tabletop manipulation will find this dataset highly relevant, especially if they operate Franka robotic arms or can apply sim-to-real transfer techniques. The scale—95,600 episodes and 27.6 million frames—supports pre-training of large transformer-based policies and provides sufficient diversity for data augmentation experiments. Teams already using the LeRobot framework benefit from immediate compatibility with observation schemas and metadata conventions, reducing integration time. Commercial entities that require permissive licensing for proprietary model training will appreciate the Apache-2.0 terms, which impose no revenue-sharing or non-commercial restrictions. Organizations with established data pipelines for Parquet and multi-view video ingestion can onboard this dataset with minimal infrastructure changes.

If your deployment targets non-Franka manipulators—such as UR5, Kinova, or custom grippers—you may face kinematic mismatches that require adapter networks, domain randomization, or additional real-robot data collection to bridge the sim-to-real gap. The dataset does not provide task labels or goal annotations in the public metadata, so teams needing supervised task classification, language-conditioned goals, or semantic segmentation must invest in annotation pipelines or heuristic labeling before training. The 15 FPS capture rate may undersample fast dynamics like impact insertion or high-speed bin picking, making supplementary high-frequency datasets necessary for those behaviors. Finally, mobile manipulation, outdoor navigation, bimanual coordination, and multi-robot scenarios receive no coverage here, so teams focused on those domains should source alternative or complementary datasets to achieve task coverage.

The dataset is hosted on Hugging Face and can be downloaded using the Hugging Face datasets library or direct HTTP access to the repository at https://huggingface.co/datasets/cadene/droid_1.0.1. The chunked Parquet structure integrates naturally with PyTorch DataLoader, Dask, and cloud object stores like S3 or GCS, enabling efficient streaming and partial loading during distributed training. Because the data was generated with LeRobot v2.1, teams already operating LeRobot-based workflows inherit compatible schemas for observations, actions, and metadata, minimizing preprocessing boilerplate. For custom pipelines, you can parse the Parquet episode files and synchronize video frames using the published metadata in meta/info.json, which specifies chunk boundaries, episode indices, and video paths. No API keys, usage metering, or vendor approvals are required, so you can scale data ingestion predictably across on-premise clusters or cloud compute.

The dataset focuses exclusively on Franka arm interactions, so kinematic and gripper differences will require transfer learning or hardware-specific fine-tuning when deploying to other platforms. Task diversity and semantic annotations are not documented in the public metadata, meaning you must infer task boundaries from state-action sequences or apply unsupervised segmentation methods if goal-conditioned training is essential. The 15 FPS sample rate, while adequate for many manipulation primitives, may miss fast transients in contact-rich or high-speed tasks, potentially necessitating augmentation with higher-rate datasets or simulation. The absence of canonical train-test splits and standardized benchmarks places the burden of evaluation protocol design on your team, which can complicate reproducibility and cross-study comparisons. Finally, the dataset does not cover mobile bases, outdoor environments, or multi-agent scenarios, so pipeline architects must source additional data if those capabilities are in scope.