⏱️ Lectura: 11 min
A robotics model trained with zero robots during its initial phase already nearly doubles the success rate of its most cited rival on new tasks. That is what Xiaomi reports with Xiaomi Robotics 1 (XR-1), a foundational model for robotic manipulation that reverses the usual order: it first learns from robot-free video at massive scale, and only afterward is fine-tuned with real data from arms and grippers.
📑 En este artículo
The bet responds to a well-known problem in robotics: unlike text or images, real robot data is expensive to collect and scarce. Xiaomi Robotics 1 tries to break that barrier by separating what the model learns about the world (robot-free pretraining) from how it applies that to a specific physical body (post-training with a robot).
TL;DR
- Xiaomi Robotics 1 (XR-1) pretrains on 100,000 hours of robot-free video (UMI data) across more than 1,700 scenarios.
- Post-training uses more than 7,200 hours of real data recorded in homes: tidying sofas, shoe racks, and kitchens.
- With less than 10 hours of data per task, XR-1 achieves 75% overall success, nearly double π0.5’s 40%.
- With less than 40 hours per task, XR-1 rises to 85% success versus 53% for π0.5, the reference model from Physical Intelligence.
- In laundry loading (folding clothes) with less than 40 hours, XR-1 reaches 100% success versus 50% for π0.5.
- XR-1 achieves the best result in all four simulation benchmarks it was evaluated on, according to Xiaomi.
- Pretraining shows a clean scaling law: action error drops consistently with more data and parameters.
- Success on real robots after post-training rises predictably with more pretraining data, with no signs of saturation.
What happened
Xiaomi published the results of Xiaomi Robotics 1, a foundational model for robotic manipulation trained in two stages, following the same paradigm used by language models: broad pretraining followed by specialized fine-tuning. The key difference is that the pretraining stage uses no robot data whatsoever.
Instead of recording robotic arms performing tasks, Xiaomi used 100,000 hours of embodiment-free trajectories captured with the UMI method (Universal Manipulation Interface), a handheld device that records human manipulation without needing a physical robot. Those trajectories cover more than 1,700 different scenarios: homes, businesses, industrial sites, and outdoor spaces.
After that phase, the team applied post-training with real robot data: more than 7,200 hours recorded in real homes, solving tasks like tidying a sofa, sorting a shoe rack, or putting away kitchen utensils. That combination (lots of robot-free video plus a moderate amount of real data) is the core of the proposal.
Context and history: why robotics did not follow the scaling law
Language and vision models have benefited for years from empirical scaling laws: more data, more parameters, and more compute produce predictably better capability. Robotics was left out of that trend because obtaining real robot data, unlike text scraped from the internet, requires physical hardware, human supervision, and real recording time. That bottleneck limited how much policy models for robots could scale.
The UMI method, originally developed in academia as a universal manipulation interface, offers a way out of that problem: capturing manipulation demonstrations with a portable gripper that a person can carry into any environment, without relying on an expensive robot during recording. Xiaomi takes that idea to a scale that had not been publicly reported until now: 100,000 hours of trajectories.
The challenge with a dataset that large is not just recording it, but labeling it. Manually annotating 100,000 hours of video is unfeasible. That is why Xiaomi built an auto-labeling pipeline: it splits each long trajectory into fixed-length segments and uses a vision-language model (VLM) to describe, for each segment, how the state of the gripper and the objects it interacts with changes.
How the two-stage training works
The training of Xiaomi Robotics 1 follows the same two-phase scheme popularized by large language models: a broad phase that builds general representations, and a fine-tuning phase that aligns those representations with a specific use case.
flowchart TD
A["100,000 hours of UMI video"] --> B["Pretraining: action from state descriptions"]
B --> C["7,200+ hours of real robot data"]
C --> D["Post-training: embodiment and instruction alignment"]
D --> E["Robot executes new tasks in unseen environments"]In the pretraining stage, the model learns to generate actions that move a scene toward the state described by the automatically generated text. It does not yet know what a specific robot is or how its joints move: it learns a general notion of how the world changes when someone manipulates objects.
In the post-training stage, that general knowledge is translated to a specific robotic body along two axes. First, embodiment alignment, which uses real robot data to map the general action-generation capability to the actual physical actuators. Second, instruction alignment, which turns the model from “generate actions given a described state change” into “understand a natural-language instruction and execute it directly”.
💭 Key point: the central finding of the paper is that action error during pretraining drops consistently as data and model size grow, and that this improvement carries over to real robot success after post-training, even in environments and with objects never seen before.
Technical details and performance
Xiaomi evaluated the post-trained model in two complementary scenarios: efficient adaptation to new tasks and standard simulation benchmarks.
For adaptation to new tasks, the team measured how long it takes XR-1 to learn complex tasks it had never seen before (packing a phone, refilling a printer, loading a washing machine, packing boxes) using only a few hours of demonstrations per task, comparing it against π0.5, the reference model from Physical Intelligence.
| Task | XR-1 <10h/task | π0.5 <10h/task | XR-1 <40h/task | π0.5 <40h/task |
|---|---|---|---|---|
| Pack phone | 70% | 30% | 80% | 40% |
| Refill printer | 70% | 20% | 60% | 20% |
| Load washing machine | 80% | 40% | 100% | 50% |
| Pack boxes | 80% | 70% | 100% | 100% |
| Overall average | 75% | 40% | 85% | 53% |
With a budget of less than 10 hours of demonstrations per task, XR-1 reaches 75% overall success, nearly double the 40% achieved by π0.5 with the same data budget. Raising the budget to less than 40 hours per task, XR-1 reaches 85% versus 53% for π0.5.
