| license: apache-2.0 | |
| pipeline_tag: robotics | |
| # ReMAP-DP: Reprojected Multi-view Aligned PointMaps for Diffusion Policy | |
| This repository contains the weights for **ReMAP-DP**, as presented in the paper [ReMAP-DP: Reprojected Multi-view Aligned PointMaps for Diffusion Policy](https://arxiv.org/abs/2603.14977). | |
| **Project Page:** [https://icr-lab.github.io/ReMAP-DP/](https://icr-lab.github.io/ReMAP-DP/) | |
| ## Description | |
| ReMAP-DP is a novel framework synergizing standardized perspective reprojection with a structure-aware dual-stream diffusion policy. By coupling re-projected views with pixel-aligned PointMaps, the architecture leverages learnable modality embeddings to fuse frozen semantic features and explicit geometric descriptors. This ensures precise implicit patch-level alignment, bridging the gap between 2D semantic reasoning and 3D spatial awareness for high-precision robotic manipulation tasks. | |
| ## Citation | |
| ```bibtex | |
| @misc{yang2026remapdpreprojectedmultiviewaligned, | |
| title={ReMAP-DP: Reprojected Multi-view Aligned PointMaps for Diffusion Policy}, | |
| author={Xinzhang Yang and Renjun Wu and Jinyan Liu and Xuesong Li}, | |
| year={2026}, | |
| eprint={2603.14977}, | |
| archivePrefix={arXiv}, | |
| primaryClass={cs.RO}, | |
| url={https://arxiv.org/abs/2603.14977}, | |
| } | |
| ``` |