Xiaomi Robotics-1: More Data Beats Bigger Models — How Xiaomi Is Reshaping Robot Training

The Data Problem in Robot AI

Large language models have it easy. They can train on vast swaths of the public internet — trillions of words from articles, books, forums, and code repositories. Robot AI faces an entirely different reality. Useful data on robot movement — how to grip, lift, rotate, and place objects — is scarce and expensive to produce. The standard approach has humans remotely guiding a physical robot through every motion, a painstaking process that is slow, costly, and often yields repetitive data from the same tasks in identical settings.

Xiaomi's new Xiaomi-Robotics-1 model aims to break this bottleneck. The model follows a scaling pattern similar to large language models: its performance improves as it trains on more data. But unlike LLMs, which can draw from the entire written internet, Xiaomi had to build its own dataset from scratch — and the way they did it is arguably the most interesting part of the research.

A Handheld Gripper Replaces Expensive Robots

To bypass the data collection bottleneck, Xiaomi largely abandoned physical robots during the data-gathering phase. Instead, the team developed portable handheld grippers equipped with cameras — a simple device that a person picks up and operates by hand. This setup allows recording manipulation tasks in kitchens, offices, stores, factory floors, and outdoor spaces without a robot being present at all.

The result was over 100,000 hours of motion recordings across more than 1,700 different environments — an unprecedented scale for robot training data. A dataset of this magnitude creates its own challenge: each recording needs a textual description that the model can learn from. Manual labeling was impractical, so Xiaomi employed another AI model to describe each motion segment in text. The team reports that the entire dataset was labeled in approximately two weeks.

This approach matters beyond Xiaomi's labs. For countries and regions where robotics research is still nascent — including Georgia and the broader Caucasus region — the ability to collect high-quality training data with inexpensive, portable tools dramatically lowers the barrier to entry. Research groups no longer need million-dollar robot fleets to contribute meaningfully to robot AI development. A handheld gripper and a camera may be enough to start building valuable training datasets.

Xiaomi then transferred that training to physical robots, including wheeled models and dual-arm systems. The model still had to account for the kinematic differences between a handheld gripper and a robot arm — a non-trivial domain adaptation challenge that the team successfully addressed.

Why More Data Matters More Than Model Size

The central finding of the Xiaomi-Robotics-1 research is both clear and consequential: while increasing model size does improve performance, adding more training data produces dramatically larger gains than adding more compute. The researchers state unequivocally that progress in robot AI will depend primarily on collecting larger and more varied datasets.

This finding aligns with what researchers demonstrated for visual data in early March 2026. For language models, the established scaling law dictates that model size and data volume should grow at roughly the same rate to make optimal use of a fixed compute budget. That balance shifts when a model must process images rather than just text — in vision-language domains, adding data helps far more than increasing model size. Xiaomi's work extends this principle to the robotics domain, suggesting that the data-centric scaling paradigm applies broadly across AI modalities.

The results with physical robots are striking. As Xiaomi increased the amount of training data, the model's success rate in unfamiliar environments rose from approximately 25 percent to 75 percent. The researchers note that they have not yet reached a point where more data stops improving performance — the scaling curve shows no signs of plateauing. This suggests that the potential for data-driven improvement in robot AI remains vast and largely untapped.

On standard robot AI benchmarks, Xiaomi-Robotics-1 posted the best results to date. In one demonstration, a robot packed a suitcase without human assistance — a task that took more than ten minutes and required the robot to navigate across an entire room. On the RoboCasa365 leaderboard, Xiaomi-Robotics-1 leads by a wide margin, especially on unseen composite tasks. On the RoboDojo benchmark, it scored approximately 58 percent higher than the runner-up, though absolute success rates remain modest — a reminder of how much work still lies ahead.

Adapting to New Tasks with Minimal Data

A foundation model's true value lies in its ability to adapt quickly to new tasks. Xiaomi tested Xiaomi-Robotics-1 on four unseen tasks: packaging a phone, loading laundry into a washing machine, feeding paper into a printer, and packing items into a box.