In simulation, Xiaomi evaluated the model on four standard benchmarks from the robotics community and reports the best average result on all four, which the team says confirms that the generalization gains observed in the real world also hold up in simulated evaluation.
Although Xiaomi did not publish the model’s code or weights, it did describe the typical inference pattern of a vision-language-action (VLA) model like this one: it receives an image of the scene and a text instruction, and returns a sequence of low-level actions for the actuator. A conceptual, unofficial example of how that inference loop is structured in a typical VLA:
vla_model = load_policy("xr1-checkpoint")
observacion = camara.capturar_frame()
instruccion = "put the dishes away in the cupboard"
accion = vla_model.predecir(
imagen=observacion,
texto=instruccion,
historial_estado=robot.estado_actual()
)
robot.ejecutar(accion)This block illustrates the standard cycle of a VLA model: observation plus instruction go in, a low-level action comes out. The interesting logic in XR-1 is not in that cycle, but in how the policy that solves it was trained.
The auto-labeling pipeline used in pretraining, which describes state transitions per video segment, can be summarized conceptually like this:
for clip in dividir_en_segmentos(trayectoria_umi, duracion_fija=SEGMENT_LEN):
descripcion = vlm.describir_transicion(
frame_inicial=clip.primer_frame,
frame_final=clip.ultimo_frame,
foco="gripper and manipulated object"
)
dataset.agregar(clip, descripcion)That pipeline is what turns 100,000 hours of raw video into a labeled corpus usable for training action generation, without relying on segment-by-segment human annotation.
How to approach this today
Xiaomi did not release weights or code for Xiaomi Robotics 1, so there is no installation command to try the model itself. What is reproducible today is the methodological part: the UMI method that feeds the pretraining is a public research project, documented at umi-gripper.github.io, with blueprints for the portable gripper and the data capture pipeline.
If you work in manipulation robotics and don’t have a fleet of robots, the practical lesson from XR-1 is replicable at smaller scale: recording human demonstrations with a portable UMI-style interface is much cheaper than operating a robotic arm, and that video can be used to pretrain action representations before touching real hardware. Verifying that this pretraining actually helps is done the same way as in the paper: measure the action error on a separate validation set as the dataset grows, and confirm that the curve drops monotonically before investing in the expensive stage of real data.
💡 Tip: before collecting real robot data, validate your embodiment-free pretraining by measuring action error on a fixed validation set; if it does not drop consistently as you add more data, the problem is in the labeling pipeline, not the amount of data.
Impact and analysis
The most relevant result from XR-1 is not that a large model performs well, but that the improvement obtained in robot-free pretraining transfers predictably to success with real robots after post-training. That is exactly what was missing to justify investing in embodiment-free data at scale: that it is not just a cheap way to generate more data, but a cheap way to generate data that actually moves the needle in the physical world.
The comparison against Physical Intelligence’s π0.5 matters because it is not a weak baseline: it is one of the most cited reference models in generalist manipulation with real robots. XR-1 nearly doubling its success rate with the same data budget per task suggests that separating pretraining (cheap, robot-free) from post-training (expensive, with a robot) is a strategy with better return per hour of collected data than training directly with robot data from scratch.
The honest limitation of the announcement is that all the evidence comes from Xiaomi itself, with no public weights or independent evaluation yet. The four simulation benchmarks where XR-1 claims to lead are also not detailed with complete figures in the published material, which makes it impossible, for now, to verify those numbers externally.
What’s next
Xiaomi Robotics 1 is presented as a foundational model meant to serve as a base for later applications, not as a finished product. The team itself already uses it on two fronts: rapid adaptation to new tasks with few hours of data, and evaluation on the community’s standard simulation benchmarks.
The natural next step, if Xiaomi follows the pattern of other foundational model labs, is to scale the embodiment-free pretraining even further (more hours, more scenarios) and measure whether the scaling curve keeps holding without saturation, as the team reports up to the current point. It also remains to be seen whether Xiaomi will integrate XR-1 into its own home robotics products, given that a good portion of the post-training data is already oriented toward household tasks.
📖 Summary on Telegram: View summary
Try it yourself: check out the UMI project at umi-gripper.github.io to see firsthand how robot-free manipulation is recorded.
Frequently Asked Questions
What is Xiaomi Robotics 1?
It is a foundational model for robotic manipulation from Xiaomi, trained in two stages: pretraining with robot-free video and post-training with real robot data.
What does “embodiment-free” or UMI mean?
It means the data is recorded without using a physical robot, with a portable gripper that a person carries by hand. UMI (Universal Manipulation Interface) is the method Xiaomi used to capture those 100,000 hours.
Did Xiaomi release the code or weights for XR-1?
No, to date Xiaomi has not published the model’s weights or code. What is public and reproducible is the UMI method on which the pretraining is based.
How does XR-1 compare to π0.5?
With less than 10 hours of data per task, XR-1 achieves 75% average success versus 40% for π0.5. With less than 40 hours, XR-1 rises to 85% versus 53% for π0.5.
How did Xiaomi label 100,000 hours of video without manual annotation?
With an automatic pipeline: it splits each trajectory into fixed-duration segments and uses a vision-language model to describe how the state of the gripper and the objects changes in each segment.
Are XR-1’s results independently verified?
Not yet. All published figures come from the Xiaomi Robotics team itself; there are no public weights or external evaluation available for now.
References
- Xiaomi Robotics: official page for the Xiaomi Robotics 1 project, with data, methodology, and results tables.
- Universal Manipulation Interface (UMI): the research project on which the embodiment-free data capture is based.
- Physical Intelligence: the lab that develops π0 and π0.5, the model used as a comparative reference.
- Wikipedia: Xiaomi: general context about the company and its hardware and robotics divisions.
📱 Enjoy this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Erhan Astam en Unsplash
0 Comments