Xiaomi has released the code and weights for Xiaomi-Robotics-1, a 5 billion-parameter vision-language-action model designed to control robots from images and natural-language instructions. Xiaomi Tech (@XiaomiTech_) publicized the open-source package on August 5th, two days after the project repository recorded the release of its post-training, inference and benchmark evaluation pipeline.
The release turns a research announcement from July into an artifact robotics teams can inspect and adapt. Xiaomi published the initial technical report on July 16th, describing a model pre-trained on more than 100,000 hours of Universal Manipulation Interface, or UMI, trajectories. UMI systems let people demonstrate manipulation tasks with handheld gripper-like hardware, making data collection less dependent on operating a specific robot.
That distinction defines Xiaomi's core bet. The 100,000-hour figure represents human-generated manipulation trajectories rather than 100,000 hours of autonomous robot operation. Xiaomi says the corpus spans more than 1,700 scenarios across homes, commercial premises, industrial sites and outdoor environments. An automated vision-language system split the recordings into clips and labeled changes in scenes, objects and gripper movements.
Xiaomi then used more than 10,000 hours of cross-embodiment data to align those learned actions with physical robots and direct instructions. That post-training mixture included more than 7,200 hours of in-house robot data collected in homes, along with filtered public robot datasets and manually annotated UMI trajectories, according to the report.
What Xiaomi released
The 5B checkpoint is available under the Apache 2.0 license. Xiaomi also published code for adapting the checkpoint to new robot data, running an inference server and evaluating versions fine-tuned for RoboCasa, RoboCasa365 and VLABench. Separate benchmark checkpoints are collected on Hugging Face.
The package is aimed at robotics researchers with their own data and hardware integration work ahead of them. Xiaomi's deployment instructions require Python 3.9 or newer, a CUDA GPU and version 4.57.1 of Hugging Face Transformers. The model uses Qwen3-VL-4B-Instruct as its vision-language backbone and accepts three camera images, a language instruction and the robot's current state. Its output contains action commands for components including dual arms and grippers.
Xiaomi provides a demonstration dataset showing the expected post-training format. The main 100,000-hour UMI corpus and the 7,200-plus hours of in-house robot recordings are outside the published package. Outside teams can fine-tune and deploy the checkpoint, though they cannot reconstruct Xiaomi's full training run from the released materials.
That boundary matters because the dataset is the central research asset. The paper says Xiaomi-Robotics-1 improved as the researchers increased both training data and parameter count, with data increases producing more pronounced gains in action-prediction accuracy than model-size increases. Opening the adaptation pipeline gives other labs a way to test the resulting policy, while Xiaomi retains the collection operation that produced its claimed advantage.
Xiaomi's benchmark claims
Xiaomi reports that Xiaomi-Robotics-1 reached a 75% average success rate across phone packing, printer refilling, laundry and box packing after receiving fewer than 10 hours of demonstrations per task on average. A pi-0.5 baseline reached 40% under the same data budget. With fewer than 40 hours per task on average, Xiaomi reported success rates of 85% for Xiaomi-Robotics-1 and 53% for pi-0.5.
Those figures come from Xiaomi's own experiments. The research team evaluated each model for 10 trials per task, leaving a relatively small number of attempts behind each percentage. The released evaluation code and checkpoints should make the simulation results easier to examine, while reproducing the real-robot claims still requires compatible machines, task setups and demonstration data.
Xiaomi also reports average success rates of 74.5% on RoboCasa, 57.4% on RoboCasa365 and 59.1% on VLABench. Xiaomi lists a 13.93% average success rate for RoboDojo in the project table, while the paper separately reports an average RoboDojo score of 20.07. The metrics measure different outputs and should not be treated as interchangeable.
Xiaomi-Robotics-1 follows Xiaomi-Robotics-0, a 4.7 billion-parameter vision-language-action model introduced in February. Xiaomi released the earlier model's full post-training pipeline in April. Xiaomi-Robotics-1 expands that approach around a much larger UMI collection and a two-stage training recipe modeled on the pre-training and alignment structure used for language models.
The open-source package gives robotics teams a usable checkpoint instead of another video-only demonstration. Xiaomi's strategic advantage remains upstream: collecting, labeling and aligning enough physical-world data to make general robot policies useful across new tasks and machines.