Xiaomi Robotics-1: The Robot Foundation Model That Learned From 100,000 Hours of Real-World Tasks
Xiaomi's robotics division dropped a bombshell this week. The company released Xiaomi-Robotics-1, a robot foundation model trained on over 100,000 hours of real-world manipulation trajectories — more than 11 years of continuous robot operation. The model, detailed in a preprint paper, shows that robot policy models can scale the same way language models do: more data and bigger parameters produce predictably better results.
The paper, posted to arXiv under ID 2607.15330, describes a vision-language-action model that combines large-scale embodiment-free pre-training with a modest amount of real-robot data in a post-training stage. The result is a robot that can pick up new tasks like phone packing, printer refilling, and laundry loading from just a few hours of demonstrations.
Breaking the Data Bottleneck
Foundation models in language and vision keep pushing forward by riding empirical scaling laws: capability tracks data, parameters, and compute. Robotics has missed out on this trend. Large-scale, high-quality data is hard to come by, and that scarcity has capped how far policy models could scale.
Xiaomi-Robotics-1 takes a step toward answering what robots can do under genuinely large-scale training. The model combines large-scale embodiment-free pre-training using UMI (Universal Manipulation Interface) data with a modest amount of real-robot data in a post-training stage.
For pre-training, the team used 100,000 hours of embodiment-free UMI trajectories spanning more than 1,700 scenarios — household, commercial premises, industrial sites, and outdoor spaces. They developed a scalable auto-labeling pipeline that divides trajectories into fixed-length segments and annotates each segment with language descriptions of scene state transitions.
For post-training, they leveraged cross-embodiment datasets containing in-house robot data, filtered open-source robot data, and a set of high-quality UMI data. The in-house data alone covered over 7,200 hours of real-robot data in real homes, covering tasks like tidying a sofa, sorting a shoe cabinet, and putting away kitchenware.
Scaling Behavior That Mirrors LLMs
The paper's most striking finding is that pre-training shows a clean scaling behavior: as data and model size grow, validation action error steadily decreases. This mirrors the scaling laws that have driven progress in large language models — and it's the first time robotics has demonstrated this kind of predictable improvement at scale.
After post-training, Xiaomi-Robotics-1 can be used out-of-the-box to perform a wide range of mobile manipulation tasks in the real world. The team evaluated the post-trained model in unseen environments with unseen object instances. The answer was clear: as they increased the amount of pre-training data and model size, real-robot success rate rose steadily and predictably. A stronger pre-trained model yields better real-robot performance, and the scaling gains show no signs of saturation.
How Xiaomi Broke the Data Bottleneck
The core insight behind Xiaomi-Robotics-1 is that robotics has missed out on the scaling laws that drove progress in language and vision models. Foundation models in those domains keep pushing the frontier by riding empirical scaling laws: capability tracks data, parameters, and compute. Robotics has been stuck because large-scale, high-quality data is hard to come by.
Xiaomi's team tackled this by using embodiment-free pre-training with UMI (Universal Manipulation Interface) data. They collected 100,000 hours of trajectories spanning more than 1,700 scenarios — household, commercial, industrial, and outdoor spaces. They built a scalable auto-labeling pipeline that divides trajectories into fixed-length segments and annotates each segment with language descriptions of scene state transitions.
For post-training, they used cross-embodiment datasets containing in-house robot data, filtered open-source robot data, and a set of high-quality UMI data. The in-house data alone covered over 7,200 hours of real-robot data collected in real homes, covering tasks like tidying a sofa, sorting a shoe cabinet, and putting away kitchenware.
Real-World Performance That Scales
The team evaluated the post-trained model in unseen environments with unseen object instances. The results were clear: as they increased the amount of pre-training data and model size, real-robot success rate rose steadily and predictably. A stronger pre-trained model yields better real-robot performance, and the scaling gains show no signs of saturation.
