Lei Jun Announces Xiaomi Robot Progress: CyberOne 02 Works Continuously for 8 Hours at Car Factory

Your Xiaomi car is being manufactured by Xiaomi robots:

Just now, Xiaomi founder Lei Jun personally released the latest progress:

Xiaomi's humanoid robot has, for the first time, taken on the challenge of 8 hours of stable operation in the car factory, fully meeting the demands of all-day production.

For Xiaomi, this counts as a highlight moment of crossing the embodied intelligence kill line for the first time.

On the other hand, it is also rather ironic: among the big companies, the first to implement the "8-hour workday" is actually a robot.

Xiaomi robot, "8-hour workday"

If you don't say it, real scenarios really are the best baseline~

In the video of the Xiaomi robot moving bricks at the factory in July, you could still see "flaws," such as shaking and getting stuck.

But in the latest footage, CyberOne 02 can already continuously complete tasks in the car assembly workshop for 8 hours:

The workstation is the sorting of the center console side cover panels in the automotive final assembly logistics area, and the cover panels are flexible workpieces, large in size and irregular in shape, with randomness in their placement posture, placing higher demands on visual positioning and grasping force control than standard structural parts.

Within 8 hours, CyberOne 02 continuously operates in step with the production line tempo, automatically adjusting its own material-picking route and stride in coordination with the incoming AMR:

When the AMR changes shifts, CyberOne 02 also waits patiently, and when there is a deviation in incoming materials or a placement failure, it can adjust on its own:

Lei Jun is very modest, defining this stage as the robot's "internship," a preliminary verification of production-line-level operational capability.

In addition to the factory, he also mentioned another scenario:

In the Xiaomi youth apartment, it has completed validation of long-horizon tasks lasting more than 30 minutes continuously, covering 3 rooms and including 50+ tasks.

It can autonomously plan and execute complete tasks covering tool preparation, sofa and desktop tidying, garbage cleanup, item classification and return, room inspection, whole-house re-inspection, and exit.

Blooming in both B-end and C-end scenarios, Xiaomi's embodied intelligence compared with the Iron Egg of back then gives a sense of "a scholar should be seen with new eyes after three days apart."

And Lei Jun also emphasized that no matter which scenario, behind it is the same embodied intelligence foundation model.

How far has Xiaomi's embodied intelligence technology progressed?

The CyberOne 02 robot "moving bricks" at the Xiaomi car factory is officially stated to be backed by Xiaomi's latest embodied foundation model, Xiaomi-Robotics-1.

And this is a typical piece of research that "starts with a problem and looks for a solution."

Embodied intelligence has long faced a real problem: it is difficult for robots to easily obtain enough training data the way language models do: to learn a certain action or a certain task, a real robot needs to repeatedly execute it in a physical environment, and then collect data through teleoperation and other methods;

But for a robot to repeatedly execute a specific task in a physical environment, it requires the robot to be able to execute this task... To this day, embodied intelligence's data loop still has no recognized efficient kick start path.

You might say, isn't human operation the best data source? To be honest, autonomous driving has already stepped in this pit once, and embodied intelligence says it will not be fooled again.

Not to mention the cost issue, even if there were a data collection product that everyone urgently needed and wore every day, how would it guarantee the proportion of efficient data in the returned data that is key to solving corner cases?

If people were specially hired to do targeted data collection, then it would return to the cost problem. So human data can be used to strengthen and assist key cases, but it is difficult to use as the main data source for improving and iterating the foundation model.

What Xiaomi XR-1 attempts to solve is precisely the question of how embodied intelligence can obtain large-scale pre-training data; translated, it is attempting to replicate the Scaling law that has already appeared in the large model field.

Specifically, in terms of model architecture, XR-1 adopts an end-to-end VLA (Vision-Language-Action) architecture and introduces a Mixture-of-Transformers (MoT) design.

To put it simply, VLA needs to connect three links: the robot obtains environmental information through vision, combines it with language instructions to understand the task goal, and finally generates executable actions.

XR-1 splits this process into two coupled core components: a vision-language backbone and an action generation head.

The vision-language backbone uses the pre-trained Qwen3-VL and is mainly responsible for understanding the scene the robot currently sees and the language instructions given by the user. It provides visual and semantic information to help the model judge the current environmental state and task requirements.

The action generation part uses a Diffusion Transformer (DiT). Conditioned on the robot's own state and the KV cache provided by the vision-language model, it generates the action sequence for a future period through Flow Matching, that is, an Action Chunk.

The division of labor in the architecture is relatively clear: Qwen3-VL is responsible for understanding the environment and the task, and DiT is responsible for converting the understanding results into continuous actions. The two work together through the MoT architecture to form an end-to-end VLA model.

With the model in place, the training stage of XR-1 adopts a two-stage paradigm. The first stage is Embodiment-free Pre-training. Xiaomi uses UMI devices to collect more than 100,000 hours of real human operation trajectories.

UMI can be understood as a low-cost handheld operation data collection device, composed of components such as a gripper and a wrist camera. The Xiaomi team used this setup to complete operations such as grasping, moving, and placing in real environments, while the system synchronously recorded visual information and operation trajectories—collecting data does not require binding to a specific robot; the core is recording the relationships among the different elements of task execution.

