Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories

🚀 Xiaomi-Robotics-1 Is Here: Scaling Robot Learning with Over 100,000 Hours of Real-World Manipulation Data

Foundation models in language and vision have advanced through scaling—using more data, larger models, and greater compute. Robotics, however, still faces a fundamental challenge: large-scale, diverse, and high-quality manipulation data is difficult and expensive to collect.

Today, we introduce Xiaomi-Robotics-1, a vision-language-action foundation model designed to study and unlock scaling in robot learning. It combines large-scale embodiment-free pre-training, real-robot embodiment alignment, and natural-language instruction alignment, enabling the model to acquire broad manipulation knowledge and transfer it efficiently to physical robots.

🌍 From Limited Robot Data to Over 100,000 Hours of Real-World Experience

Xiaomi-Robotics-1 is pre-trained on more than 100,000 hours of real-world UMI manipulation trajectories, covering over 1,700 scenarios across homes, commercial spaces, industrial sites, and outdoor environments.

Because these trajectories are collected without depending on a specific robot embodiment, the model can learn from a much broader range of objects, environments, and manipulation behaviors than would be practical using robot-collected data alone.

Our experiments show a clear scaling trend:

📉 Increasing either the amount of pre-training data or the model size consistently reduces action prediction error.

🧠 From Raw Manipulation Trajectories to Language-Guided Actions

Manually annotating more than 100,000 hours of data is impractical. To address this, we build a scalable VLM-based automatic annotation pipeline.

Each long manipulation trajectory is divided into shorter clips. A vision-language model then describes how the grippers and objects change within each clip. Xiaomi-Robotics-1 is trained to generate actions that produce the state transition described by the language instruction.

This allows large amounts of unstructured manipulation data to be converted into training examples that connect visual observations, language goals, and actions.

🤖 From Embodiment-Free Knowledge to Real-Robot Skills

After pre-training, Xiaomi-Robotics-1 is aligned with physical robots using cross-embodiment robot datasets, including more than 7,200 hours of in-house real-robot data collected in real homes.

The model also undergoes instruction alignment, allowing it to follow natural-language commands rather than relying only on descriptions of state transitions.

After alignment, Xiaomi-Robotics-1 can directly perform a wide range of mobile manipulation tasks, including tasks involving previously unseen environments and object instances.

Most importantly, the benefits of large-scale pre-training transfer to physical robots:

📈 Models trained with more pre-training data or larger model sizes achieve higher real-robot success rates, with no clear sign of performance saturation.

⚡ Adapting to New Tasks with Only a Few Hours of Demonstrations

Xiaomi-Robotics-1 can be adapted to complex downstream tasks using only a small amount of task-specific robot data.

We evaluate it on four real-world tasks: phone packing, printer refilling, laundry loading, and box packing.

With an average of fewer than 10 hours of demonstrations per task, Xiaomi-Robotics-1 achieves:

📈 75% overall success rate, compared with 40% for π0.5 under the same data budget.

When the training data is increased to an average of fewer than 40 hours per task, the overall success rate reaches:

📈 85%

These results show that large-scale pre-training substantially improves data efficiency when learning new robot skills.

🏆 State-of-the-Art Results Across Four Simulation Benchmarks

We further evaluate Xiaomi-Robotics-1 on four challenging simulation benchmarks:

📈 RoboCasa: 74.5% average success rate
📈 RoboCasa365: 57.4% average success rate
📈 VLABench: 59.1% average success rate
📈 RoboDojo: 13.93% average success rate

Xiaomi-Robotics-1 achieves the best reported performance across all four benchmarks, demonstrating strong generalization across different environments, tasks, and evaluation settings.

Xiaomi-Robotics-1 shows that large-scale embodiment-free manipulation data can serve as an effective foundation for general-purpose robot learning.

By learning from broad real-world experience, aligning with physical robot embodiments, and adapting to new tasks with limited demonstrations, Xiaomi-Robotics-1 takes an important step toward robot foundation models that improve with scale and generalize across tasks, environments, and embodiments. ✨