Panoptic view of ICAT platform in the real world. ICAT is 6 by 5 meters in size, and 10 intelligent robots are used in ICAT now for autonomous driving studies. Heterogeneous computing devices are equipped for Federated Machine Learning (FedML) research and multi-agent vehicle computing. 

Traffic lights and signs are now integrated seamlessly with CV2X technologies and enable interactions with traffic agents. Leveraging a down-scaled infrastructure system, multiple traffic signs and traffic lights are set up for realistic traffic simulations, where the traffic signs include speed limit, yield sign, construction zone, stop sign, and do not enter signs.

Digital twin system of ICAT. (a) ICAT's map is designed within RoadRunner, which is a 3D interactive autonomous driving system builder. (b) The designed map then can be seamlessly imported into CARLA simulations. (c) SUMO simulations are also enabled with the ICAT road system. The merits of leveraging a digital twin for our proposed ICAT platform can be summarized in three aspects. 

1.  Algorithm testing. Faster algorithm iteration can be achieved by efficient simulations. 

2. Traffic management. Managing a group of interacted traffic agents with a centralized autonomous driving algorithm in the real world hardly guarantees safety, the counterpart in a digital twin can release such a safety burden. 

3. Remote testing. With advanced visualization capabilities, the information of testing in the real world can be accurately reflected on a remote end.

An ICAT centralized multi-agent traffic management system testing in a Python simulation, where simulated 10 traffic agents could be controlled by a centralized planning algorithm to avoid collisions and improve the throughput. The planned trajectories for different robots are marked in different colors. The simulation can be transferred to a ROS-based real-world experiment without extra effort.

An illustration of ROS-based traffic management real-world experiments. (a) 4 robots are running a ROS-based traffic management demo. (b) Real-world demos can be also monitored in RViz, where the ICAT road is published as a Path message (the tilted lines are caused by the automatic connection of section start and end points in the Path message), the planned car trajectories from the central manager are rendered as different colors. (c) We have tested up to 6 robots in the ICAT real-world traffic system.

Decentralized Autonomous Driving differs from traffic management methods in that it doesn't rely on externally received command information, the traffic agents solely rely on their own sensors and computing devices to conduct perception and data processing. Thanks to heterogeneous on-board computing devices, the ICAT robots with ARM computing platforms leverage Scalable Open Architecture For the Embedded Edge (SOAFEE) and Autoware OpenAD Kit to support autonomous driving. SOAFEE and OpenAD Kit serve as a baseline for autonomous driving and enable researchers to quickly and easily modify the perception, planning, and control algorithms.