InSpatio, a startup founded in 2025 by SenseTime and specializing in spatial intelligence models, opened what it calls the first simulation-based embodied AI data collection center in China in Hangzhou on October 29. This center is aimed at producing one million hours of data monthly, which equates to approximately 100,000 hours per day.
According to InSpatio, robots lack not only data volume but also diversity. For instance, a robot can reliably grasp an object on a fixed shelf, but its efficiency decreases if the shelf rotates, the target object is obscured by another row of drinks, or the robot has to reach from another side.
Data from the Chinese Academy of Information and Communications Technology indicates that over 70 robot training arenas were operational in China by the end of June, with plans for another 40. However, many of these focus on simple tasks such as folding clothes and cleaning tables. Any change in layout requires rebooking the venue and operators, leading to increased costs.
InSpatio's approach changes the calculation formula from 'people multiplied by hours' to 'environment configurations multiplied by parallel computing.' The system starts with a small number of demonstrations performed by real robots or via teleoperation, and then combines motion planning, policy runs, and automated physics checks to generate closed trajectories in hundreds of parallel environments. These trajectories are accompanied by synchronized RGB, depth, segmentation, 6D pose, and action labels.
Scene assets are sourced from two channels. One channel utilizes a partnership with Hesai's Kosmo 3D world engine, which reportedly can scan a 100 square meter room in about 15 minutes with millimeter-level 3DGS reconstruction. InSpatio then segments this static scan into individual objects with defined physical properties. The second channel uses InSpatio's proprietary world simulator, Topos Pro, which converts single images, multiple images, video, LiDAR data, or aerial photography into object-level scenes where the table, chair, and equipment properties, including collision properties, can be selected, replaced, and assigned.
Scene processing involves separating target objects into individual instances, filling in hidden geometry and restoring the background, as well as calibrating mass, center of mass, and friction for each task. The company also defines three levels of service: L1 for navigation, L2 for dexterous grasping with millimeter-level geometry and dense occlusion, and L3 for contact and articulated dynamics, such as opening drawers. It is claimed that the advantage of simulation increases at higher levels, although final policies are still verified on real robots.
