← Back to feed

Xiaomi released Xiaomi-Robotics-1 on July 15, a Vision-Language-Action model built around a single thesis: the data problem in robotics can be solved the same way it was solved in language. You scale pre-training on broad, diverse data, then specialise with targeted fine-tuning.

The obstacle has always been that robot training data is hard to collect in volume. Language models scaled by ingesting text from across the internet. Video models scaled on YouTube. Robotics had no equivalent source. The result was policy models capped at what sparse, expensive teleoperation datasets could support.

Xiaomi-Robotics-1 attacks this with embodiment-free pre-training.

Two-Stage Training

Pre-training is about breadth. The model trains on 100,000 hours of UMI (Universal Manipulation Interface) trajectories spanning more than 1,700 scenarios: household, commercial premises, industrial sites, and outdoor spaces. An automated labeling pipeline segments each trajectory into fixed-length chunks and annotates each segment with language descriptions of scene-state transitions. No robot is involved at this stage. The model learns general representations for action generation from human video demonstrations.

Post-training is about precision. The pre-trained model is fine-tuned on cross-embodiment datasets that include over 7,200 hours of real-robot data collected in actual homes, covering tasks like tidying sofas, sorting shoe cabinets, and putting away kitchenware. This data receives manual annotation — temporal segments and instruction prompts — at higher quality than the auto-labeled pre-training data. The post-training stage aligns the broad pre-trained representations to specific robot embodiments and instruction-following.

The Scaling Question

The core question driving the paper: does capability improve predictably as training data increases for robot policies, the same way it does for language models?

Language and vision models have followed empirical scaling laws for years. Robotics has not had enough high-quality data to test whether the same holds. By pre-training at 100K-hour scale with diverse scenario coverage, Xiaomi-Robotics-1 is positioned to characterize the scaling curve — the relationship between data volume and capability — in a way that smaller datasets cannot.

Context

This is the successor to Xiaomi-Robotics-0 (February 2026), an open-source VLA model focused on real-time inference on consumer hardware. Robotics-1 is substantially larger in data scale and more research-oriented.

The 100,000-hour pre-training figure is the largest reported for a VLA model. For comparison, LingBot-VLA 2.0 filtered 90,000 hours down to 50,000 for its training run. Xiaomi-Robotics-1 takes the opposite approach: start large, use the broad distribution, and rely on post-training to add precision.

Code and weights are available on GitHub under XiaomiRobotics. The project page is at robotics.xiaomi.com.

Key Numbers

• Pre-training: 100,000 hours of UMI trajectories, 1,700+ scenarios
• Post-training: 7,200+ hours real-robot data collected in real homes
• Tasks: sofa tidying, shoe cabinet sorting, kitchenware, and more
• Released: July 15, 2026
• Predecessor: Xiaomi-Robotics-0 (Feb 2026), 587 GitHub stars, Apache 2.0