ByteDance Seed GR-3 and ByteMini
2025-07-21 · Publisher : ByteDance Seed · Domain : Robotics
ByteDance Seed releases GR-3, a large vision-language-action model that generalizes to unseen objects and instructions, and ByteMini, a two-arm mobile robot designed to run it.
What changes
- GR-3 is a large-scale VLA co-trained with web vision-language data.
- Efficient fine-tuning from human trajectories collected in VR.
- ByteMini: a bimanual mobile robot built for GR-3.
- GR-3 beats π0 on many hard tasks, per the authors.
### What changes On July 21, 2025, ByteDance's Seed team published the technical report for GR-3, successor to GR-2. This time it is a large vision-language-action (VLA) model, paired with an in-house robot, ByteMini. ### Specifications - **Generalization**: unseen objects, environments and instructions, including instructions involving abstract concepts. - **Training recipe**: co-training with web-scale vision-language data, efficient fine-tuning from human trajectories collected with VR headsets, and imitation learning on robot trajectories. - **Tasks**: long-horizon and dexterous tasks, bimanual manipulation and mobile movement. - **ByteMini**: a bimanual mobile robot designed for flexibility and reliability, running GR-3. - **Results**: GR-3 outperforms π0 on a wide range of challenging real-world tasks (results published by the authors). ### Why it matters ByteDance moves from research model to model-plus-robot, like Figure or 1X. Using VR to collect cheap human demonstrations helps reduce reliance on teleoperation. ### Key takeaway GR-3 confirms the rise of large Chinese tech groups in generalist robot models.
Links
Tags : VLA, Manipulation mobile, Bimanuel, ByteDance