PRBonn/RAP

What it solves

RAP addresses the challenge of 3D point cloud registration—the process of aligning multiple unposed point clouds into a single common coordinate frame. This is a critical step for tasks like 3D reconstruction, robot localization, and multi-session map merging, especially in scenarios with low overlap between scans or when using different sensor modalities.

How it works

Instead of the traditional approach of matching individual points (correspondence matching) to estimate transformations between pairs of clouds, RAP treats registration as a conditional generation problem. It uses a learned continuous, point-wise velocity field (via flow matching) to transport noisy points into a registered scene. The system employs a lightweight local feature extractor and enforces rigidity during test-time to recover the final pose of each view.

Who it’s for

This tool is designed for researchers and engineers working in robotics, 3D computer vision, and SLAM (Simultaneous Localization and Mapping), specifically those needing to merge maps from multiple robots or relocalize agents in 3D environments.

Highlights

• Generative Approach: Directly generates the registered point cloud rather than relying on pairwise transformation optimization.
• Robustness: Achieves state-of-the-art results on benchmarks, particularly in low-overlap situations.
• Generalization: Works across different scales and sensor modalities.
• Broad Application: Supports downstream tasks including relocalization and multi-robot SLAM.