Behind Hundreds of Robot Training Grounds, Embodied AI Recalculates the Ledger

A year ago, one of the most common questions in embodied AI was: where will robot data come from? Today, the question has quietly shifted to: is the data collected at such high cost actually useful? That change marks a new phase in which the industry is moving from data scarcity to scrutinizing data returns.

In March 2025, Beijing's first humanoid robot data training center was unveiled at Shougang Park, according to media reports. The planned 3,000-square-meter facility is expected to house 108 humanoid robots for collecting and training data needed by embodied AI.

As more training grounds are built, data supply is no longer the sole bottleneck. Evaluating data quality, lowering collection costs, and improving model generalization are becoming more practical challenges. For embodied AI companies, the return on data investment will shape the next stage of competition. Whether these facilities can deliver useful data at scale remains the key test.