Kunlun Wanwei spins out robotics unit and trains Riemann-1.0 on 232,000 hours of embodied data

创业邦 · wechat · 2026-07-21

From world models to robots

Kunlun Wanwei’s internal team, now operating as Riemann Dynamics, has shifted from game-world generation and video models toward embodied AI and robotics.

What they built

They introduced Riemann-1.0, a world action model aimed at long-horizon tasks, cross-scenario generalization, and cross-embodiment transfer.
The team argues robots still lack a unified API like ChatGPT, and wants to create a cross-robot interface so different hardware can reuse the same model.
They say the model jointly learns robot actions and the environment changes those actions cause, allowing the robot to observe results, update its state, and revise plans.

Data and scaling

The model is trained on 232,000 hours of multimodal embodied interaction data.
That includes 200,000+ hours of first-person human video, 12,000 hours of high-precision demonstration data, and 20,000 hours of real-robot plus simulation trajectories.
The team says robotics still lacks an internet-scale data source comparable to CommonCrawl or ImageNet, so it is leaning on crowdsourced first-person capture.

Product and business direction

The company is testing domestic commercialization through a robot API for hardware makers who lack strong model teams.
Abroad, it is exploring paid demand in Japan, especially for hotels, elder care, airports, and retail.
The long-term bet is that robots could move from selling hardware to selling services, with a new token-like economy around robotic capabilities.

Original post →

Kunlun Wanwei spins out robotics unit and trains Riemann-1.0 on 232,000 hours of embodied data

More from Companies & People