3DWay-15B

This repository contains the 15B model checkpoint of 3DWay: Generalizing Robot Manipulation via 3D Consistent Waypoints, accepted at ECCV 2026.

This model is introduced in the paper 3DWay: Generalizing Robot Manipulation via 3D Consistent Waypoints. Code: https://github.com/ziqin-h/3DWay.

Overview

3DWay introduces 3D Consistent Waypoints, an explicit and actionable trajectory representation for robot manipulation. Given multi-view RGB observations and a language instruction, the model predicts multi-view consistent 2D waypoints, which are geometrically triangulated into 3D trajectories.

The resulting trajectories can be directly executed for manipulation tasks or used as geometric priors for foundation Vision-Language-Action (VLA) models.

Project Page

https://ziqin-h.github.io/3DWay/

Citation

@inproceedings{3dway2026,
  title     = {3DWay: Generalizing Robot Manipulation via 3D Consistent Waypoints},
  author    = {Huang, Ziqin and Li, Yingyue and Zhang, Chenyangguang and Zhang, Ruida and Chen, Yuxin and Wang, Gu and Liu, Xingyu and Tomizuka, Masayoshi and Ji, Xiangyang},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}