Cooperative Adaptive Cruise Control: A Gated Recurrent Unit Approach

Embedded artificial intelligence solutions are promising controllers for future sustainable and automated road vehicles. This study presents a deep learning-based approach combined with vehicle communication technology for the design of a real-time cooperative adaptive cruise control (CACC). A particular type of recurrent neural network has been selected, namely a gated recurrent unit (GRU). GRU exhibits improved learning performance in control problems such as the CACC since it avoids the vanishing gradient problems that characterize long time series. A GRU has been trained using ad-hoc CACC datasets build-up according to an optimal control policy, i.e. dynamic programming (DP), for a battery electric vehicle. In particular, DP optimizes the longitudinal speed trajectory of the Ego (Following) vehicle in CACC so to achieve energy savings and passenger comfort improvement. Results demonstrate that the Ego vehicle controlled by the trained GRU can achieve an ecofriendly driving in CACC without compromising passenger comfort and safety requirements. Unlike DP, GRU holds strong real-time potential. The performance of the proposed GRU approach for CACC is verified by benchmarking with the optimal performance obtained off-line using DP in several driving missions.

Automated summary

This study proposes a deep learning-based approach combined with vehicle communication technology for real-time cooperative adaptive cruise control (CACC) and trains a gated recurrent unit (GRU) for control. Experimental results demonstrate that the trained GRU can achieve ecofriendly driving in CACC without compromising passenger comfort and safety requirements.