Deep Reinforcement Learning for Adaptive Exploration of Unknown Environments

TL;DR

This paper introduces an adaptive deep reinforcement learning method for UAVs that efficiently balances exploration and exploitation in unknown environments, enabling more effective search for areas of interest without prior maps.

Contribution

It develops a unified DRL-based approach with environment map segmentation and extended algorithms to improve autonomous exploration and target search in unknown settings.

Findings

Outperforms baselines in environment coverage

Navigates efficiently in random environments

Covers more areas of interest in fewer steps

Abstract

Performing autonomous exploration is essential for unmanned aerial vehicles (UAVs) operating in unknown environments. Often, these missions start with building a map for the environment via pure exploration and subsequently using (i.e. exploiting) the generated map for downstream navigation tasks. Accomplishing these navigation tasks in two separate steps is not always possible or even disadvantageous for UAVs deployed in outdoor and dynamically changing environments. Current exploration approaches either use a priori human-generated maps or use heuristics such as frontier-based exploration. Other approaches use learning but focus only on learning policies for specific tasks by either using sample inefficient random exploration or by making impractical assumptions about full map availability. In this paper, we develop an adaptive exploration approach to trade off between exploration and exploitation in one single step for UAVs searching for areas of interest (AoIs) in unknown environments using Deep Reinforcement Learning (DRL). The proposed approach uses a map segmentation technique to decompose the environment map into smaller, tractable maps. Then, a simple information gain function is repeatedly computed to determine the best target region to search during each iteration of the process. DDQN and A2C algorithms are extended with a stack of LSTM layers and trained to generate optimal policies for the exploration and exploitation, respectively. We tested our approach in 3 different tasks against 4 baselines. The results demonstrate that our proposed approach is capable of navigating through randomly generated environments and covering more AoI in less time steps compared to the baselines.