ByteDance Research GR-2

2024-10-08 · Publisher : ByteDance Research · Domain : Robotics

ByteDance Research unveils GR-2, a robot agent pretrained on 38 million web video clips to learn world dynamics, then fine-tuned to generate both video and actions.

### Background On October 8, 2024, ByteDance Research, the research arm of TikTok's parent company, released GR-2, a generative video-language-action model for robot manipulation. The idea: first learn how the world evolves by watching huge amounts of internet video, then learn to act. ### Specifications - **Pretraining**: 38 million video clips and over 50 billion tokens from the web. - **Fine-tuning**: training on robot trajectories to generate both future video and actions. - **Results**: 97.7% average success rate across more than 100 tasks (scores published by the authors), with good generalization to unseen backgrounds, environments, objects and tasks. - **Scaling**: the authors observe that performance improves with model size. ### Why it matters GR-2 illustrates an alternative to the classic VLA: instead of starting from a vision-language model, it relies on video generation to acquire physical intuition. This approach foreshadows the world-action models that multiplied in 2025 and 2026. ### Key takeaway ByteDance established itself in 2024 as a serious player in robot models; GR-3 followed in July 2025.

Tags : VLA, Vidéo générative, Manipulation, ByteDance