Xiaomi-Robotics-1: What 100,000 Hours of Training Data Actually Prove

The short answer

Xiaomi-Robotics-1 is a substantial first-party robot-learning result, but its headline needs careful parsing. Xiaomi says the model was pretrained on more than 100,000 hours of real-world manipulation trajectories. Those trajectories were captured with handheld Universal Manipulation Interface (UMI) devices, not by robots operating for 100,000 hours. Robot embodiment alignment came later through a roughly 10,000-hour cross-embodiment post-training mixture, including more than 7,200 hours of in-house mobile-manipulator and dual-arm robot data (technical report).

The technical report then tests three different propositions:

whether larger UMI datasets and larger model variants reduce pretraining error;
whether those scaling gains transfer to four out-of-the-box real-robot tasks after post-training;
whether the pretrained policy can be fine-tuned efficiently for four separate dexterous tasks.

Xiaomi reports positive results for all three. The strongest real-robot evidence includes declared task sets, model and data ablations, success rates, and, for the fine-tuning study, ten trials per task. The paper also reports leading results across four simulation benchmarks.

Important evidence remains unavailable. The results have not been independently reproduced. The 100,000-hour dataset is not public. Trial-level real-robot logs, intervention rules, failure videos, confidence intervals, hardware details, and safety events are not published. The artifact record did improve on August 3: Xiaomi released Apache-2.0 code and checkpoints for post-training, inference, and three simulation evaluation paths. That makes outside testing possible, but it does not retroactively open the proprietary corpus or reproduce the physical trials.

The defensible conclusion is: Xiaomi reports that large-scale handheld manipulation data improved its own post-trained robot policy under bounded company-run tests. The released artifacts now support outside simulation tests, but they are not proof of general robot intelligence, production reliability, or complete reproducibility.

Evidence at a glance

Question	Published evidence	Classification
What is the 100,000-hour corpus?	Real-world manipulation trajectories captured with handheld UMI devices	Large proprietary pretraining corpus, not robot runtime
How is it connected to robots?	About 10,000 hours of post-training data, including more than 7,200 hours from mobile manipulators and dual-arm robots	Cross-embodiment alignment mixture
Does more pretraining data help?	Xiaomi reports out-of-the-box real-robot success rising from 26% without action pretraining to 75% with the full corpus	First-party four-task scaling study
Does model size help?	With 20,000 pretraining hours, Xiaomi reports 61% for 2B, 75% for 5B, and 79% for 10B variants	First-party model-size ablation
Can it learn new tasks efficiently?	Average 75% success across four dexterous tasks with less than ten hours of fine-tuning data per task on average	Forty company-run trials per model condition
Does it lead simulation benchmarks?	Xiaomi reports top aggregate results on RoboCasa, RoboCasa365, VLABench, and RoboDojo	First-party benchmark evaluation
Are code and weights downloadable?	Apache-2.0 post-training and inference code, a 5B checkpoint, a small demo dataset, and three simulation evaluation paths	Public artifacts, with important data and protocol gaps
Is it production deployed?	No customer or recurring operating evidence in the cited sources	Not established

The 100,000-hour number is UMI data, not robot operating time

UMI is a handheld data-collection system designed to capture human manipulation trajectories. An operator moves a gripper-like device through tasks while cameras and motion tracking record observations and actions. This avoids dedicating a physical robot to every collected trajectory and can expand the range of environments, objects, and behaviors represented in training data. Our guide to robot training data collection compares UMI with direct teleoperation, egocentric video, language labels, interventions, EEG, and EMG.

Xiaomi says its UMI corpus covers homes, commercial spaces, industrial sites, and outdoor environments. It divides each trajectory into fixed-length segments and uses an automatic vision-language labeling pipeline to describe the scene transition in each segment. The model learns to predict actions that transform the current scene toward the described target state.

This design has two important consequences.

First, 100,000 hours is a data-volume claim, not an autonomy or deployment claim. It does not mean Xiaomi robots completed 100,000 autonomous operating hours. It also does not reveal how many hours remain after quality filtering, how much repetition exists, how operators and tasks were sampled, or how train and test environments were separated.

