Action-Geometry Prediction with 3D Geometric Prior for Bimanual Manipulation
Nov 1, 2025·
,,,,,·
0 min read
Chongyang Xu
Haipeng Li
Cheng Shen
Haoqiang Fan
Ziliang Feng
Shuaicheng Liu
Abstract
By building upon pre-trained 3D geometric foundation models like VGGT or pi3, this work unifies geometry-aware latents with semantic features to jointly predict future action sequences and 3D scene evolutions, enabling robust, coordination-aware bimanual manipulation directly from RGB observations.
Type
Publication
Submitted to CVPR 2026
Authors
Ph.D. Student @ Sichuan University
Embodied AI Intern @ Tongyi Robotics
Embodied AI Intern @ Tongyi Robotics
Hello! 👋
I’m Chongyang, a researcher who’s into physical AI & robotics, equally passionate about sports, music, humanities, and sociology. I’m doing multimodal learning and reinforcement learning in the grandest simulator of all: life — one episode at a time, learning what’s worth the strife.
I’m openly seeking collaborations — if you have any research ideas or projects, feel free to reach out!
Education 🎓
I’ve been studying at Sichuan University for 7 years and have fallen deeply in love with Chengdu. I received my B.Eng. in Software Engineering and am now pursuing my Ph.D. in Computer Science.