Deep Reinforcement Learning for Autonomous Navigation of Mobile Robots in Indoor Environments
Abstract
Conventional autonomous navigation framework for mobile robots is highly modularized with
various subsystems such as localization, perception, mapping, planning and control. Although these provide easy interpretation, they are highly dependent on a known map of the robot’s surroundings for navigating in a cluttered environment. Local planners such as DWA require a map with all obstacles in the surroundings to calculate an optimal collision-free trajectory to the goal. Planning and tracking a collision-free path without knowing the obstacle locations is a challenging task.
Since the advent of deep learning techniques, the field of deep reinforcement learning has
proven to be a powerful learning framework for robotic tasks. Deep Reinforcement Learning
has demonstrated wide success in various complex computer games such as Go and StarCraft
which have high dimensional state and action spaces. However, it has rarely been used in real
world applications due to the Sim-2-Real challenges in transferring the trained RL policy into the
real-world.
In this work, we propose a novel framework for autonomously navigating a mobile robot
in a cluttered space without known localization of the obstacles in its surroundings using deep
reinforcement learning techniques. The proposed method is a modular and scalable approach due to a strategic design of the training environment. It uses constrained space and randomization techniques to learn an effective reinforcement learning policy in lesser simulation training time. The state vector consists of the target location in the mobile robot coordinate frame and additionally a 36-dimensional lidar vector for obstacle avoidance task. We demonstrate the optimal discrete action policy on a Turtlebot in the real-world. We have also addressed some key challenges in robot pose estimation for autonomous driving tasks.
Citation
Vaidya, Gargi Yatin (2021). Deep Reinforcement Learning for Autonomous Navigation of Mobile Robots in Indoor Environments. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /196085.