cadene/droid_1.0.1 Franka robot dataset for VLA training

cadene/droid_1.0.1 is a large-scale robotics demonstration dataset containing 95,600 episodes of Franka robot arm behavior, totaling 27.6 million frames captured at 15 frames per second. The dataset was built using the LeRobot framework version 2.1 and stores each episode as a parquet file alongside synchronized video recordings, providing state trajectories, action sequences, and visual observations suitable for training manipulation policies, vision-language-action models, and world models. The data is organized into 95 chunks of 1,000 episodes each, enabling scalable data loading for distributed training workflows, and the entire collection is hosted on Hugging Face with over 429,000 downloads to date. The dataset focuses exclusively on the Franka robot platform, making it particularly valuable for teams already deploying Franka arms in research labs or production environments where transfer learning from this morphology will accelerate policy development.

cadene/droid_1.0.1 is released under the Apache License 2.0, a permissive open-source license that explicitly permits commercial use, modification, and distribution without royalty payments or copyleft obligations. This means robotics companies can train proprietary models on the dataset, deploy those models in commercial products, and sell integrated systems without needing to open-source their derivatives or seek additional permissions from the dataset authors. The Apache-2.0 license requires only that you include a copy of the license text and preserve any copyright notices when redistributing the dataset itself, but these requirements do not extend to trained model weights or robotic systems that merely use the data during development. Legal teams appreciate this license for its clarity and industry-standard language, which simplifies IP audits and reduces procurement friction when building supply chains around open datasets.

Teams operating Franka robot arms and building manipulation policies for tabletop tasks should prioritize this dataset, especially those seeking large-scale demonstration data under a permissive license that allows commercial deployment. The dataset is particularly well-suited for research labs training vision-language-action models that require millions of state-action pairs from a consistent robot morphology, as well as startups developing behavior cloning systems where data volume and licensing clarity are critical procurement criteria. Organizations already integrated with the LeRobot ecosystem will benefit from plug-and-play compatibility with existing data loaders and training scripts, reducing engineering overhead during the data ingestion phase. The 15 fps frame rate makes the dataset appropriate for manipulation tasks that do not require high-frequency control, such as pick-and-place operations, assembly sequences, and object rearrangement scenarios where human-scale motion dynamics are sufficient for policy learning.

cadene/droid_1.0.1 is not suitable for teams requiring multi-robot generalization, as the dataset contains only Franka arm demonstrations and will not provide the morphological diversity needed to train policies that transfer across UR5, xArm, or custom manipulator designs. Organizations building high-frequency control systems for tasks like dynamic manipulation, throwing, or contact-rich assembly may find the 15 fps capture rate insufficient to model the temporal dynamics of rapid motions or impact events. Teams that need explicit task labels or semantic annotations for task-conditioned policies will face additional annotation costs, since the dataset metadata does not include task categories or goal descriptions for the 95,600 episodes. If your deployment environment involves sensor modalities beyond RGB video—such as depth cameras, tactile arrays, or force-torque sensors—you should verify that the parquet schema includes those signals before committing to this dataset, as the Hugging Face metadata does not specify the full modality breakdown. Finally, organizations standardizing on non-PyTorch training frameworks may encounter integration friction with the LeRobot data format and should budget engineering time for building custom data loaders.

You can download cadene/droid_1.0.1 directly from Hugging Face using the datasets library or the Hugging Face CLI, with the data organized into 95 chunk directories containing 1,000 episodes each in parquet format plus synchronized video files. The LeRobot codebase provides native data loaders that parse the parquet schema and yield batches of observations, actions, and metadata for PyTorch training loops, minimizing the need for custom preprocessing code if you are already using that framework. Each episode is stored at the path pattern data/chunk-000/episode-000000.parquet, allowing you to load subsets of episodes for experimentation or distribute chunk loading across multiple GPU nodes for large-scale training runs. If you are using TensorFlow, JAX, or a custom training stack, you will need to write an adapter that reads the parquet files using libraries like PyArrow or Pandas, extracts the observation tensors and action vectors according to the LeRobot schema, and feeds them into your model's input pipeline.

After downloading cadene/droid_1.0.1, your data engineering team should first verify the integrity of all 95 chunk directories and confirm that episode counts match the documented 95,600 total, checking for corrupted parquet files or missing video streams that could cause training crashes. Inspect a sample of episodes to understand the observation space dimensions, action vector structure, and any metadata fields specific to the LeRobot schema, ensuring compatibility with your model architecture's input requirements. Because the dataset does not include pre-defined validation or test splits, you must manually partition episodes into training, validation, and test sets by withholding entire chunks to prevent data leakage, a critical step that requires coordination with your ML research team to establish statistically valid evaluation protocols. Run exploratory data analysis on action distributions, episode lengths, and visual scene characteristics to identify potential dataset biases or coverage gaps relative to your deployment environment, and consider visualizing trajectory rollouts to detect labeling errors or low-quality demonstrations that may degrade policy performance. Finally, execute a small-scale training run on a subset of episodes to benchmark data loading throughput and confirm that your infrastructure can sustain the I/O demands of streaming 27.6 million frames during full-scale experiments.