The most consequential competition in Chinese robotics in 2025-2026 is not which humanoid has the most degrees of freedom or the highest joint torque. It is which company accumulates the most and best-quality training data for robot AI. Embodied AI, AI that controls physical bodies in the physical world, is fundamentally a data problem: the models that will make humanoid robots genuinely useful are trained on millions of demonstration trajectories, sensor readings, and task executions. The company with the best data will build the best AI, which will make the best robots, which will generate more data. Several Chinese companies are racing to 10 million data entries, the threshold that industry observers estimate is necessary for meaningful task generalization. This article covers who is in that race, what the data consists of, and why 10 million entries matters.
What Embodied AI Training Data Is
A robot AI training entry is a recorded episode of robot operation: sensor inputs (camera images, joint positions, force/torque readings, IMU data) time-synchronized with robot actions (joint position commands, gripper commands) and outcomes (task success, failure, collision). One entry might be a 30-second pick-and-place task: the robot reaches for an object, grasps it, moves to a target location, and places it, approximately 15,000 data points at 500 Hz recording rate. Scaling to 10 million entries requires generating this data across diverse objects (thousands of object types), diverse environments (different lighting, backgrounds, occlusion patterns), and diverse robot starting states (different initial joint configurations). The diversity is what drives generalization, a model trained on 10,000 variations of the same task generalizes far better than one trained 10,000 times on the exact same conditions.
The Race: Who Has What
AgiBot: 10 million synthetic entries from Genie Sim 3.0 (2025), plus approximately 500,000 real-world entries from deployed units. AgiBot is ahead on synthetic data scale. Zhiyuan Robotics: announced a 10 million entry target for its embodied AI model, training underway as of mid-2025, mix of real and synthetic. Zhiyuan’s differentiation is physical data from its research deployment partnerships. Unitree: 50,000 real-world trajectories for UnifoLM-VLA-0 (disclosed), likely more proprietary data not disclosed. Unitree leads in real-world data quality for manipulation tasks covered by UnifoLM-VLA-0. Fourier: N1 open-source community generates additional data from 500+ build teams, distributed data collection at low cost, unclear quality and curation. UBTECH: BrainNet training data from Walker S2 deployments in Siemens partnership environments, industrial task data at production scale, proprietary.
Synthetic vs Real Data
Synthetic data (from simulation) scales cheaply and covers scenarios that are rare or dangerous in real operation. The limitation is sim-to-real gap: physics simulators imperfectly model contact dynamics (the way objects slip, deform, and stick during manipulation), which means policies trained purely on synthetic data fail on contact-rich tasks in the real world. Real data is expensive to collect and scales slowly, but captures true contact dynamics. The hybrid approach, synthetic data for coverage and scale, real data for contact dynamics fine-tuning, is the current consensus in Chinese robotics research. AgiBot’s 10M synthetic + 500K real ratio is consistent with this hybrid approach.
Data Quality vs Quantity
10 million low-quality entries (robot teleoperation with random errors, inconsistent task definitions, poor camera calibration) is less valuable than 1 million high-quality entries (expert demonstration, verified success labels, calibrated sensors, diverse but controlled environments). The data quality competition is not tracked by public announcements, companies do not disclose their quality filtering ratios. Industry observers estimate that 30-50% of raw collected entries are discarded during quality filtering. If AgiBot discards 40% of its synthetic data and 20% of its real data, its effective training dataset is 6M synthetic + 400K real, still the largest disclosed in China, but with significant uncertainty about quality.
Open Data as Strategy
AgiBot released 1M synthetic entries as open source. Fourier released N1 build data. These releases accelerate the broader Chinese embodied AI sector, more researchers building on Chinese robot platforms, more papers citing Chinese data, more international adoption of Chinese robot AI standards. The strategic calculation: the open-released data is not proprietary competitive data. It is the commodity data that helps build the overall Chinese embodied AI sector, which creates demand for Chinese robot hardware. Proprietary data, the industrial task data from production deployments at CATL, automotive plants, and electronics manufacturers, is retained.
AgiBot data context: AgiBot Genie Sim 3.0. Unitree AI context: Unitree UnifoLM-VLA-0. Full ranking: Top 40 China Robot Rankings 2025.
Sources
- AgiBot: Genie Sim 3.0 data scale disclosure (2025)
- Zhiyuan Robotics: 10M data entry target announcement (2025)
- Unitree: UnifoLM-VLA-0 training data disclosure (March 2026)
- CRIA: Embodied AI data sector report (2025)
