Active-SpCoSLAM

Implementation of Active-SpCoSLAM that is the active semantic mapping method.
This repository includes the source codes used for the experiments in our paper on SMC 2023.

Abstract of Active-SpCoSLAM

We proposed Active-SpCoSLAM, which simultaneously learns spatial concepts and maps based on information gain (IG). IG is derived from the following graphical model: IG is calculated separately for IG related to simultaneous localization and mapping (SLAM), IG related to location concepts, and finally a weighted sum. The mapping between language and location is then done using an image captioning model called ClipCap.

Execution environment

• Ubuntu 20.04
• Python 3.8.10
• ROS noetic
• Image captioning system: ClipCap

Preparation

• Please refer to the file external_packages.txt to download packages of gmapping, A* algorithm and simulation environment
• Please refer to the file requirements.txt to install requirements
• Please refer to the file /Active-SpCoSLAM/models/model_for_clipcap/download_weights.txt to download weights for ClipCap
• Please refer to the file /Active-SpCoSLAM/models/model_for_places/download_weights.txt to download weights for PlacesCNN
• Please refer to the file code_replace_procedure.txt to modify gmapping and A* algorithm package

Execution procedure

• Launch the file slam_gmapping/gmapping/launch/slam_gmapping_pr2.launch
• Launch the file Active_SpCoSLAM/scripts/pub_global_pose.py
• Launch the file Astar/launch/astar.launch
• Execute the python file /Active-SpCoSLAM/scripts/main.py