Second, UMI actions do not directly match every robot. A handheld gripper and a mobile dual-arm system have different kinematics, observations, action spaces, delays, and failure modes. Xiaomi therefore uses a separate post-training stage to bridge from UMI action generation to robot embodiments and from descriptive state changes to imperative instructions such as a person might give a robot.

This is one form of the embodiment gap between shared data and target-robot execution. The post-training mixture addresses alignment, while calibration, controller behavior, physical trials, contact, safety, and recovery still need target-specific evidence.

What enters post-training

The report describes about 10,000 hours of cross-embodiment post-training trajectories. Its main components are:

more than 7,200 hours of in-house data from mobile manipulators and dual-arm robots across household environments and tasks;
more than 1,000 hours of UMI data relabeled with imperative instructions;
public robot datasets including Bridge V2, RT-1, DROID, and Open X-Embodiment;
vision-language data used alongside action data during training.

These categories should not be added to the 100,000-hour headline as if every hour were equivalent. Handheld trajectories, robot demonstrations, public datasets, and vision-language examples have different embodiments, labels, collection protocols, and uses in the training objective.

The paper gives mixture ratios and high-level composition, which is useful. It does not release the proprietary trajectories, detailed source distribution, collection demographics, duplicate analysis, filtering thresholds, robot hardware bill of materials, or per-task hour allocation. An outside team cannot yet inspect whether measured gains come primarily from data volume, diversity, labeling, quality control, architecture, compute, or some combination.

How the model works

Xiaomi-Robotics-1 couples a pretrained vision-language model with a diffusion transformer using a Mixture-of-Transformers architecture. The vision-language component encodes images and language. It also predicts candidate action chunks and scores them. The diffusion transformer generates a continuous action chunk through flow matching while conditioning on the observation, instruction, and robot state.

The paper studies 2-billion, 5-billion, and 10-billion-parameter variants. These sizes refer to the vision-language component, while the diffusion transformer is held at roughly 230 million parameters. The comparison is therefore more controlled than changing the entire system at once, although training compute and inference latency for each variant are not fully normalized into a practical robot-deployment comparison.

For an introduction to these model families, read Robot Foundation Models. Our LeRobot v0.6.0 practical guide explains how model availability, evaluation tools, robot integrations, and human correction fit into a usable development stack. The robot benchmark evidence audit provides a reusable checklist for deciding when success rates from different tasks, embodiments, and trial protocols can be compared.

Real-robot scaling: the strongest result and its boundary

Xiaomi evaluates out-of-the-box behavior on four tasks in unseen environments after post-training:

desktop cleaning;
clothes folding;
sofa tidying;
shoe tidying.

For the 5-billion-parameter model, the report says overall success rose from 26% with no action pretraining to 53% with 12.5% of the UMI corpus, then continued upward to 75% with the full corpus. The paper says doubling the last tested data slice from 50% to 100% added six percentage points, with no observed saturation inside the tested range.

A model-size study starts from checkpoints trained on 20,000 hours of UMI data. Xiaomi reports overall success of 61% for 2B, 75% for 5B, and 79% for 10B. Three of four tasks improved consistently with scale, while sofa tidying was similar for the 5B and 10B variants.

This is meaningful because the paper connects a pretraining ablation to real robot outcomes rather than reporting only validation loss. But it is still a bounded company evaluation. The main text does not provide a complete trial count for this out-of-the-box scaling figure, task-by-task confidence intervals, randomization procedure, intervention policy, reset rules, failure taxonomy, or all-run video archive. The four tasks are related household manipulation tests on Xiaomi’s own systems, not a cross-laboratory or cross-industry evaluation.

The result supports this statement: within Xiaomi’s tested pipeline, more UMI pretraining data and a larger model generally improved aggregate success after robot alignment. It does not support a universal claim that adding any manipulation data will improve any robot task.

Fine-tuning: 75% means 30 successes in 40 task trials

The paper separately evaluates adaptation to four dexterous tasks:

removing clothes from a hanger;
placing a pillow into a pillowcase;
folding a fitted sheet;
hanging clothes on a rack.

Each model is evaluated for ten trials per task. In the low-data condition, Xiaomi-Robotics-1 uses less than ten hours of fine-tuning data per task on average and reports 75% average success plus 90% average milestone progress across the four tasks. With ten trials for each of four tasks, a 75% average corresponds to 30 successful trials out of 40 if tasks are equally weighted, as the protocol states.

