Map Imagination Like Blind Humans: Group Diffusion Model for Robotic Map Generation

TL;DR

This work tackles the challenge of building large-scale 3D point cloud maps from severely limited perception, inspired by blind humans' mental cartography. It introduces a Group Diffusion Model (GDM) that partitions a map into groups and applies diffusion-denoising within each group, using a two-stage process: Stage 1 generates central points from path data, and Stage 2 performs group-wise denoising to yield a detailed map. The method yields reasonable maps from path data alone and further benefits from sparse LiDAR cues, significantly reducing sensor dependency compared to traditional LiDAR/vision-based approaches. Practically, this enables robots to imagine and generate basic maps with minimal onboard sensing, potentially supporting navigation and planning in sensor-constrained scenarios. The approach combines diffusion theory with a sparse Unet backbone and demonstrates robustness across shapes and large-scale extents.

Abstract

Can robots imagine or generate maps like humans do, especially when only limited information can be perceived like blind people? To address this challenging task, we propose a novel group diffusion model (GDM) based architecture for robots to generate point cloud maps with very limited input information.Inspired from the blind humans' natural capability of imagining or generating mental maps, the proposed method can generate maps without visual perception data or depth data. With additional limited super-sparse spatial positioning data, like the extra contact-based positioning information the blind individuals can obtain, the map generation quality can be improved even more.Experiments on public datasets are conducted, and the results indicate that our method can generate reasonable maps solely based on path data, and produce even more refined maps upon incorporating exiguous LiDAR data.Compared to conventional mapping approaches, our novel method significantly mitigates sensor dependency, enabling the robots to imagine and generate elementary maps without heavy onboard sensory devices.

Figures (6)

• Figure 1: The architecture of our proposed two-stage map generation method. While stage 1 generate central points $C'$ and noisy map $P_T$, stage 2 employ the denoising process to generate large scale map $P_0$.
• Figure 2: The group diffusion model works by dividing the original map into several groups. The diffusion process and the denoising process are separately applied to these group points. We add a certain amount of spacing between each group in this picture, while actual map is presented at the top of the picture without these spacing intervals.
• Figure 3: Generating map from a path and limited LiDAR points. Given both path data and limited LiDAR points, we first estimate their normals and width $w$, and then generate one point per meter along these normals, extending up to a distance of $w$ meters. Finally, we employ the diffusion process to add noise to the points in order to obtain $P_T$ and then we utilize the denoising process to obtain a detailed map $P_0$.
• Figure 4: Comparason of three map generation modes using our proposed two-stage map generation architecture. The length of this slected path is about 6.4km, and the height range is [0, 78m]. The key difference among the three modes lies in their input data: Mode 1 utilizes solely the path, Mode 2 incorporates both the path and a random width $w$, while Mode 3 employs the path alongside exiguous sampled LiDAR point clouds.
• Figure 5: The error distances color scale of the generated map compared to ground truth. We compare three types of map sequences (Seq) using the Mode 3. Seq I has the minimal error since it contains fewer outlier points.