Training Data for Unitree Robotics

Unitree's business model creates a unique dynamic in the training data market. By pricing the G1 humanoid at roughly $16,000 — an order of magnitude less than competitors — they are democratizing access to humanoid hardware. But this accessibility means their robots will be deployed in far more diverse settings than any single company's internal data collection can cover.

Research labs buying G1 units for university projects need data tailored to their specific research domains. Commercial customers deploying G1 for patrol, inspection, or light logistics need data from their specific operating environments. This creates a long-tail data distribution that Unitree cannot possibly serve from in-house collection alone.

The locomotion challenge is particularly relevant for Unitree. Their reinforcement learning-based locomotion controllers are primarily trained in simulation using Isaac Gym. While these policies transfer reasonably well to flat indoor surfaces, the outdoor terrain traversal that many customers want — grass, gravel, slopes, construction debris — requires real-world data to calibrate the sim-to-real gap. Each terrain type has unique contact dynamics that simulation engines model imprecisely.

For manipulation, the G1's relatively compact form factor means it operates in different workspace geometry than larger humanoids like Atlas or Figure 02. Training data needs to match this specific embodiment's reach envelope, force capabilities, and camera viewpoints. Claru's ability to collect data using standardized protocols but in diverse physical environments addresses both the embodiment-specific and environment-diversity requirements simultaneously.

Unitree's Go2 quadruped has become the most popular research quadruped worldwide due to its combination of capability and price. This massive installed base creates a community data opportunity — hundreds of labs are training Go2 controllers for different tasks, but their data remains siloed. A distributed data collection framework that standardizes recording protocols across this user community could create the most diverse quadruped locomotion dataset ever assembled.

The social navigation dimension is increasingly important as Unitree robots move from research labs into service roles. Security patrol, companion assistance, and delivery applications all require robots to navigate among pedestrians — predicting human movement, maintaining appropriate distances, and responding to social cues. This demands data collected in crowded real-world environments like shopping malls, office buildings, and public sidewalks.