黎曼动力携手光轮智能与诺亦腾机器人，剑指2026年百万小时具身智能数据建设

TL;DR - Riemann Dynamics announced strategic partnerships with Lightwheel Intelligence and Noitom Robotics (Aug 6) to build a closed-loop embodied-AI data pipeline, targeting 1 million hours of embodied data collected and trained by end-2026. It matters because it shifts embodied AI from single-model gains toward model-driven data production, benchmarking, and real-robot deployment feedback.

• Two flagship models anchor the effort: Riemann-1.0, an embodied world-action model trained on large-scale human video, and Matrix-Game 3.5, an interactive world model aimed at long-term memory, continuous interaction, and open-world simulation.
• Riemann-1.0 is claimed to rank first on the RoboCasa-365 household benchmark with a 62.6% average success rate, reported as 8.4 percentage points above the prior leading level.
• With Lightwheel, Riemann-1.0 and Matrix-Game 3.5 will be adapted/validated against the EgoSuite human-data platform, RoboFinals evaluation platform, and RoboStack deployment-feedback platform, so model capability gaps drive targeted data production rather than raw data scaling.
• With Noitom, the focus is motion capture plus force/tactile feedback — body motion trajectories, joint states, contact forces, and interacting-object states — with explicit emphasis on high-precision temporal and spatial synchronization of multimodal streams to improve fine force control, long-tail tasks, and cross-embodiment generalization.