Three Chinese companies join forces to build the world's first 1 million hours of embodied data

Riemann Dynamics, which recently achieved dual SOTA in simulation and real-world testing, has partnered with Noitom Robotics and Lightwheel AI to build an embodied intelligence data foundation. The three companies will integrate collection, simulation, training, and evaluation into a unified pipeline, with Riemann Dynamics aiming to collect 1 million hours of robot training data by end of 2026, targeting double the current global estimate of high-quality embodied interaction data (about 500,000 hours). Riemann Dynamics has already trained its Riemann-1.0 model on 232,000 hours of data, validating the approach. The collaboration addresses the industry's data bottleneck: insufficient quantity and strict quality requirements, as embodied data comes from varied sources like real robot collection, teleoperation, wearables, and simulation, each with limitations. Noitom Robotics, a leading motion capture company with 70% global market share, provides real human action data; Lightwheel AI, the first embodied data unicorn, offers simulation and evaluation platforms; Riemann Dynamics, spun from Kunlun Wanwei's Matrix world model team, handles model training. The partnership aims to create an open, standardized data ecosystem, with shared standards and open-source data, similar to NVIDIA's Physical AI Data Factory and Waymo's closed-loop approach in autonomous driving.