Vision-Based Navigation System for Autonomous Ground Robots in Dynamic Environments

Abstract

Autonomous Ground Robots (AGRs) in unstructured settings are an unresolved issue that 
necessitates an intelligent agent's capacity to identify and respond to possible impediments within its vicinity. 
Barriers include automobiles, people, or fixed items in an ordered setting, such as highway or urban navigation, 
and unexpected static and dynamic obstacles in informal settings like woodland roads. The second condition is 
often more challenging to manage, owing to its greater uncertainty. This work presents a vision-based 
dynamical methodology for route planning and guidance of a quadruped AGR operating in a disordered setting, 
especially on a forest road. The computational dynamics methodology employs a recurring neural network that 
utilizes a sensor for data acquisition, generating sequences of prior depth sensor measurements and forecasting 
upcoming measurements over a limited temporal interval. The research evaluates the methodology against 
other leading techniques in obstacle-driven path planning methods. It conducts ablation tests to assess the 
effects of architectural modifications on the model elements. This shows that the approach attains enhanced 
performance in effectively creating collision-free paths for the clever agents.