Xiaomi Robotics 1

Xiaomi-Robotics-1 (XR-1) is a Vision-Language-Action (VLA) model for robot manipulation. It couples a pre-trained Qwen3-VL vision-language model with a Diffusion Transformer action head through a Mixture-of-Transformers design. The DiT matches the VLM in layer count but uses a smaller hidden size for faster inference. The released checkpoint is a 5B-parameter model.

Training has two stages. Pre-training uses over 100K hours of embodiment-free UMI (handheld gripper) trajectories across 1,700+ scenarios. A VLM-based pipeline auto-labels clips with descriptions of state transitions, which avoids manual annotation. Post-training then uses over 10K hours of cross-embodiment data, including in-house robot data, filtered open robot data and manually annotated UMI data. This aligns the model to specific robot embodiments and to natural-language instructions.

Xiaomi reports that XR-1 can do mobile manipulation in unseen environments with unseen objects, and can adapt to new tasks from a few hours of demonstrations. It reports results on RoboCasa (74.5%), RoboCasa365 (57.4%), VLABench (59.1%) and RoboDojo (13.93%), and ranked first on the RoboCasa365 and RoboDojo leaderboards as of 15 July 2026. These figures are self-reported. The code is Apache-2.0 and covers post-training, inference and benchmark evaluation. It is built on Hugging Face transformers pinned to 4.57.1 and is aimed at consumer GPUs. It is intended for robotics researchers and developers.