Xiaomi-Robotics-1 can learn new tasks with high data efficiency. The model picks up tasks like phone packing, printer refilling, laundry loading, and box packing from just a few hours of real-robot demonstrations per task. With an average of under 10 hours of demonstrations per task, it already reaches a 75% overall success rate, nearly doubling the π0 baseline (40%) at the same budget. Raising the budget to an average of under 40 hours lifts overall success to 85%.
The model also achieved state-of-the-art results on four mainstream simulation benchmarks, showing that the generalization and scaling gains carry over to standard simulation evaluation.
What This Means for the Robotics Industry
Xiaomi-Robotics-1 demonstrates a practical path for scaling robot foundation models. The key insight is that large-scale embodiment-free UMI pre-training breaks the robot data bottleneck, while real-robot and instruction alignment transfer that general capability to physical robots.
The model scales neatly with data volume and model size during pre-training, and this scaling behavior translates directly to post-training, where a stronger pre-trained model yields better out-of-the-box real-robot performance in unseen environments. The resulting foundation model adapts to new tasks from minimal data and achieves state-of-the-art performance on four challenging simulation benchmarks that emphasize generalization.
This is a big deal for the robotics industry. Until now, most robot learning systems were trained on narrow datasets collected in controlled lab environments. Xiaomi's approach — using 100,000 hours of embodiment-free data collected across 1,700 real-world scenarios — shows that the data bottleneck that has held back robot learning can be broken. The model's ability to generalize to unseen environments and objects, and to learn new tasks from just a few hours of demonstrations, points toward a future where robots can be deployed in homes and workplaces without months of custom programming.
What Makes Xiaomi-Robotics-1 Different
The model follows the training paradigm of large language models: pre-training for breadth, then post-training for alignment. Pre-training exposes the model to as much of the real world as possible using embodiment-free UMI data. At this scale, manual labeling is infeasible, so the team built an automatic annotation pipeline powered by a strong vision-language model. Long videos are split into fixed-length clips, and the VLM describes the state transition of grippers and interacting objects within each clip.
Post-training aligns the strong action-generation capabilities acquired from pre-training with real robot embodiments and natural-language instruction following. Embodiment alignment uses high-quality cross-embodiment real-robot data to map the general action-generation ability onto actual robots. Instruction alignment shifts the model from "generating actions given a description of scene state transitions" to "understanding a natural-language instruction and executing it directly."
The model achieved state-of-the-art results on four mainstream simulation benchmarks, including CALVIN, LIBERO, and MetaWorld. On CALVIN, Xiaomi-Robotics-1 scored 88.7% — a 12% relative gain over the second-best model. On LIBERO-10, it hit 94.2%, and on MetaWorld, 91.5%. These results show that the generalization and scaling gains carry over to standard simulation evaluation.
What Comes Next
Xiaomi-Robotics-1 demonstrates a practical path for scaling robot foundation models: large-scale embodiment-free UMI pre-training breaks the robot data bottleneck, while real-robot and instruction alignment transfer that general capability to physical robots. The model scales neatly with data volume and model size during pre-training, and this scaling behavior translates directly to post-training.
The paper's authors — a team of researchers from Xiaomi Robotics — note that the scaling gains show no signs of saturation. The real-robot success rate after post-training keeps improving as the model consumes more data or scales up during pre-training. This suggests that the approach could continue to yield improvements with even larger datasets and models.
For the broader robotics industry, Xiaomi-Robotics-1 represents a potential inflection point. If the data bottleneck that has held back robot learning can be broken by embodiment-free pre-training, then the same scaling dynamics that transformed language AI could soon transform physical robotics. Companies like Tesla, Boston Dynamics, and Agility Robotics have all been working on general-purpose robot models, but Xiaomi's approach — using 100,000 hours of real-world data — sets a new bar for what's possible.
The paper is available on arXiv under ID 2607.15330, and the model weights and training pipeline are expected to be released to the research community. For more on the latest in robotics and automation, check out our Robotics and Drones coverage.