After completing pre-training, XR-1 enters the second stage: Cross-embodiment Post-training: through these data, the model gradually adapts to the specific robot's body structure, motion capabilities, and control methods, while also learning the task instructions that humans are more accustomed to using.

This is the true core point of Xiaomi XR-1—adaptation, giving these embodiment-free action intentions a physical interface that all robots can share.

XR-1's choice is to unify action labels into delta EE pose (relative incremental end-effector pose), meaning it only describes how much the end effector translates along which direction and how much it rotates around which axis.

The same type of operation intention corresponds to different joint trajectories on different robots, but their delta EE poses are similar.

The model learns on this shared representation, so post-training does not need to maintain a separate action vocabulary for each robot.

However, delta EE pose only unifies the writing format of actions and cannot solve kinematic feasibility. Post-training can use real robot data to expose these infeasible mappings and let the model learn to avoid them.

This is the second layer of embodiment alignment: the shared representation provides a unified coordinate framework, and real-machine data constrains the instructions within this framework to the subspace that the specific robot can actually reach.

Adaptation at the network level uses hierarchical fine-tuning.

The post-training data is about 10,000 hours, of which Xiaomi's self-collected real-machine data exceeds 7,200 hours. This scale naturally cannot support training from scratch, but its role is calibration: the last few layers of the vision encoder undergo adaptive fine-tuning, shifting features from being oriented toward semantic categories to being oriented toward physical interaction, paying more attention to graspable parts and spatial relationships;

at the same time, the action generation head DiT adjusts the output distribution on cross-embodiment data, making the generated delta EE pose sequences closer to the reachable range of the target embodiment.

As for the main body of knowledge in the Qwen3-VL backbone, it is chosen to be retained, with only interface-level adjustments.

So the actual meaning of "gradual adaptation" is: use delta EE pose to establish a cross-embodiment action writing layer, use real-machine data to constrain this layer to each robot's reachable space, use hierarchical fine-tuning to retain pre-training knowledge while correcting interface-level distribution shifts, and then use imperative-sentence instructions to switch the trigger mechanism to human language.

Judging from the experimental results disclosed in the paper, XR-1 has already demonstrated a certain generalization capability.

In simulation tests, XR-1 achieved a success rate of 57.4%–57.6% on RoboCasa365, higher than the previous best result of 46.6%; on RoboDojo, the average score reached 20.07, also a clear improvement over the previous result of 13.07.

In few-shot fine-tuning tests, XR-1 used on average less than 10 hours of demonstration data per task, and the average success rate on four complex tasks reached 75%, while the comparison baseline π0.5 given in the paper was 40%.

In addition, the paper also demonstrated room-level long-horizon operation tasks lasting more than 10 minutes, such as autonomously completing suitcase packing. Such tasks involve multiple consecutive operation steps and place higher demands on the model's task understanding, action planning, and execution coherence—

Doesn't this echo the validation deployment in Xiaomi Youth Apartments that Mr. Lei mentioned?

To summarize finally, the most valuable explorations of Xiaomi Embodied AI are reflected in:

pragmatic choices in the data paradigm, and verification of the transferability of Scaling behavior.

If this trend continues to hold at larger scales, it means the robotics field can also follow the engineering paradigm of "pre-trained foundation → downstream adaptation."

Xiaomi initially crosses the embodied intelligence "kill line"

English text is being prepared.

For example, the 100,000 hours of UMI data still actually has a gap from real deployment.

Training on data collected by fixtures has currently only preliminarily verified that transfer is possible. On the one hand, the diversity and quality of the data have not been independently verified; on the other hand, more than a quarter of the 10,000 hours of cross-embodiment data used in the paper relies on open-source datasets, and it has not yet been fully tested on more complex tasks and bodies.

In addition, based solely on the officially released videos and public papers, it is really impossible to determine whether Xiaomi has already established a data flywheel for embodied intelligence.

In the public materials released in July, Xiaomi's robot had quite a few stalls and jitters.

But the new continuous-operation video is wide-angle and sped up, making details hard to see.

Even if all the previous problems have been solved, it is still impossible to determine whether this was achieved through manual optimization of individual cases or through a data-driven approach.

However, for Xiaomi's embodied intelligence business, this is still a milestone—

Figure (bipedal humanoid BMW factory project), AgiBot (wheeled robot consumer electronics scenarios), and UBTech (bipedal humanoid Zeekr factory) all achieved continuous operation validation in similar non-standard positions before Xiaomi.

But across the entire embodied intelligence track, there are only these few players.

More accurately, Xiaomi has run into the first tier in landing validation across several comprehensive dimensions: bipedal full-humanoid form, coordinated dexterous manipulation of the upper and lower body, and complex non-standard long-horizon tasks.

And the kill line for embodied intelligence in industrial scenarios is far more demanding than for C-end users:

Standardized workstations can never be faster or more precise than automated robotic arms, nor more cost-advantageous, so embodied intelligence can only replace positions that robotic arms cannot handle—highly non-standard positions requiring dexterous manipulation.

This includes final assembly, material transfer and organization, and so on in automotive production workshops.

The successful validation of 8 consecutive hours of operation in an automotive factory is the first time Xiaomi has crossed the "kill line" of embodied intelligence.

This article comes from the WeChat public account "Intelligent Vehicle Reference" (ID: AI4Auto), author: Jia Haonan, published by 36Kr with authorization.