In March 2026, AgiBot announced that its AgiBot World dataset had crossed 1 million real-world robot action trajectories. That number sounds abstract until you understand what it takes to generate it — and what it enables. This is not synthetic simulation data. Every trajectory comes from a physical robot executing a physical task in a real environment, with every joint angle, force reading, and camera frame logged at millisecond resolution.
Understanding this dataset is critical for anyone evaluating Chinese humanoid robot vendors in 2025-2026. The 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 collecting data. Each robot performs manipulation tasks — picking, placing, inserting, assembling — while human operators provide teleoperated demonstrations through a haptic control interface. The operator moves; the robot mirrors; the system logs every action.
One trajectory covers one complete task execution: from initial arm position through grasp through placement through release. A trajectory for a simple pick-and-place operation might contain 3,000 to 8,000 time-step records depending on speed. A complex assembly trajectory — inserting a connector into a PCB socket while applying controlled 2N insertion force — might contain 15,000 records.
At 100 robots running 16 hours per day, AgiBot generates approximately 8,000 to 12,000 trajectories per day. Reaching 1 million trajectories required roughly 100 to 125 days of continuous operation at this scale. The implication: AgiBot has been running this data collection infrastructure since at least mid-2025.
Why Trajectory Count Matters for Robot Intelligence
Modern robot manipulation policies are trained on demonstration data using imitation learning — the robot learns to reproduce what the human operator demonstrated. The quality of the learned policy scales with two variables: the diversity of the training data and the quantity.
Diversity means the same task demonstrated across different lighting conditions, object positions, surface textures, and starting configurations. A robot trained on 10,000 trajectories of the same grasp task executed identically will fail when the object is rotated 15 degrees. A robot trained on 100,000 varied demonstrations of the same task handles rotation, displacement, and partial occlusion without retraining.
Quantity drives the generalization that diversity alone cannot provide. With enough demonstrations, the policy network begins to learn physical priors — what contact forces feel like, how objects respond to gravity, when to slow down and apply more controlled force. These priors transfer across tasks in ways that cannot be engineered manually.
Tesla’s Optimus team has discussed targeting 1 billion simulation steps for training. AgiBot’s 1 million real-world trajectories are not directly comparable — real data and simulation data have fundamentally different noise characteristics and physical accuracy — but the scale signal is the same. Both companies are betting that data quantity is the primary lever for making robots genuinely useful in unstructured environments.
The Compounding Advantage
The data flywheel works like this: more data produces better policies. Better policies mean robots can execute tasks more reliably. More reliable robots can be deployed in more environments. More deployment environments generate more diverse training data. That data improves the next model generation. The cycle repeats, and each iteration widens the gap between the data-rich leader and every follower.
AgiBot’s 10,000 cumulative humanoid units shipped by March 2026 — confirmed in People’s Daily reporting — represent 10,000 data collection endpoints, each contributing trajectories from real customer sites across automotive, electronics, and logistics. This is the critical advantage of shipping volume early: every deployed robot is also a data sensor.
Unitree, with similar shipment volumes, faces the same opportunity. But Unitree’s G1 and H1 are purchased by a much more diverse customer base — researchers, developers, hobbyists — who may not contribute structured training data back to Unitree’s central dataset. AgiBot’s closer integration with enterprise customers through deployment partnerships creates a more controlled data collection loop.
What AgiBot World Contains
AgiBot released a subset of the AgiBot World dataset publicly in late 2025, covering approximately 100,000 trajectories across 100 manipulation task categories. The categories include: object sorting, peg-in-hole insertion, cable routing, container opening, drawer manipulation, garment folding, and multi-step assembly sequences. Each category contains 500 to 2,000 demonstrations with full 6-axis force/torque data, RGB-D camera streams from three viewpoints, and joint state logs.
The public release is both a research contribution and a marketing move. Academic labs that build on AgiBot World data for their research will publish results that validate AgiBot’s platform. Engineers at competitor companies will study the dataset structure and task coverage as a benchmark for their own programs.
Implications for Procurement and Competition
For industrial buyers evaluating humanoid platforms, the trajectory dataset is a due diligence question. Ask vendors: how many trajectories have you collected? On what task categories? What fraction comes from real environments versus simulation? What is your data collection rate per month?
A vendor with fewer than 100,000 real-world trajectories in 2026 is still in early-stage data development. Their robots will require extensive on-site fine-tuning for each new deployment environment. A vendor with 1 million+ diverse real-world trajectories can deploy policies that generalize out of the box for standard task categories — significantly reducing integration time and cost.
For AgiBot’s competitors, the 1 million trajectory milestone sets a concrete bar. Unitree, UBTECH, Leju, and Fourier Intelligence all need to answer the question of how their data programs scale. Hardware capability across the top Chinese humanoid brands is increasingly comparable. The differentiating factor for the next 24 months will be data scale and policy quality.
For more on the competitive dynamics of China’s humanoid robot leaders, see our Top 40 China Robot Rankings 2025 and our overview of 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 H1-H2 2025
