Xiaomi-Robotics-1 Brings Scaling Laws to Robot VLA Models with 100K+ Hours of Real Trajectories

Introduction

Robot learning has not enjoyed the same data abundance as language or vision. Web text and images can be collected at massive scale, but manipulation data requires physical interaction, careful setup, and often expensive robot time. Xiaomi-Robotics-1 is an attempt to test a central question for embodied AI: if robot policies are trained on much larger and more diverse real-world trajectories, do they scale in a predictable and useful way?

The paper presents Xiaomi-Robotics-1 as a vision-language-action foundation model for mobile manipulation. Its pre-training stage uses more than 100,000 hours of real-world UMI manipulation trajectories across over 1,700 scenarios, including homes, commercial spaces, industrial sites, and outdoor environments. The model is then aligned with real robot embodiments and imperative natural-language instructions.

Key Points

• A much larger real-world data base: Instead of relying only on trajectories collected from a specific robot platform, the model is pre-trained on embodiment-free UMI manipulation data. This gives it exposure to a broader range of objects, scenes, and interaction patterns.
• Automatic language supervision: Manually annotating 100,000+ hours of trajectories would be unrealistic. The team splits long trajectories into clips and uses a vision-language model to describe changes in grippers and objects. These descriptions become language conditions for learning actions.
• A two-stage training recipe: Pre-training builds general action-generation capability from state-transition descriptions. Post-training then adapts those capabilities to real robots using cross-embodiment datasets, including more than 7,200 hours of in-house real-robot data collected in real homes.
• Scaling behavior transfers: Experiments report that increasing either pre-training data or model size reduces action prediction error. More importantly, stronger pre-training also leads to better real-robot performance after alignment, with no clear saturation reported in the provided results.
• High data efficiency for new tasks: On four real-world tasks—phone packing, printer refilling, laundry loading, and box packing—the model reaches a 75% overall success rate with fewer than 10 hours of demonstrations per task on average, compared with 40% for π0.5 under the same data budget. With fewer than 40 hours per task on average, success rises to 85%.
• Strong simulation benchmark results: The reported average success rates are 74.5% on RoboCasa, 57.4% on RoboCasa365, 59.1% on VLABench, and 13.93% on RoboDojo, described in the source as the best reported performance across all four benchmarks.

Why It Matters

The main contribution is not merely a new robot policy, but a data-centric path for scaling embodied intelligence. Xiaomi-Robotics-1 suggests that large amounts of real manipulation experience, even when not tied to one robot body, can become a useful foundation once converted into language-conditioned training examples.

That does not mean general-purpose household robots are solved. Real deployment still involves safety, long-horizon planning, recovery from failure, hardware constraints, and evaluation gaps between benchmark tasks and open-ended environments. The reported results show strong trends, not the end of the problem.

Still, the work is an important signal: progress in robotics may increasingly depend on scalable data collection and automated annotation pipelines, not only on better robot hardware or isolated task demonstrations.

Source: Hugging Face Daily Papers