Xiaomi compares against π0.5 and an internally implemented GR00T-N1.5 condition. It reports 40% average success for π0.5 and 27.5% for GR00T-N1.5 in the same low-data setting. These figures are more useful than an unbounded demo, but they should not be read as a general model leaderboard.

The report says the baseline models were fine-tuned within Xiaomi’s asynchronous setting. Hardware, implementation choices, checkpoint selection, preprocessing, and optimization can materially affect a comparison. Forty trials per condition also produce coarse percentages: one trial changes the aggregate by 2.5 points. No confidence intervals or independent replication are supplied.

Simulation results need one published-number caution

Xiaomi reports leading aggregate results on RoboCasa, RoboCasa365, VLABench, and RoboDojo. The paper’s body reports 57.4% average success on RoboCasa365 and says this is 10.8 percentage points above the prior best. The abstract reports 57.6% and compares it with 46.6%.

That 0.2-point inconsistency is small enough not to change the ranking claim, but it matters for exact reporting. Until Xiaomi corrects or explains it, 57.4% in the results section and 57.6% in the abstract should be retained as two conflicting first-party values, not silently merged.

Simulation benchmarks can test many tasks under repeatable conditions. RoboCasa365, for example, evaluates 50 tasks spanning seen atomic tasks, seen composite tasks, and unseen composite tasks. Xiaomi says it trained on the officially released 100 demonstrations per task. Even so, simulation success does not establish real-world safety, sensing robustness, gripper durability, latency, intervention rate, or production economics.

The suitcase demonstration is evidence of sequence length, not reliability

The project page shows a room-level mobile manipulation sequence in which the robot packs a suitcase over more than ten minutes. Xiaomi describes it as autonomous.

A long continuous sequence can reveal more than a montage because navigation, object handling, transitions, and recovery must remain coordinated. One published run still lacks a denominator. The sources do not state how many full attempts were made, whether any failed attempts were excluded, what interventions or resets were allowed, whether the environment was represented in training, or how the same checkpoint performs across different rooms.

The right classification is first-party autonomous demonstration, not a pilot, commercial service, or recurring deployment. The evidence labels in the humanoid robot availability tracker apply equally well here: a demonstration can be technically informative without proving sustained field operation.

Artifact availability: what Xiaomi released on August 3

The original July 21 audit found a paper, a project page, and a README-only repository. Xiaomi materially changed that state on August 3. Its release commit added:

Apache-2.0 Python code for post-training and client-server inference;
a public, ungated Xiaomi-Robotics-1-5B checkpoint;
a small example post-training dataset with synchronized egocentric and wrist-camera videos;
fine-tuned checkpoints and evaluation launchers for RoboCasa, RoboCasa365, and VLABench;
company-generated RoboCasa logs and a RoboCasa365 episode summary.

The released workflow is more substantial than a model file alone. The post-training guide specifies the input format, training launcher, checkpoint conversion, and runtime client. It requires Python 3.9 or newer, a CUDA GPU, the Qwen3-VL-4B-Instruct backbone, FlashAttention, and exactly version 4.57.1 of Transformers, because other versions are not verified by Xiaomi. The runtime accepts three camera images, an instruction, and robot state, then returns action outputs.

The evaluation release also improves the denominator record. Xiaomi’s RoboCasa guide reports 24 tasks with 100 episodes each and provides per-task logs. Its listed per-task values average 74.2%, slightly below the paper’s 74.5% aggregate. Xiaomi warns that results may vary across GPU machines. The RoboCasa365 guide specifies 50 trials across each of 50 tasks and reports 1,432 successes in 2,500 episodes, or 57.28%, compared with the paper body’s 57.4% rounded result and the abstract’s 57.6% value.

This is now a runnable research release, not merely an announced one. It exposes enough material for outside teams to inspect the implementation, fine-tune the 5B model, and attempt three simulation evaluations. It does not include the pretraining pipeline or 100,000-hour corpus, the complete 10,000-hour alignment mixture, the 2B and 10B variants, a RoboDojo evaluation path, or the physical-robot evaluation stack. No independent replication was identified during this update.

