Xiaomi-Robotics-1 Shows That More Data Beats Bigger Models for Training Robots to Move

Xiaomi’s new Robotics-1 system demonstrates that scaling motion data — not model parameters — is the most reliable lever for teaching robots dexterous movement. Rather than deploying expensive physical biped or arm systems, the team used handheld grippers equipped with cameras to record more than 100,000 hours of manipulation across kitchens, offices, factory floors and outdoor spaces. The dataset was fully labeled in roughly two weeks by another AI model, compressing a process that previously cost months of human annotation.

The result is a robot AI that follows spoken and written commands in unfamiliar environments, adapts to new tasks with little additional training, and generalizes across different physical platforms — including wheeled models and dual-arm systems. The key finding is counterintuitive for anyone who has watched the LLM race: increasing the model size helped, but doubling the data produced much bigger gains. That mirrors earlier results showing that visual data scales better than compute for embodied AI.

Embodied AI has a data problem that LLMs do not. Language models ingest the internet; robot models need physical recordings of joints, forces, contacts and outcomes. Previous robotics datasets have been small, narrow and expensive, partly because training a real robot to execute 100,000 grasps can take years of wall-clock time and frequent hardware repairs. Xiaomi sidestepped that bottleneck by treating humans with handheld grippers as teleoperated data generators rather than trying to build autonomous collectors.

Why scaling motion data matters more than model parameters

The Xiaomi team combined its own apartment recordings with open-source robot datasets and annotated UMI data for post-training. That hybrid approach meant the model saw both first-person manipulation inside realistic homes and structured lab demonstrations, giving it a broader prior distribution than any single-source dataset. The post-training step used supervised fine-tuning rather than reinforcement from human feedback, which is cheaper and faster to iterate when the base dataset is already diverse.

Performance gains tracked data volume closely, not parameter count. When the researchers doubled the training set, task success rates improved materially; when they kept data constant and increased model size, improvements were smaller and hit diminishing returns sooner. That finding aligns with a broader shift in robot AI research: the community is beginning to treat dataset curation as a first-class engineering problem, with papers on synthetic data generation, sim-to-real transfer and human teleoperation competing with pure architecture hunting.

The commercial implication is that companies with large manufacturing or logistics operations — Xiaomi included — can bootstrap their own robotics data by simply filming humans doing the work. The AI labeling pipeline then turns raw video into training pairs at scale. That lowers the barrier for any organization that already has physical infrastructure and wants to add embodied AI, because the expensive part becomes compute, not hardware or human labor.

What Xiaomi’s approach means for the robotics data race

The robotics industry has been waiting for a clear formula that lets it escape the simulation gap — the tendency of models trained in simulators to fail on real hardware because physics and textures do not transfer perfectly. Xiaomi’s result is not a complete solution because it still requires real-world demonstrations, but it shows that the demonstrations do not need to come from robots. If humans with cheap grippers can generate enough diverse data, the robotics data problem becomes a logistics and labeling problem rather than a hardware design problem.

Competitors such as Tesla’s Optimus team, Boston Dynamics and Figure AI have bet heavily on scaling hardware fleet size to generate data. Xiaomi suggests a different route is possible: treat data generation as an offline, crowdsourced process and reserve expensive hardware only for the final fine-tuning and validation stages. That approach is especially attractive for consumer and commercial robots that must operate in unstructured environments — kitchens, offices, warehouses — where simulators struggle to reproduce noise, clutter and edge cases.

The remaining open question is whether data-can-do-it generalizes to bipedal locomotion and heavy manipulation, which require whole-body coordination that handheld grippers cannot capture. Xiaomi tested on wheeled and dual-arm systems but did not release a humanoid. If the data-scaling thesis holds for full-body control, it would reshape how hardware companies allocate capital: fewer robots, more cameras and better labeling pipelines. That would be a structural advantage for software-first AI labs over traditional automation builders.

Source: The Decoder