RoboSense and BitRobot have released an open dataset built specifically for embodied AI training, capturing real work in three messy commercial environments with depth, colour and motion recorded together.
The release came during CoRL 2026 in Austin, where the two companies announced a strategic cooperation covering 3D perception hardware, distributed data collection and an open robotics development network.
What is in the dataset
The first output of the cooperation is RoboCap LiDAR-100, released by BitRobot at the conference. It covers restaurant kitchens, vehicle repair and car detailing, with a total collection volume of 100 hours.
The distinguishing feature is what gets recorded alongside the images. Each episode captures lidar depth information, RGB imagery and inertial measurement data simultaneously, so a model is trained on geometry and motion rather than inferring three dimensions from two.
The dataset is live on Hugging Face. The public page shows the first 50 hours open, containing 1,964 task episodes across 37 task categories, with the remaining 50 hours scheduled for later release.
Why these three environments
Restaurant kitchens, repair bays and detailing shops share a set of properties that laboratory datasets systematically exclude. Surfaces are cluttered and partially occluded. Objects are wet, hot, greasy or fragile. Tools are picked up and put down in arbitrary order. The human operator’s body blocks the camera for much of the task.
Those are precisely the conditions where vision-only models fail. A model trained on clean tabletop manipulation learns to grasp objects it can see clearly. A model trained on a kitchen learns what to do when the pan is behind a stack of plates and the handle is out of frame.
RoboSense also used the conference to show a new spatial camera category at BitRobot’s booth, positioning its 3D perception hardware as an input layer for robot learning rather than purely an automotive sensor.
The partner network behind it
BitRobot has previously established ecosystem partnerships with Hugging Face and Unitree, and co-initiated RoboColosseum, an open evaluation programme that puts robot models through real-world testing rather than benchmark suites.
For RoboSense, the move extends a lidar business into data infrastructure. Hardware margins compress as sensor volumes rise. Owning part of the data pipeline that trains the models those sensors feed is a way to stay in the value chain as the automotive lidar price war continues.
For anyone training embodied models outside China, the practical consequence is immediate. A hundred hours of messy, physically accurate, multi-sensor commercial work is a different asset class from another simulated grasping set, and it is downloadable.
Editor’s note: This is an adapted translation of the original Gasgoo report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.gasgoo.com/apps/robot/news/.