机器之心 (微信公众号)T2·Models
72 pts
Researchers from Fudan University and Shanghai Innovation Institute propose F2G-Pose, which performs direct RGB-D category-level 6D pose estimation in a single forward pass, jointly predicting object pose, size, and full 3D shape without per-instance CAD models or reference images. Accepted to NeurIPS 2026 with code released, it offers a more direct spatial perception route for robotic grasping and manipulation.