Volumetric Occupancy Mapping With Probabilistic Depth Completion for Robotic Navigation

TL;DR

This work tackles safe robotic navigation in unknown cluttered indoor environments where commodity RGB-D sensors produce missing depth data on shiny or distant surfaces. It introduces probabilistic depth completion that jointly predicts dense depth and per-pixel uncertainty from RGB-D inputs and feeds these into an online occupancy mapping pipeline. A two-decoder network learns both depth and uncertainty, using surface normals and occlusion boundaries, and is trained with a Bayesian-inspired loss; its outputs are integrated into a multiresolution occupancy framework to improve free-space estimation. Experiments on synthetic interiors and real-world RGB-D data show that combining completed depth with reliable uncertainty yields more complete free-space maps and faster exploration without compromising reconstruction safety, enhancing robotic navigation capabilities in unknown environments.

Abstract

In robotic applications, a key requirement for safe and efficient motion planning is the ability to map obstacle-free space in unknown, cluttered 3D environments. However, commodity-grade RGB-D cameras commonly used for sensing fail to register valid depth values on shiny, glossy, bright, or distant surfaces, leading to missing data in the map. To address this issue, we propose a framework leveraging probabilistic depth completion as an additional input for spatial mapping. We introduce a deep learning architecture providing uncertainty estimates for the depth completion of RGB-D images. Our pipeline exploits the inferred missing depth values and depth uncertainty to complement raw depth images and improve the speed and quality of free space mapping. Evaluations on synthetic data show that our approach maps significantly more correct free space with relatively low error when compared against using raw data alone in different indoor environments; thereby producing more complete maps that can be directly used for robotic navigation tasks. The performance of our framework is validated using real-world data.

Figures (10)

• Figure 1: Overview of our approach for mapping with depth completion. Top: Our depth completion network takes raw RGB-D images to predict the completed depth and depth uncertainty, which are used as additional inputs for probabilistic 3D mapping. Bottom: By leveraging the network completions to complement the original raw depth data (right), we obtain more complete maps of free space when compared against using the raw depth alone (left). Darker shades of blue indicate areas of lower occupancy probability.
• Figure 2: Overview of our proposed approach. We leverage a network for depth completion with uncertainty to improve the input for probabilistic tracking and mapping.
• Figure 3: Our probabilistic depth completion system pipeline including the training framework with different training loss components (\ref{['SS:loss_function']}). Given an input RGB-D image, we predict surface normals and boundaries and pass them to the probabilistic depth completion network (\ref{['F:architecture']}) to predict depth and associated uncertainty. Black in the depth images indicates missing information.
• Figure 4: Our architecture for depth completion with uncertainty. We extend the network of Huang2019 with a second output decoder for uncertainty prediction. Our network takes as input raw RGB-D, surface normals and boundaries (left), and outputs completed depth (top-right) and pixelwise uncertainty (bottom-right).
• Figure 5: (a) Inverse sensor model for fusing new data into a map. Occupancy probability as a function of the difference $d_r$ from a query point to depth measured along a ray. (b) Measurement uncertainty model for mapping with raw depth in supereight 'MultiresOFusion' Funk_arXiv2020. Standard deviation $\sigma$ given depth measurement $z$.