In March 2026, AgiBot announced its AgiBot World dataset crossed 1 million real-world robot action trajectories. Every trajectory comes from a physical humanoid robot executing a real task, with joint angles, force readings, and camera frames logged at millisecond resolution. This is not simulation data. For engineers evaluating Chinese humanoid vendors in 2026, the competitive gap between brands is no longer primarily about hardware. It is about data.
How AgiBot Generates Trajectories at Scale
AgiBot’s Shanghai facility runs 100 humanoid robots simultaneously. Human operators demonstrate tasks via haptic teleoperation and the robot mirrors every motion. A simple pick-and-place trajectory contains 3,000 to 8,000 time-step records. A complex assembly trajectory with controlled 2 N insertion force runs to 15,000 records. At 100 robots running 16 hours per day, AgiBot generates 8,000 to 12,000 trajectories daily. Crossing 1 million required approximately 100 to 125 days of continuous operation — this infrastructure has been running since at least mid-2025.
Why Trajectory Count Matters
Manipulation policies train on demonstration data via imitation learning. Policy quality scales with diversity and quantity. Diversity means the same task demonstrated across different lighting, object positions, and surface textures. A robot trained on 10,000 identical demonstrations fails when the object rotates 15 degrees. Trained on 100,000 varied demonstrations, it handles rotation, displacement, and partial occlusion without retraining. Quantity drives the generalization that diversity alone cannot provide. With enough demonstrations, the policy network learns physical priors: what contact forces feel like, how objects respond to gravity, when to apply controlled force. These priors transfer across tasks in ways that cannot be engineered manually.
The Compounding Data Flywheel
More data produces better policies. Better policies enable more reliable robots. More reliable robots deploy in more environments. More environments generate more diverse data. Each iteration widens the gap between the data-rich leader and every follower. AgiBot’s 10,000 cumulative units shipped by March 2026 represent 10,000 data collection endpoints across automotive, electronics, and logistics customer sites. Every deployed robot is also a sensor. Unitree has similar shipment volumes but sells to researchers and developers who may not feed structured training data back to a central dataset. AgiBot’s enterprise model creates a tighter data loop.
What AgiBot World Contains
A public subset of 100,000 trajectories was released in late 2025 across 100 manipulation categories: object sorting, peg-in-hole insertion, cable routing, container opening, garment folding, and multi-step assembly. Each category includes 500 to 2,000 demonstrations with 6-axis force/torque data, RGB-D from three viewpoints, and full joint state logs.
Implications for Procurement
Ask vendors: how many trajectories collected? On what task categories? What fraction from real versus simulation? What is monthly data collection rate? A vendor with fewer than 100,000 real-world trajectories in 2026 still requires extensive on-site fine-tuning per deployment. A vendor with 1 million plus diverse real trajectories can deploy generalizing policies out of the box, significantly reducing integration time and cost. The trajectory dataset is now a procurement due diligence item alongside payload and price.
See Top 40 China Robot Rankings 2025 and robot manufacturers in China.
Sources
- People’s Daily EN: AgiBot ships 10,000 cumulative humanoid units (March 2026)
- AgiBot World dataset release notes (November 2025)
- ESM China: China humanoid robot shipment data 2025
