Heterogeneous Multi-agent Collaboration with Reinforcement Learning

In recent times, the rapid advancements in autonomous technology have led to surging demands for effective transportation systems especially in emergency scenarios. This may be attributed to the fact that typical vehicles are usually homogeneous and have limitations such as being less adaptable to complex tasks and having restricted mobility in urgent situations. In response to this pressing challenge, this project delves into the collaboration of heterogeneous multi-agents to improve the adaptability, flexibility, robustness, and efficiency of the overall system. Compared to homogeneous agents, heterogeneous agents offer benefits such as cost saving and reduced size for each. Furthermore, they have more comprehensive perceptions attributing to the complement of their diverse features and functions, potentially enabling greater flexibility, cooperation, and resilience in tackling unexpected challenges and diverse tasks.

The proposed project focuses on exploring heterogeneous multi-agent collaboration, utilizing cloud service providers especially the Amazon Web Services (AWS). This involves researching and implementing ways to facilitate communication, data sharing, and coordination through the AWS cloud computing, Robot Operating System (ROS), and Reinforcement Learning (RL). In the specific contents of emergency cases, we will take the advantages of various capabilities of multiple intelligent agents to simultaneously achieve features such as search, rescue, and monitor. To achieve these features, we will explore and employ Amazon key services and technologies including open-source packages, RoboMaker, and IoT (Internet of Things) services. We will start by implementing these features on individual AWS small cars, and then expand the implementation to multiple AWS small cars.

On the AWS (Amazon Web Services) platform, the objectives of the heterogeneous multi-agent collaboration involve the following aspects:

• Creatively involve AWS cloud computing, robot operating System (ROS), and reinforcement learning (RL). RL can be used to train various agents to collaboratively perform tasks within an environment, and AWS can provide the computing resources needed for training and deployment. AWS, in conjunction with ROS, enables the interface with the robotic hardware and the coordination of robot behaviors within their environment. This allows for the development of intelligent and adaptive robotic systems.
•  Innovational use of advanced technologies such as lidar, depth sensors in supply of multi-dimensional information for task execution which can greatly improve the system’s perception and intelligent decision-making capabilities. In the case of autonomous driving, combination of data from different sensors or even implementation of lidar cameras on driving tracks, can provide more comprehensive environmental perception and information collection, thereby promoting the collaboration between cars and other objects and enhancing the accuracy, safety, and robustness of the autonomous driving execution.

This Dachuang project will lead to students’ signature work, multi-agent system demo, and smart applications. Some other outcomes could be patent and conference paper submissions with students serving as first authors or co-authors.

• Be familiar to ROS programming, and able to set up frame work of task allocate RL algorithm.
• Able to design simulation environment, robot communication system based on ROS; and develop task allocate RL algorithm.

1 year