Xiaomi’s robotics team has published Xiaomi-Robotics-1, a vision-language-action model the company says is ready to run on physical robots without additional training. Xiaomi posted the technical writeup on its robotics research site on July 20, 2026, alongside an arXiv preprint describing the model’s construction. The headline claim is not the model’s task list. It is how little robot data Xiaomi needed to build it.
The model learns in two stages, mirroring how large language models are built. A pretraining phase absorbs a huge volume of general manipulation footage. A much smaller post-training phase then adapts that general knowledge to a specific robot body and to following spoken instructions.
The pretraining corpus totals 100,000 hours of manipulation footage gathered across more than 1,700 environments, spanning houses, shops, factories, and outdoor sites. None of that footage came from a robot. Xiaomi calls the approach embodiment-free, built on UMI rigs, short for universal manipulation interface, handheld devices that record human hand and gripper motion without a robot arm attached. A vision-language model then auto-labels the clips by describing how objects and grippers change state, avoiding the need for humans to hand-annotate 100,000 hours of footage.
Only after that stage does an actual robot enter the pipeline. Xiaomi’s post-training set adds a bit over 7,200 hours of data collected from robots working inside real homes, on tasks such as folding items on a sofa, organizing a shoe rack, and stocking kitchenware, combined with filtered public robot datasets. That works out to roughly one hour of real-robot data for every fourteen hours of human-video pretraining, a ratio that pushes the expensive part of the pipeline, physical robot operation, well behind the cheap part, recorded human motion.
That ratio is the real story here. Physical Intelligence’s pi-0.5 and NVIDIA’s GR00T line have both chased general-purpose robot foundation models, but neither has published a pretraining-to-post-training split skewed this heavily toward embodiment-free capture. If Xiaomi’s scaling pattern holds up outside its own labs, it argues that robotics’ data bottleneck was never robot hardware availability. It was cheap, embodiment-agnostic footage the field had been under-collecting relative to compute and parameters.
Xiaomi reports that pretraining scale predicts downstream competence: as it added pretraining hours and model size, both validation error and real-robot task success improved with no sign of leveling off. On efficient task adaptation, the company says XR-1 reached a 75 percent success rate across four household tasks, phone packing, printer refilling, laundry loading, and box packing, using under 10 hours of demonstration per task, against 40 percent for the pi-0.5 baseline at the same budget. Extending the budget to under 40 hours per task lifted XR-1 to 85 percent against 53 percent for pi-0.5.
On four simulation suites, RoboCasa, RoboCasa365, VLABench, and RoboDojo, Xiaomi reports its model ahead of the next-best system in every case. The largest relative gain, 58.3 percent, came on RoboDojo, where XR-1’s absolute score of 13.93 percent still trails far behind a more saturated benchmark like RoboCasa, where it posted 74.5 percent.
Every comparison in this release is self-reported. Xiaomi ran its own model against pi-0.5 and its own simulation suite; no independent lab, competition, or third-party leaderboard has verified these numbers, and the writeup includes no failure-mode breakdown or safety evaluation. The company also has not disclosed the compute budget behind the 100,000-hour pretraining run, which makes any cost comparison against Physical Intelligence or NVIDIA’s programs impossible to run today.
Teams evaluating robot foundation models for warehouse or home-service pilots should treat XR-1’s reported numbers as a hypothesis to test on their own hardware rather than a settled benchmark result, and should watch for independent replication of the embodiment-free scaling curve before committing a 2026 deployment budget to the approach.
Reported by Xiaomi on July 20, 2026.
