Dataset profile
IPEC-COMMUNITY/bridge_orig_lerobot Dataset Profile
IPEC-COMMUNITY/bridge_orig_lerobot is a large-scale robotic manipulation dataset containing 53,192 episodes across 19,974 distinct tasks, totaling 1,893,026 frames of WidowX robot video captured at 5 fps. Released under the Apache-2.0 license, this dataset enables commercial training of vision-language-action models, behavior cloning policies, and world models for tabletop manipulation scenarios. Robotics teams building VLA systems can use this data to learn diverse pick-and-place, push, and arrangement behaviors in kitchen and tabletop environments, with clear licensing that supports both research and production deployment.
Quick facts
- Scale
- 53,192 episodes, 1.89M frames
- License
- Apache-2.0
- Format
- Parquet + video (LeRobot v2.0)
- Modality
- Video
- Robot platform
- WidowX
- Commercial use
- Permitted
Dataset composition and structure
The dataset comprises 53,192 demonstration episodes spanning 19,974 unique manipulation tasks performed by a WidowX robotic arm. Each episode is stored in Parquet format with accompanying video files, organized into 54 chunks of approximately 1,000 episodes each for efficient streaming and loading. Video data is captured at 5 frames per second, yielding 1,893,026 total frames across 212,768 video files. The tasks cover common tabletop manipulation primitives including object grasping, placement, pushing, and multi-step arrangement sequences in kitchen and workspace settings. All data follows the LeRobot v2.0 codebase structure, ensuring compatibility with modern vision-language-action training pipelines. The dataset includes only a training split, with the full range of episodes available for model development. Episode metadata, robot configuration, and structural information are documented in meta/info.json, providing teams with the technical specifications needed for loader implementation and data pipeline integration.
Licensing and commercial deployment rights
IPEC-COMMUNITY/bridge_orig_lerobot is distributed under the Apache-2.0 license, one of the most permissive open-source licenses available for machine learning datasets. This licensing grants teams full rights to use, modify, and distribute derivatives of the data in both research and commercial products without royalty obligations. Robotics companies training production VLA models, behavior cloning systems, or world models can deploy policies learned from this dataset in commercial hardware without licensing restrictions or attribution requirements beyond those specified in Apache-2.0 terms. The permissive license removes procurement friction common with restrictive academic datasets, enabling faster path-to-production for manipulation AI. Teams building foundation models for physical AI or offering robot-learning-as-a-service can incorporate this data into training corpora that power revenue-generating products. Legal and compliance teams will find the Apache-2.0 terms straightforward to review, as the license is widely adopted across enterprise AI infrastructure and presents minimal liability surface compared to custom or academic-only licenses.
Procurement considerations for manipulation teams
With over 738,000 downloads, this dataset represents one of the most widely adopted manipulation training corpora in the robotics community, indicating both quality and integration maturity. The LeRobot v2.0 format ensures compatibility with Hugging Face datasets infrastructure, enabling teams to stream data directly into PyTorch or JAX training loops without custom ETL pipelines. The WidowX platform is a common low-cost research arm, making the kinematic and visual domain directly transferable for teams prototyping on similar hardware or using sim-to-real techniques with WidowX digital twins. However, procurement teams should note that the dataset size category is listed as unknown in metadata, and while episode counts are documented, storage requirements and bandwidth needs for the full 212,768 video files should be estimated before pipeline deployment. The chunked structure with 54 segments supports incremental download and distributed training, but teams with limited storage should plan for multi-terabyte capacity. Organizations seeking vendor support, custom licensing, or extended versions of this data should evaluate whether community-maintained datasets align with their support and SLA requirements, or whether commercially supported alternatives better fit enterprise procurement policies.
Known limitations and scope boundaries
The dataset is limited to a single robot platform, the WidowX arm, which constrains direct transfer to other manipulator morphologies without domain adaptation or fine-tuning on target hardware. Tasks are confined to tabletop and kitchen environments, so teams building manipulation policies for warehouse, outdoor, or human-scale environments will find limited direct applicability without supplementary data. The 5 fps capture rate is suitable for relatively slow manipulation tasks but may undersample fast dynamics like rapid grasping or impact events, potentially limiting policy performance on time-sensitive tasks. Episode success labels and task conditioning information are not explicitly surfaced in the provided metadata, so teams requiring success-filtered demonstrations or language-conditioned training may need to parse task identifiers or augment with external annotations. The dataset originated from the Bridge data collection effort and was converted to LeRobot format by the community, meaning provenance and collection protocols follow the original Bridge project methodology. Teams requiring calibration data, force-torque measurements, or proprioceptive sensor streams beyond visual observations should verify whether their required modalities are present in the Parquet episode files, as the high-level metadata highlights video as the primary modality.
