Riemann-1.0 Learns Household Tasks from 200,000 Hours of Human Video

A new trend is emerging in methods to teach robots household tasks. According to a report by Quantum Position, Riemann Dynamics has officially released its general-purpose robot operation model, Riemann-1.0. The model was trained on a dataset totaling 232,000 hours, including over 200,000 hours of first-person human videos. On the robotic household benchmark “RoboCasa-365,” Riemann-1.0 achieved a success rate of 62.6%, surpassing the previous state-of-the-art (SOTA) by 8.4 points and claiming the top spot in rankings.

Riemann Dynamics is a subsidiary established by Kunlun to focus on embodied intelligence, and this marks its first model release. Since its official debut at WAIC 2026, the model’s performance has drawn significant attention.

How Human Videos Transform Robotics

The standout feature of Riemann-1.0 lies in its heavy reliance on human videos as training data. The developers collected 232,000 hours of video footage, primarily first-person perspectives of humans cooking, folding clothes, and cleaning tables. This was supplemented with over 12,000 hours of high-quality data captured via UMI (Universal Manipulation Interface) and exoskeleton gloves, as well as more than 20,000 hours of trajectory data from real and simulated robots. Altogether, the dataset covers 41 types of robot hardware and thousands of interaction methods.

This approach is underpinned by the integration of two methodologies: VLA (Vision-Language-Action) and world models. VLA excels at direct action outputs but is prone to errors, whereas world models are adept at predictions but are slower and more resource-intensive. A review published by NVIDIA’s Seattle Robotics Lab in June highlighted that the World Action Model (WAM) has rapidly evolved from a niche area of VLA research. It suggested that next-generation foundational robotic models would likely be hybrids of these approaches. Similarly, at ICML 2026, multiple papers, including those from Yann LeCun’s team, explored the direction of learning latent action-world models from unlabeled, in-the-wild videos.

Riemann-1.0 is the first domestically developed model to fully implement this hybrid approach as an engineering solution, with validations extending from simulation benchmarks to real-world robots.

Transforming 232,000 Hours of Video into Skills

Human videos do not inherently contain labels for robotic movements. For example, a video of a person folding clothes does not include data about the angles of robotic arm joints or the gripping force of grippers. To tackle this challenge, Riemann-1.0 employs a three-stage data processing pipeline.

The first stage, referred to as “preparation,” involves a proprietary automated human data processing pipeline. This pipeline performs tasks such as correcting fisheye lens distortion, segmenting and labeling video actions using VLM, and conducting quality checks to filter out unsuitable data. Six factors are eliminated: scenes with too many people, meaningless movements, excessive obstructions, incomplete tasks, non-first-person views, and hands moving out of the frame. Additionally, the system constructs 3D hand models and converts motion trajectories into global coordinates, making the “hand movements” in the videos machine-readable as 3D coordinates.

In the second stage, the system adopts a fully causal motion-video integration modeling framework based on diffusion models. This enables the unified learning of visual dynamics, environmental states, and robot action sequences within a single generative process. The model not only recognizes “what is in front of it” but also learns “how actions change the environment and what the world will look like next.”

Finally, in the third stage, the system performs motion alignment using a small amount of real robot data, converting the physical intuition derived from human videos into executable control signals.

Real-World Robot Performance

RoboCasa-365 is a comprehensive test for robotic household tasks, where many models achieve success rates below 50%. Riemann-1.0 scored 62.6%. In demo videos, the robot is shown performing a series of actions: picking up a spoon stuck to a table, emptying leftover soup from a bowl before cleaning it, and finally wiping the table. The model’s ability to close the gap between benchmark scores and actual performance has been highly praised.

Similar approaches have been attempted before. For instance, Being-H0.7 utilized 200,000 hours of first-person video, while Generalist AI’s GEN-1 incorporated 500,000 hours of wearable device data. What sets Riemann-1.0 apart is its ability to successfully implement this paradigm at an engineering level and validate it in both benchmarks and real-world applications.

Editorial Opinion

The launch of Riemann-1.0 provides a potential solution to the “trade-off between data quality and quantity” in the field of embodied intelligence. The successful demonstration of a pipeline that transforms large-scale, low-cost human daily activity videos into robot control signals without action labels could influence the industry’s scaling strategies. In the next 3–6 months, it is highly likely that competitors will introduce similar models adopting this approach.

In the medium to long term, if the method of transferring human behavior data to robots becomes more widespread, the versatility of robots could improve dramatically. However, the ethical acquisition of data and privacy protection—particularly concerning first-person footage collected via wearable devices—are areas where legal frameworks have yet to catch up. This is an issue that demands urgent discussion within the industry.

From our perspective, it remains to be seen whether a hybrid of VLA and world models will become the standard architecture for foundational robotic models or if a new approach will emerge. The degree to which Riemann-1.0’s methods can generalize across different robotic hardware will be a key indicator of its practical utility.

References

• “看了20万小时「人类干活实录」，机器人悟了”, by 一水 — 量子位, 2026-07-19T08:32:36.000Z (ARR)
• Source URL: https://www.qbitai.com/2026/07/454592.html