Xiaomi claims robotics scaling laws hold with new foundation model

Xiaomi has released Xiaomi-Robotics-1, a foundation model designed to address the persistent data scarcity that has historically constrained the scaling of robotics policy models. By leveraging large-scale embodiment-free pre-training on 100,000 hours of trajectories across more than 1,700 scenarios, the company aims to replicate the empirical scaling laws that have driven advances in language and vision artificial intelligence. The model combines this extensive pre-training with a post-training phase utilising 7,200 hours of real-robot data to align the system with physical embodiments and natural-language instructions.

The pre-training phase addresses the infeasibility of manually annotating such vast datasets by employing an automatic annotation pipeline powered by a vision-language model. This system segments long videos into fixed-length clips and generates language descriptions of scene state transitions, allowing the model to learn action generation that drives scenes toward described states. Xiaomi reports that this approach yields a clean scaling behaviour, where validation action error decreases steadily as both the volume of pre-training data and the model size increase.

Post-training focuses on two alignment axes: embodiment alignment and instruction alignment. Using cross-embodiment datasets that include in-house collected data from tasks such as tidying sofas, sorting shoe cabinets, and putting away kitchenware, the model maps its general action-generation capabilities onto actual robots. This stage shifts the model from generating actions based on state descriptions to executing direct natural-language instructions, enabling it to perform mobile manipulation tasks in real-world environments out-of-the-box.

In evaluations of unseen environments and object instances, Xiaomi found that the scaling behaviours observed during pre-training transfer directly to real-robot performance. The company states that real-robot success rates rise predictably as pre-training data and model size grow, with no signs of saturation in the observed gains. This suggests that a stronger pre-trained model consistently yields better performance in downstream applications, providing a practical pathway for scaling robot foundation models.

The model demonstrates high data efficiency when adapting to new tasks, requiring under 10 hours of demonstration data per task to achieve a 75 per cent overall success rate. This performance nearly doubles the 40 per cent success rate of the π0.5 baseline under the same data budget. When the demonstration budget is increased to under 40 hours, the overall success rate rises to 85 per cent. Additionally, Xiaomi-Robotics-1 achieved state-of-the-art results on four mainstream simulation benchmarks, highlighting its generalisation capabilities.