Riemann Dynamics Partners with Lightwheel AI and Noitom Robotics to Build 1 Million Hours of Embodied Data by 2026

The collaboration revolves around Riemann's Riemann-1.0 embodied world action model and Matrix-Game 3.5 interactive world model. With Lightwheel AI, the two sides will adapt and validate these models against Lightwheel's EgoSuite human data platform, RoboFinals evaluation platform and RoboStack deployment feedback platform. Riemann will use the performance of Riemann-1.0 in tasks such as mobile manipulation, object interaction, home service and long-horizon tasks to identify specific data requirements and capability gaps. Lightwheel's EgoSuite platform will then organize large-scale, high-quality human behavior video and multimodal data production to feed back into the model training pipeline.

With Noitom Robotics, Riemann will cooperate on motion capture, human-robot interaction, multimodal behavior data collection and data engineering. Noitom's long-term experience in human motion capture, force feedback collection and multimodal behavior acquisition will support Riemann-1.0 in long-tail tasks, complex manipulation, fine force control and cross-embodiment generalization. The two companies will jointly collect training data including human motion trajectories, joint states, contact force feedback and interaction object states, with a focus on solving high-precision temporal and spatial synchronization of multimodal data.

Riemann-1.0 is Riemann's embodied world action model for Physical AI. It achieved a 62.6% average success rate on the RoboCasa-365 benchmark, ranking first and surpassing the previous industry-leading level by 8.4 percentage points. Matrix-Game 3.5 is an interactive world model system aimed at building long-term memory, continuous interaction and open-world simulation. Li Yangguang, co-founder of Riemann, said the company wants to let model capability needs directly participate in the data ecosystem and use real deployment feedback to drive model iteration.

Riemann said it will continue to identify data gaps based on real-task performance of Riemann series models, and accelerate the iteration of general embodied intelligence models through multimodal data platforms, evaluation systems and robot deployment feedback. The company also plans to complete the collection and training of one million hours of embodied data by the end of 2026, and build an infrastructure covering data, evaluation, model iteration and real-machine deployment.