Xiaomi-Robotics-1: Scaling Robot Foundation Models via Embodiment-Free Pre-training

Xiaomi-Robotics-1 breaks the robot data bottleneck using embodiment-free pre-training

Xiaomi-Robotics-1 is a robot foundation model that scales performance by utilizing large-scale, embodiment-free (UMI) data to overcome the scarcity of high-quality real-robot trajectories. By combining 100,000 hours of pre-training data with a targeted post-training alignment phase, the model demonstrates a predictable scaling law: as pre-training data and model size increase, the success rate for real-world robot tasks rises steadily.

Two-Stage Training Paradigm: Pre-training and Post-training

Xiaomi-Robotics-1 follows a training architecture similar to Large Language Models (LLMs), splitting the learning process into breadth-focused pre-training and alignment-focused post-training.

Stage 1: Embodiment-Free Pre-training

To avoid the limitations of collecting data from a single robot type, Xiaomi used 100,000 hours of embodiment-free (UMI) trajectories across 1,700 scenarios, including industrial, commercial, and household spaces.

Because manual labeling at this scale is impossible, Xiaomi developed an auto-labeling pipeline using a Vision-Language Model (VLM). The VLM divides long videos into fixed-length segments and annotates them with language descriptions of scene state transitions (e.g., how a gripper interacts with an object). This allows the model to learn general action generation based on desired state changes rather than specific robot hardware.

Stage 2: Post-training Alignment

Post-training transforms general action-generation capabilities into usable robot skills through two primary axes:

Embodiment Alignment: Using cross-embodiment real-robot data (including 7,200 hours of in-house data from real homes), the model maps general actions to the specific physical constraints of the Xiaomi robot.
Instruction Alignment: The model is shifted from following "state transition descriptions" to understanding and executing natural-language instructions.

Performance and Scaling Results

Evaluation of Xiaomi-Robotics-1 shows that scaling gains from the pre-training phase transfer directly to physical performance. A stronger pre-trained model consistently yields a higher success rate in unseen environments with unseen objects.

Efficient Task Adaptation

Xiaomi-Robotics-1 can adapt to complex new tasks with very high data efficiency. When compared to the $\pi$ 0.5 baseline, XR-1 significantly outperforms it with minimal demonstrations:

Task	<10 h/task (XR-1)	<10 h/task ($\pi$ 0.5)	<40 h/task (XR-1)	<40 h/task ($\pi$ 0.5)
Phone Packing	70%	30%	80%	40%
Printer Refilling	70%	20%	60%	20%
Laundry Loading	80%	40%	100%	50%
Box Packing	80%	70%	100%	100%
Overall	75%	40%	85%	53%

Simulation Benchmarks

XR-1 achieved state-of-the-art (SOTA) results across four major simulation benchmarks, demonstrating strong generalization capabilities:

RoboCasa: 74.5% success rate (+2.6% gain over 2nd best)
RoboCasa365: 57.4% success rate (+23.2% gain over 2nd best)
VLABench: 59.1% success rate (+11.1% gain over 2nd best)
RoboDojo: 13.93% success rate (+58.3% gain over 2nd best)

Technical Analysis and Community Insights

Technical observers note that the tasks demonstrated—such as luggage packing and laundry loading—involve several "classically hard" robotics problems, including the coordination of two hands, handling deformable objects (clothing), and interacting with thin affordances like bag zippers.

Community discussion highlighted several key perspectives on the model's implications:

"The paradigm that scales to bringing truly useful robots to virtually all domains of life [is] just as LLMs have scaled to that degree across virtually all domains of knowledge work."

While some users expressed skepticism regarding the "uncut" nature of promotional footage, others pointed out the efficiency of the model size, with some noting it appears to be around 10B parameters, which is relatively small compared to modern LLMs.

There are currently questions regarding the availability of the model and datasets, as some users reported that the official GitHub and Hugging Face repositories contain assets for "XiaomiRobotics-0" but not yet for the "Xiaomi-Robotics-1" model.