Xiaomi Launches XR-1 Robotics Foundation Model Trained on 100K+ Hours
121507090301 · reddit · 2026-08-06
Xiaomi Releases Xiaomi-Robotics-1 (XR-1)
Xiaomi has released XR-1, a Vision-Language-Action (VLA) model designed for out-of-the-box mobile manipulation and efficient adaptation to unseen environments and new tasks.
Core Features
- Massive Real-World Data: Trained on over 100K hours of real-world manipulation trajectories.
- Two-Stage Training Paradigm: Inspired by LLMs, it uses pre-training for breadth followed by post-training for alignment. The pre-training scaling behavior reliably transfers to real-world robot performance with no signs of saturation.
- Hybrid Architecture: Couples a pre-trained VLM (Qwen3-VL) with a Diffusion-Transformer (DiT) via a Mixture-of-Transformers (MoT). The DiT matches the VLM in layer count but uses a smaller hidden size for faster inference.
Related event: Xiaomi Releases and Open-Sources Robotics Foundation Model XR-1(5 posts)→
More from Embodied
- Minimax video model runs locally: 4080 generates 1280x720 in ~30 mins — cocktailpeanut · 2026-08-06
- Developer working on reading experience for book scanner — aaraalto · 2026-08-06
- Three New Low-Cost Robots Unveiled Starting at $1,688 — Scobleizer · 2026-08-06
- NVIDIA Explains Cosmos World Models: Teaching Robots to Dream — nvidia · 2026-08-06
- Gemini Robotics 2 Tested: Successfully Finds Two Yellow Rubber Ducks — bousmalis · 2026-08-06
- US Faces Shortage of Over 340k Robotics Roles Amid Talent Crunch — LexSokolin · 2026-08-06