ad1t7a/10Kh-RealOmin-OpenData: Large-Scale Real-World Dual-Arm Robotics Dataset

The 10Kh-RealOmin-OpenData dataset is a large-scale collection of over 10,000 hours and more than one million video clips documenting dual-arm robotic manipulation tasks in real household environments. Published by ad1t7a on Hugging Face, it is presented as the largest open-source embodied intelligence dataset available, with data gathered from over 3,000 distinct homes and nearly 10,000 fine-grained manipulation targets. The dataset is designed to support training of vision-language-action models, world models, and reinforcement-learning agents that must generalize across diverse in-home service scenarios.

This dataset is released under the Creative Commons Attribution-ShareAlike 4.0 International license, which permits commercial use with two key obligations: you must provide appropriate attribution to the original dataset creators, and any derivative works—including trained models or modified datasets—must be shared under the same CC-BY-SA-4.0 terms. For robotics companies, this means that if you train a VLA model on this data and then distribute that model publicly or commercially, the model weights and associated training artifacts must also carry the CC-BY-SA-4.0 license. Internal use, research prototypes, and service deployments where the model is not redistributed are generally permissible, but legal review is recommended to ensure compliance with your organization's IP strategy.

Teams building vision-language-action models or world models for in-home service robots will find this dataset especially valuable because of its scale, diversity, and authentic household scenarios. The 3,000+ home environments and nearly 10,000 manipulation targets provide the variation needed to train models that generalize beyond lab-controlled settings. Dual-arm manipulation researchers benefit directly from the embodiment match, while reinforcement-learning practitioners gain access to long-horizon task sequences captured in video. Organizations with share-alike-compatible licensing strategies and sufficient infrastructure to handle terabyte-scale video datasets should prioritize this resource. Teams targeting warehouse automation, outdoor agriculture, or single-arm mobile manipulators may need to augment with domain-specific data to achieve comparable performance.

This dataset is not ideal if your deployment environment differs significantly from residential households—warehouse, hospital, or outdoor agricultural settings will exhibit different object distributions, lighting, and spatial layouts that are underrepresented here. Single-arm or mobile-manipulator teams will need to invest in domain adaptation since all clips assume a dual-arm embodiment. If your policy requires tactile feedback, force-torque sensing, or proprioceptive signals, the video-only modality will necessitate fusion with additional sensor datasets or simulation. Finally, organizations with strict proprietary-model requirements may find the CC-BY-SA-4.0 share-alike clause incompatible with their IP strategy, especially if the roadmap includes commercial distribution of pre-trained manipulation models.

Given that this dataset exceeds one terabyte, procurement and ML engineering teams should plan for substantial download time, storage capacity, and bandwidth costs, particularly when moving data into cloud training environments. Inspecting a representative sample subset before committing to a full download is critical to confirm frame rate, resolution, camera angles, and clip length align with existing data pipelines. Platforms like Truelabel offer partial-download and streaming capabilities that reduce upfront infrastructure burden and allow iterative exploration. Teams should also budget for data versioning and backup, since re-downloading the full dataset after an accidental deletion can incur significant time and cost penalties.

The dataset's Hugging Face page provides a high-level description emphasizing scale, household diversity, and dual-arm embodiment, but detailed technical specifications—such as frame rate, resolution, camera intrinsics, action-label schemas, or task taxonomies—are not exhaustively documented in the truncated description available. Teams should inspect sample clips early to reverse-engineer these parameters and confirm compatibility with training frameworks. If you source through Truelabel, you gain access to structured metadata exports, compliance summaries, and provenance records that streamline internal review and satisfy audit requirements. Always verify that privacy and consent documentation for the 3,000+ household environments meets your jurisdiction's data-protection regulations before deploying models trained on this data in production.