Xiaomi has released details of Xiaomi-Robotics-1, a robot AI model built around a large dataset of manipulation recordings. The Decoder reports that the model improved more from additional data than from simply increasing model size.
Robot AI has a different scaling problem from language models because useful movement data is scarce. Xiaomi addressed that by collecting demonstrations with portable handheld grippers and cameras, rather than relying mainly on physical robots. The company says this produced more than 100,000 hours of motion recordings from varied environments.
The dataset then had to be labeled. Xiaomi used another AI model to describe motion segments in text, reportedly finishing the labeling process in about two weeks. The resulting model was transferred to wheeled and dual-arm robots.
In tests, success in unfamiliar environments rose from about 25 percent to 75 percent as training data increased. Xiaomi says the model posted strong benchmark results, but the broader lesson is more cautious: robotics progress may depend less on ever-larger models than on collecting varied, transferable real-world data.