[Paper Review] SGoLAM: Simultaneous Goal Localization and Mapping for Multi-Object Goal Navigation

Junho Kim, Eun Sun Lee+3 more|arXiv (Cornell University)|Oct 14, 2021

Multimodal Machine Learning Applications32 references4 citations

TL;DR

SGoLAM proposes a training-free, modular approach for multi-object goal navigation using RGB-D and GPS/Compass sensors, combining classical projective geometry for simultaneous goal localization and occupancy mapping. It achieves state-of-the-art performance on par with learning-based methods, ranking 2nd in the CVPR 2021 MultiON challenge.

ABSTRACT

We present SGoLAM, short for simultaneous goal localization and mapping, which is a simple and efficient algorithm for Multi-Object Goal navigation. Given an agent equipped with an RGB-D camera and a GPS/Compass sensor, our objective is to have the agent navigate to a sequence of target objects in realistic 3D environments. Our pipeline fully leverages the strength of classical approaches for visual navigation, by decomposing the problem into two key components: mapping and goal localization. The mapping module converts the depth observations into an occupancy map, and the goal localization module marks the locations of goal objects. The agent's policy is determined using the information provided by the two modules: if a current goal is found, plan towards the goal and otherwise, perform exploration. As our approach does not require any training of neural networks, it could be used in an off-the-shelf manner, and amenable for fast generalization in new, unseen environments. Nonetheless, our approach performs on par with the state-of-the-art learning-based approaches. SGoLAM is ranked 2nd in the CVPR 2021 MultiON (Multi-Object Goal Navigation) challenge. We have made our code publicly available at \emph{https://github.com/eunsunlee/SGoLAM}.

Motivation & Objective

• To address the challenge of navigating to a sequence of target objects in 3D environments without requiring training on new environments.
• To improve generalization and transferability in unseen environments by avoiding end-to-end deep learning.
• To decompose multi-object navigation into modular components: mapping and goal localization, leveraging classical geometry.
• To achieve competitive performance against learning-based methods while maintaining fast deployment and adaptability.

Proposed method

• The mapping module converts depth observations into a 3D occupancy map using projective geometry, maintaining spatial consistency across frames.
• The goal localization module identifies and marks the 2D positions of target objects in the top-down map using RGB-D features and object detection.
• The agent’s policy selects actions based on a hybrid strategy: if a goal is detected, it plans directly toward it; otherwise, it performs exploration.
• The system uses a top-down grid representation for both map and goal locations, enabling efficient planning with classical pathfinding algorithms.
• All components are trained or optimized independently using classical computer vision and geometry—no neural network training is required.
• The method is modular and reusable, allowing plug-and-play deployment in new environments without retraining.

Experimental results

Research questions

• RQ1 Can a classical, non-learning-based approach achieve performance competitive with state-of-the-art learning-based methods in multi-object goal navigation?
• RQ2 How does a modular decomposition of mapping and goal localization improve generalization in unseen 3D environments?
• RQ3 To what extent can training-free navigation systems outperform learning-based baselines in success and progress metrics?
• RQ4 What is the trade-off between success rate and path efficiency in classical vs. learning-based navigation for multi-object tasks?

Key findings

• SGoLAM achieved a success rate of 0.62, significantly outperforming all baselines including NoMap (RNN) at 0.05 and VisMemoryMap at 0.43.
• The method achieved a progress metric of 0.71, surpassing the next best baseline (AuxTaskMap) at 0.70.
• In terms of path efficiency, SGoLAM scored 0.39 on PPL and 0.34 on SPL, ranking below the best learning-based method (AuxTaskMap) but still highly competitive.
• Despite not using any neural network training, SGoLAM demonstrated strong generalization and fast deployment in unseen environments.
• The results confirm that classical, geometry-based approaches can achieve high success rates in complex, long-horizon navigation tasks.
• The performance gap in path efficiency highlights a known limitation of classical planners, but SGoLAM remains highly effective overall.