SLAMOps: SLAM as an MLOps

Date 1
Nov. 2021 - on going

Purpose of the project

Inspired by the concept of MLOps, design the concept of SLAMOps to a level where the entire process of SLAM can be modularized/automated and applied to real services.
Phase 0: SLAM on Docker
Docker containerization to run SLAM algorithm in virtualized environment
Phase 1: Hyper-parameter tuning
Configure SLAM mapping hyper-parameter tuning pipeline using Katib in Kubeflow
Phase 2: Experiment Pipeline
Develop many ideas for SLAM and run them in isolated Docker containers
Logs output as tasks are executed, outputs saved when tasks are completed, and
Modularized to ensure results are reproducible and reusable
Phase 3: Continuous Mapping
Consider the maps generated as a result of SLAM as one deep learning model
The world we live in is constantly changing, so maps need to be constantly updated.
We envision a concept of continuous mapping that includes concepts such as map update, fleet management, etc.
Phase 4: Localization Serving
Localization, which identifies one's location on a mapped map, is similar to the task of performing inference on a model trained in deep learning.
As the number of robots to be managed increases, it is necessary to introduce the concept of large-scale distributed processing to stably provide localization services.

Contributions

Studying MLOps
Watching MLOps lectures by Infrun's Ho Yeon Song
Watching MLOps lectures by Jaeyeon Kim from Fast Campus
Participated in the MLOps KR online conference
Initial conceptual design of SLAMOps
Phase 1: Hyper-parameter tuning
Implementing metrics to evaluate the quality of the maps resulting from SLAM