Related pages
Use these to move from category-level context into specific task, dataset, format, and comparison detail.
FAQ
What is the IPEC-COMMUNITY/bridge_orig_lerobot dataset?
IPEC-COMMUNITY/bridge_orig_lerobot is a robotic manipulation dataset containing 53,192 teleoperation episodes of a WidowX robot arm performing 19,974 distinct tabletop tasks. The dataset provides 1,893,026 frames of video data at 5 fps, structured in the LeRobot v2.0 format for compatibility with modern vision-language-action training frameworks. Originally derived from the Bridge dataset and converted for the LeRobot ecosystem, it offers a large-scale resource for training behavior cloning policies, VLA models, and manipulation world models under the permissive Apache-2.0 license.
Can I use this dataset to train commercial robotic products?
Yes, the Apache-2.0 license explicitly permits commercial use, modification, and distribution of derivative works without royalty payments or restrictive attribution requirements. Robotics companies can train production models on this data and deploy the resulting policies in commercial hardware, software-as-a-service platforms, or licensed AI products. The permissive terms remove common procurement blockers associated with academic-only or non-commercial licenses, making this dataset suitable for startups, enterprise robotics divisions, and foundation model vendors building revenue-generating physical AI systems.
Who should use the IPEC-COMMUNITY/bridge_orig_lerobot dataset?
This dataset is ideal for robotics teams building vision-language-action models, imitation learning systems, or world models for tabletop manipulation tasks on WidowX or kinematically similar arms. Research labs prototyping behavior cloning architectures, companies developing general-purpose manipulation policies, and ML engineers training transformer-based robot controllers will find the scale and format well-suited to their pipelines. Teams using Hugging Face infrastructure, LeRobot tooling, or PyTorch-based training loops benefit from native compatibility. The dataset is also valuable for benchmark evaluations, ablation studies on manipulation datasets, and pre-training before fine-tuning on proprietary task distributions.
When is this dataset NOT the right choice for my project?
Teams building policies for non-tabletop environments such as warehouses, outdoor settings, or human-scale manipulation should look elsewhere, as the task distribution is confined to kitchen and desk scenarios. If your robot platform differs significantly from the WidowX morphology, direct transfer may be poor without domain adaptation or supplementary data from your target hardware. Projects requiring high-frequency control, force sensing, tactile feedback, or proprioceptive modalities beyond video will find this dataset insufficient, as video is the primary modality and the 5 fps capture rate limits temporal resolution. Organizations that require vendor support, service-level agreements, or custom licensing terms may find community-maintained datasets misaligned with enterprise procurement policies.
How is the dataset structured for training pipelines?
Episodes are stored in Parquet files organized into 54 chunks of approximately 1,000 episodes each, located at data/chunk-XXX/episode_YYYYYY.parquet. Video files accompany each episode, totaling 212,768 video assets across the dataset. Metadata is provided in meta/info.json with specifications including robot type, frame counts, FPS, and chunk structure. This organization supports efficient streaming, distributed loading, and incremental downloads, allowing teams to load subsets during development or stream from Hugging Face datasets without downloading the full corpus. The LeRobot v2.0 schema ensures compatibility with standard dataloaders and the broader LeRobot training ecosystem.
What are the storage and bandwidth requirements for using this dataset?
While the metadata does not specify total dataset size, teams should plan for multi-terabyte storage given the 212,768 video files and 1,893,026 frames. The chunked structure with 54 segments enables incremental download, so prototyping can begin with a subset of chunks before scaling to the full dataset. Network bandwidth for initial download and ongoing streaming should be estimated based on video resolution and compression, particularly for distributed training across multiple nodes. Teams with storage constraints can leverage Hugging Face's streaming mode to load data on-demand during training, though this introduces network latency into the data pipeline and requires stable high-bandwidth connections for efficient throughput.
Need data like IPEC-COMMUNITY/bridge_orig_lerobot Dataset Profile?
If your project needs similar modality, scale, or licensing, truelabel can surface comparable open datasets or match you with capture partners that deliver to spec.
Access dataset on Hugging Face