I work in an intersection of AI, robotics, and space exploration. I build systems for you ヾ(^▽^*)) so feel free to ask me if you have any question!
Currently I am a PhD candidate at the University of Hong Kong supervised by Hongyang Li and co-supervised by Ping Luo.
- ⚡ Recent works: Train a generalized policy / video generation model for robotics manipulation.
- GR1-Training is the current SOTA policy on CALVIN benchmark for multi-task robotics manipulation (ABC-D split).
- MimicTest integrates some of my policies, including Florence Policy developed on Microsoft's Florence2 VLM, which is trained with VQA, OCR, detection and segmentation tasks on 900M images.
- Unlike OpenVLA and RT2, Florence2 is much smaller with 0.23B (Florence-2-base) or 0.7B (Florence-2-large) parameters.
- Florence policy directly generates continuous actions with a linear action head / a diffusion transformer action head.
- I am contributing to the HuggingFace LeRobot toolbox, including integration of the Florence policy.
- I am participating the 1x challenge to train video generation model for robotics manipulation.
- I developed some VLM ranged from 2B~8B, please check My HuggingFace Page.
- 📫 Reach me: My email or Wechat: StarRingSpace.
- 🌱 My profile: Linkedin and my personal website.