What the evidence does not establish

The current sources do not establish:

public access to the 100,000-hour pretraining dataset;
public access to the full pretraining pipeline and proprietary alignment data;
released 2B and 10B variants or a RoboDojo evaluation path;
independent reproduction of real-robot or simulation results;
trial-level logs for the out-of-the-box tasks;
intervention, teleoperation, reset, or excluded-run counts;
success under safety-rated speed and force constraints;
cross-robot transfer without Xiaomi’s post-training stage;
production uptime, maintenance burden, throughput, or economics;
a customer pilot, commercial product, or recurring deployment.

A large training corpus can improve a model while leaving these product and operating questions unanswered. For broader context on what turns a learned policy into a physical system, see What Is Physical AI? and Embodied AI.

What to watch next

The next evidence gates are concrete:

Independent artifact run: successful installation, checkpoint loading, fine-tuning, and simulation evaluation by an outside team.
Remaining evaluation release: the RoboDojo path, baseline configurations, physical-robot code, and trial-level physical results.
Data documentation: collection distribution, filtering, deduplication, quality checks, consent, and train-test separation.
Real-robot protocol: denominators for every task, intervention and reset rules, all-run or sampled failure video, and confidence intervals.
Cross-laboratory testing: matched evaluation by an independent team on disclosed hardware.
Field evidence: sustained operating hours, task throughput, safety events, interventions, maintenance, and version history.

Verdict

Xiaomi-Robotics-1 advances the public evidence for scaling manipulation policies in four ways. It separates large-scale UMI pretraining from robot alignment, connects pretraining scale to real-robot task success, reports low-data fine-tuning with a clear ten-trial-per-task protocol, and now provides code, checkpoints, and simulation evaluation paths that outside teams can test.

The result should remain bounded. The 100,000 hours are handheld manipulation trajectories, not robot autonomy hours. The real-robot evaluations are company-run and narrow. The proprietary corpus and full physical evaluation record remain closed. The released simulation records also preserve small differences among the paper and repository aggregates rather than eliminating them.

The strongest current reading is that Xiaomi has moved from an inspectable paper to a runnable, partly reproducible research release. The next test is whether outside teams can match the simulation results and extend the evidence to disclosed physical-robot protocols.

Frequently asked questions

What is Xiaomi-Robotics-1?

Xiaomi-Robotics-1 is a vision-language-action model trained in two stages. Xiaomi first pretrains it on more than 100,000 hours of manipulation trajectories captured with handheld UMI devices, then aligns it to mobile manipulators and dual-arm robots with about 10,000 hours of cross-embodiment post-training data.

Is the 100,000-hour dataset robot data?

Not in the usual sense. The pretraining trajectories were captured with handheld Universal Manipulation Interface devices rather than physical robots. Xiaomi separately reports more than 7,200 hours of in-house data from mobile manipulators and dual-arm robots in the post-training mixture.

Did Xiaomi-Robotics-1 achieve a 75% success rate?

Xiaomi reports several 75% results under different protocols. Its 5-billion-parameter model reached 75% across four out-of-the-box real-robot tasks when pretrained on the full UMI dataset. In a separate fine-tuning study, the model averaged 75% across four dexterous tasks with less than ten hours of task data on average and ten trials per task. These are first-party evaluations, not independent results.

Are Xiaomi-Robotics-1 code and model weights available?

Yes, in a bounded form. On August 3, 2026, Xiaomi released Apache-2.0 code for post-training and inference, a public 5-billion-parameter checkpoint, a small post-training demonstration dataset, and evaluation paths for RoboCasa, RoboCasa365, and VLABench. The proprietary 100,000-hour UMI corpus, full real-robot data, and a complete independent reproduction are still unavailable.

Does Xiaomi-Robotics-1 prove that more data always improves robot performance?

No. Xiaomi reports monotonic gains within its tested data and model scales, including four company-run real-robot tasks. The result supports a scaling relationship under that protocol, but it does not establish the same return across other robots, data sources, tasks, safety constraints, or production environments.

Is Xiaomi-Robotics-1 deployed in production?

No production deployment is established by the cited sources. The report describes research evaluations, simulation benchmarks, and company demonstrations. It does not identify a customer deployment, recurring operating record, commercial access model, intervention rate, or safety qualification.