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

The scaling of embodied intelligence has yet to materialize, and the industry consensus points to data as the issue. Robot training data cannot be sourced from the internet like large language models; it typically remains in the thousands to tens of thousands of hours, with only a handful of models reaching hundreds of thousands of hours.

Physical Intelligence used over 10,000 hours of data to train π0, NVIDIA's GR00T 1.7 publicly disclosed about 40,000 hours, and Google DeepMind, in collaboration with over 20 institutions, managed to gather just over 1 million trajectories. These scales are merely a starting point in the world of large models.

Not only is data quantity insufficient, but quality requirements are also higher. Real-robot data, human videos, and simulation data each have their shortcomings, and they are scattered across different players, leading to a disconnect between collection, training, and evaluation, making it difficult to form a closed loop.

Riemann Dynamics, Lightwheel AI, and Noitom Robotics have announced a joint effort to build an embodied intelligence data foundation, connecting collection, simulation, training, and evaluation into a single pipeline. Riemann Dynamics is responsible for model training, Noitom provides real action data, and Lightwheel AI offers simulation and evaluation.

Riemann Dynamics was formerly the Matrix world model team within Kunlun Wanwei, which developed the Matrix-Game series, with the latest version supporting 720p long-sequence interaction. The team discovered that game world models and robot World Action Models address the same problem: how actions change the world.

Riemann-1.0 extends Matrix's modeling capabilities by incorporating robot actions into a unified framework, enabling planning and action in open worlds. However, moving from 232,000 hours to 1 million hours requires industry chain collaboration, not just algorithmic improvements.

Riemann Dynamics aims to complete 1 million hours of data collection by the end of 2026, striving to become the first company globally to surpass this scale. According to a report by iyiou, current global high-quality embodied interaction data is only about 500,000 hours, making the target twice the existing stock.

Fang Han, Chairman of Kunlun Wanwei, stated that the differentiation in embodied intelligence model capabilities lies in who can first collect training data at the million-hour, ten-million-hour, or even hundred-million-hour level. A million hours is not the end but the beginning.

This goal is more like an industry experiment: to verify whether robot capabilities can scale along with data once scaling truly begins.

In the early days of autonomous driving, the focus was on mileage, but it later became clear that the real differentiator was how quickly long-tail scenarios could be converted into training data. Waymo, through its internal closed loop, has accumulated over 200 million miles of real-world driving experience and billions of miles in simulation.

The robotics industry cannot simply copy Waymo's model because robots have not yet entered homes and factories at scale, lacking a natural data entry point, and the cost of failure is high. Therefore, building a closed loop through collaboration becomes a more practical choice, as NVIDIA has done by partnering with Google DeepMind and others to build the Physical AI Data Factory.

The collaboration among Riemann Dynamics and the other two parties follows the same logic, and they will jointly establish data standards, annotation standards, and evaluation benchmarks, with some data open-sourced to accelerate industry-wide progress.