With less than ten hours of training data per task, the model achieved an average success rate of 75 percent. For comparison, a competing model from Physical Intelligence managed only 40 percent on the same benchmark. Xiaomi reports that its model performed especially well with soft, deformable materials such as paper and on tasks requiring movement across a room — precisely the kinds of challenges that have historically proven difficult for rigid, scripted robot systems.

This few-shot adaptation capability is particularly relevant for practical deployment scenarios. In real-world applications — whether in a Georgian warehouse, a Batumi hotel, or a Tbilisi factory — robots will need to learn new tasks quickly with minimal demonstration data. Xiaomi-Robotics-1's performance suggests that this vision is increasingly attainable.

The Landscape of Robot AI Research

Xiaomi's approach is one of several competing strategies in the rapidly evolving robot AI landscape. A May survey of World Action Models reviewed approximately 100 studies and found that teleoperation data — the traditional approach — is precise but costly and confined to limited settings. Xiaomi's handheld grippers directly address these limitations.

Nvidia, Carnegie Mellon University, and UC Berkeley took a different route with the ENPIRE project, using AI coding agents to autonomously teach a fleet of eight robots how to grasp objects. Nvidia is also pursuing synthetic training data generation, aiming to transform robotics' data problem into a compute problem — a fundamentally different philosophy from Xiaomi's real-world data collection approach.

China's BAAI research institute is exploring yet another direction. Their Orca world model learns from 125,000 hours of video without any action labels and still matches the specialized pi0.5 model on five manipulation tasks. Where Xiaomi automates labeling and applies it at scale, Orca attempts to eliminate labels entirely — a more radical approach that could prove valuable if it scales.

Physical Intelligence serves as a reference point for both camps. The company drew criticism in April after claiming its pi0.7 model could generalize when it may have been recalling similar training data. Xiaomi's finding that data volume matters more than compute makes this question harder to resolve — if a model has seen enough varied data, distinguishing genuine generalization from sophisticated pattern matching becomes increasingly difficult.

What This Means for Georgia and Emerging Tech Ecosystems

For SiTech Georgia and the broader Georgian tech community, Xiaomi's research carries several important implications. First, the democratization of data collection — using simple handheld tools instead of expensive robots — means that smaller teams and startups can participate in robot AI development. A Batumi-based startup could theoretically collect manipulation data in local settings, contributing to global robotics datasets or building specialized models for local applications.

Second, the "data over compute" finding is encouraging for ecosystems with limited access to massive GPU clusters. If data quality and quantity can partially compensate for computational constraints, teams in emerging tech hubs can compete on innovation rather than infrastructure spending. This is a welcome shift from the prevailing narrative that AI research increasingly belongs only to those with billion-dollar compute budgets.

Third, Xiaomi's open-source commitment — the model and code will be released on GitHub and Hugging Face — means immediate access for researchers and developers worldwide. The team recently released Xiaomi-Robotics-0 under an open license, focused on fast, real-time operation. Georgian universities, AI labs, and independent developers can experiment with these models, build upon them, and potentially contribute improvements back to the community.

The Road Ahead

Xiaomi-Robotics-1 represents a significant step forward, but it also highlights how much remains to be done. Absolute success rates on benchmarks, while leading, are still below what would be required for reliable commercial deployment in unstructured environments. The model's strengths — handling deformable objects, navigating rooms, adapting to new tasks — are real, but gaps remain.

The research also fits into Xiaomi's broader strategic picture. When the company introduced its MiMo-V2 models in March, it stated that those AI agents could eventually control robots. Xiaomi-Robotics-1 is a key piece of that vision — a bridge between language-capable AI agents and physical robotic systems that can act in the real world.

Perhaps the most profound takeaway is this: the era of robot AI is no longer primarily an engineering problem of building bigger models. It is increasingly a data problem — one of collecting, labeling, and scaling diverse real-world interaction data. Xiaomi has shown that with creativity in data collection and a commitment to scale, even the physical constraints of robotics can be addressed through data-centric AI. For researchers, startups, and tech communities worldwide — including Georgia — this opens a door that many thought was firmly closed.

📖 Source