We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hacker News·16d ago·5 min read
Xiaomi-Robotics-1
TL;DR
Xiaomi-Robotics-1 is a robot foundation model trained on over 100,000 hours of diverse manipulation data, addressing the scarcity of large-scale robotics training. It combines embodiment-free pre-training with real-robot data to enhance its capabilities.
✦ Why It Matters
Engineers can implement Xiaomi-Robotics-1 to enhance robotic task performance in real-world environments today.
Key Takeaways
- Utilize large-scale UMI data for pre-training robotic models.
- Implement two-stage training for better alignment with real-world tasks.
- Leverage efficient adaptation for new tasks with minimal data.
- Monitor scaling behavior to predict real-robot performance improvements.
How It Works
Xiaomi-Robotics-1 employs a scalable auto-labeling pipeline that annotates UMI trajectories with language descriptions, allowing the model to learn action generation effectively. The pre-training phase focuses on broad exposure to diverse environments, while the post-training phase aligns the model's capabilities with real robot actions and natural language instructions.