Xiaomi-Robotics-1 (robotics.xiaomi.com)

Xiaomi has announced the development of Xiaomi-Robotics-1, a groundbreaking robot foundation model that utilizes over 100,000 hours of embodiment-free (UMI) pre-training data combined with real-robot data in a post-training phase. This innovative approach addresses the scarcity of large-scale, high-quality datasets that has historically limited the scalability of robotic policy models. By leveraging a scalable auto-labeling pipeline, the model learns action generation from diverse tasks across various environments without the need for extensive manual annotation. As the amount of pre-training data and model size increases, the model demonstrates consistent improvements in real-robot performance, showing clear scaling behavior in terms of success rates for tasks such as phone packing and laundry loading. This development is significant for the AI/ML community as it paves the way for more effective robot applications by showcasing the potential of large-scale training methodologies applied to robotics. Xiaomi-Robotics-1 can achieve impressive task success rates, with a 75% rate after just a few hours of training on new tasks, significantly outperforming previous models. Additionally, it has achieved state-of-the-art results on multiple simulation benchmarks, indicating that its scaling and generalization benefits extend beyond theoretical constructs into practical, real-world robotics applications. This model’s ability to efficiently transfer learned behaviors into physical robots marks a pivotal step for advanced automation.