Dynamic obstacle avoidance in mobile robots using adversarial deep learning

Abstract

In modern days, getting autonomous mobile robots to work in dynamic human environments and to help people with completing tasks such as delivery is a ubiquitous issue among roboticists. The implementation of mobile robot technology has become very popular locally in recent years. YOTEL Singapore, Haidilao restaurant, Jewel Changi Airport, Free the Robot restaurant are some examples of how robots may be integrated into businesses to expedite tasks such as item delivery, surveillance, and cooking. Lately, owing to the plight of COVID-19, more robots are deployed in hospitals to carry out disinfection activity and to do telepresence for the purpose of preventing the spread of infectious disease between people. In all these implementations, an autonomous mobile robot is not able to avoid the issue of navigation in a human crowd as robots integrate into our daily lives. The problem of obstacle avoidance arises when a robot attempts path planning to generate a collision-free motion trajectory across a certain period. In the presence of other moving obstacles, there exists an element of uncertainty where the obstacle may cross the generated motion trajectory and potentially result in collision with the robot. For a robot operating in a human environment, it becomes more important for it to develop an ‘awareness’ to predict surrounding motions and preempt sudden movements so that path planning will be safer without accidents. This entails robots that are able to respond within human reaction time to do obstacle avoidance and have an understanding of human movement and intentions. Current solutions require relatively longer period and a good amount of computation power to process information of the surroundings before the robot makes a calculated move. This problem is further exacerbated with a continually changing, dynamic environment such as a human crowd. Therefore, there is a need to develop a more dynamic and robust solution to tackle navigation in a dynamic environment, taking a busy hospital setting for example, where a robot is required to deliver items while navigating among crowd of people and nurses traversing quickly. The proposed solution is to develop an adversarial deep learning-trained neural network model that can navigate to goals as far as five meters while avoiding any potential obstacles, static and dynamic, that may come in the way. A Pybullet and ROS-integrated adversarial deep learning framework is developed for training an actual robot software on obstacle avoidance tasks. A robot agent will learn based on information of the goal’s whereabouts, its own movement speed, collision sensor, and laser scanner, to output movement velocity.