Some key features of EI-Drive include:
- Dynamic Deadline-driven Edge Computing: The edge-computing module is an innovative task-scheduling model that enables multiple applications to conduct their computation across different edge units with different computational capacities, while coping with time-varying connectivity latency. We employ the idea of dynamic deadline scheduling that can adapt to the changes of the environment and the running time of tasks associated with the computational complexity of different models, thereby designing an environment-dependent adaptive dynamic deadline model for the downstream tasks and proactively selecting proper learning models and computational edge units.
- Reinforcement Learning-based Behavior Planning Module: EI-drive integrates the traditional autonomous driving planning modules with RL-based approaches to account for the sophisticated the behavior planning problem.
- Realistic Sensing and Perception Module: Besides the sensing and perception system in CARLA and OpenCDA, EI-drive provides various ways to simulate the disturbance and latency in different sensing modes based on the empirical results. EI-drive also provides various approaches to fuse heterogeneous sensing sources and representative approaches to simplified high-dimensional perception data.
- Reinforcement Learning (RL) Friendly Support: EI-drive provides multiple RL-friendly modules to support the execution and evaluation of different RL algorithms. For example, EI-drive provides a data center module to record the trajectories of agents. All collected data can be stored in various formats and easily be queried for downstream applications such as offline RL and inverse RL. EI-drive also provides a human-control module to collect human-driven behavior trajectories for studying mixed-road driving.
Notably, EI-Drive has enabled researchers to test smart intersection scenarios with modules that represent edge-computing aided cooperative vehicle and infrastructure systems. These edge-computing modules/devices would offer additional computational capacity and enable vehicles with a more global view of dynamic traffic activities of a full intersection. Our testing results corroborate Dr. Shao's tests and prove that these edge-computing modules can improve the accuracy of drive path projections, making it safer and more efficient to operate automated vehicles than in the absence of connected infrastructure support. Connected vehicle infrastructure, such as what is modeled in the EI-Drive platform could prove useful for all automated vehicles, and also may have near-term applications in aiding emergency vehicles or other service vehicles while navigating busy intersections.