Xiaomi releases robotics AI model Xiaomi-Robotics-1, more data beats bigger models

Xiaomi released its robotics AI model Xiaomi-Robotics-1, which follows scaling laws similar to large language models but gains performance mainly from data volume. To build the dataset, Xiaomi used portable hand-held grippers with cameras operated by humans, collecting over 100,000 hours of motion data in kitchens, offices, and factories. Another AI model auto-annotated the data in about two weeks. Tests show the performance gain from increasing data far exceeds that from scaling model parameters: in unfamiliar environments, success rates rose from 25% to 75% without hitting a plateau. On standard robotics AI benchmarks, Xiaomi-Robotics-1 achieves state-of-the-art results. On four new tasks (e.g., packing a phone, loading a washing machine), it hit 75% average success rate with less than 10 hours of training data, dwarfing Physical Intelligence's 40%. The model accepts voice or text commands and can be adapted to new tasks with minimal training. Xiaomi plans to open-source the model and code on GitHub and Hugging Face.