Maps representations in SLAM problem for simple environments Ref.No. : SSTCAC22025

Maps representations in SLAM problem for simple environments

Research Field: Artificial intelligence

Research Content:

This project deals with the problem of localization and mapping in dynamic unknown environments with hierarchical and compositional objects. Previously proposed solutions for map representation, like point clouds, meshes, occupancy grids, NeRF, and others while having benefits still lack the necessary flexibility, adaptability, and resource efficiency.

Thus, we need to search for an appropriate data structure that deals with both geometric and semantic information in the dynamic environment and efficiently allocates available memory and computational resources.

To simplify the problem, we sidestep from a complex 3d environment to simpler with fewer dimensions, but complex enough, environments. The main idea is that this simplicity would promote a faster search for new algorithms and data structures for map representation. Furthermore, even a 1d binary environment is suitable to implement and test ideas from biological localization and visual navigation. We hypothesize, that learning-based solutions in simpler environments will give fruitful insights and algorithms for a real-life complex 3D environment.

Research Progress:

Artificial 1D environments were created with dynamic objects and the applicability of exciting approaches were analyzed. The main problems are formulated, and the work on finding the appropriate solution is ongoing. Many ideas were explored, like lossy compression, pattern matching in compressed representations, suffix trees, action-perception cycle from neuroscience, minimal description length principle from algorithmic information theory, POMDP, neural representations, and more, but it is still hard to develop an appropriate algorithm that is not purely hand-crafted but incorporates learning.

Cooperation Needs:

Research assistance, sharing of experience, and cooperation. Discussions and constructive feedback will be beneficial for a project.

Benefits:

The project will provide a better understanding and new directions of research on the mapping problem for visual navigation. Optimistically, the proposed solutions will be suitable for a 3D environment with little or no upgrading, and bring closer the solution to the navigation problem, like in self-driving cars, or autonomous drone flights.