SciGroveBeta
Robotics

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

Jun Guo, Piaopiao Jin

Featured July 23, 2026

Original Paper

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

This robot brain learns how to do many tasks by watching over 100,000 hours of real-world actions, using a smart system to automatically describe what's happening, making it super good at new jobs with little training.

In depth
The paper introduces Xiaomi-Robotics-1, a foundational vision-language-action model trained on an unprecedented scale of over 100,000 hours of real-world robot manipulation data. A key innovation is their auto-labeling pipeline that automatically generates natural language descriptions for trajectory segments, overcoming the data bottleneck. This enables a two-stage training process, first learning general action generation from state transitions, then aligning these capabilities to imperative human instructions and diverse robot embodiments, demonstrating strong scaling and efficient adaptation to new tasks.

Key Takeaways

  • 1
    The model is pre-trained on over 100,000 hours of real-world manipulation trajectories, enabled by a novel auto-labeling pipeline for language annotations.
  • 2
    A two-stage training process involves pre-training for general action generation from state transitions, followed by post-training to align skills to diverse robot embodiments and imperative human instructions.
  • 3
    The model exhibits robust scaling behavior, achieving state-of-the-art performance in unseen environments and efficiently adapting to novel, complex tasks with minimal fine-tuning data.

Conceptual Flow

HIGH LEVEL
1
Methodology (The "Logic")

The robot learns in two main steps: first, it watches tons of videos to understand how things change, then it learns to follow human commands using that knowledge.

Huge Video Library
Auto-Generated Descriptions
Learn General Skills
Robot Brain (General)
Human Commands
Robot Actions
2
Results (The "Impact")

By learning from so much data, the robot gets much better at new tasks and works well even in places it's never seen before.

Old Robot Brain
New Robot Brain
Compare Performance
Better Success Rate
Works in New Places
Learns Faster

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.