Kosmo, Ropedia Partner on Embodied AI Training Data
The collaboration pairs Hesai's 3D spatial capture with Ropedia's action and annotation pipeline to address a data gap in robot training.
Updated on Nov. 17, 2026 in Robotics

Kosmo, the 3D world data engine under Hesai Technology, has partnered with physical AI infrastructure company Ropedia on an end-to-end training data pipeline for embodied AI. The collaboration spans spatial data capture, structured processing, standardized delivery, and model validation.
Why it matters
Embodied AI models need data that combines precise spatial geometry with human-object interaction, and current 3D reconstruction and first-person video each supply only one of these. The partnership targets that gap by joining the two data types into a single pipeline.
The Ropedia Xperience 2 project comprises 1,589 scanned scenes, 74 tasks, and 4,230 first-person records totaling 131 hours of footage, all captured by Kosmo One. Whether the dataset improves downstream model performance has not been reported.
The players
Hesai Technology
Hesai Technology is a Chinese lidar and sensing company whose Kosmo division builds 3D world data infrastructure for robotics and autonomous systems.
Ropedia
Ropedia is a physical AI infrastructure company that has served more than 1,800 data agencies and develops embodied AI training data tools.
The details
Kosmo handles spatial capture and 3D reconstruction, while Ropedia manages action recording, 4D reconstruction, and semantic annotation using its HOMIE Gen2 and Experience Engine. The two companies state they will refine collection and delivery based on failure cases observed during model training.
Timeline
October 14: Hesai launched Kosmo One and Kosmo World.
November 12: The Ropedia Xperience 2 research project launched.
The Tech Race
The collaboration extends Ropedia's HOMIE Gen2 and Experience Engine pipeline by adding Kosmo's 3D spatial capture as an upstream data source. It positions both companies against data suppliers serving OpenAI, Meta, DeepMind, NVIDIA, Google, and Microsoft, which the partners name as target clients.
For developers building robots and embodied AI systems, the pipeline offers a single source for both spatial geometry and human-object interaction data. Whether it becomes available to smaller teams beyond the named target clients has not been stated.
The takeaway
Robot training depends on data that captures both where objects sit in space and how people interact with them, and this partnership combines the two. Progress in embodied AI will depend on whether such pipelines produce measurable gains in deployed models.
Further reading
For related coverage of machines learning from physical environments, see the Robotics section.







