Field experiments and validation of heavy duty truck cruise control deep reinforcement learning

Building control systems for heavy duty trucks have historically been dependent on availability of the details of the mechanical configuration of each target truck. This article investigates transfer and robustness of continuous control systems learned using model free deep-RL as an alternative; a configuration agnostic strategy for control system development. For this purpose, deep-RL cruise control policies are developed and validated in simulation and field experiments using two differently configured trucks; full-size Volvo and Freightliner trucks. Their performance are validated for step, ramp, and sinusoidal reference speed trajectories to stimulate steady-state and transient behavior, and to test speed-tracking for low, high, and variable accelerations. The robustness of these controllers were validated for unmodeled gravity effects and for operating the controllers outside of the engine command training distribution bounds. In addition, the controllers were validated against a classical model-based controller.

Automated summary

This article investigates an alternative strategy for developing control systems for heavy duty trucks, which is configuration agnostic and relies on model-free deep reinforcement learning. The study validates the performance and robustness of the control systems through simulation and field experiments on two differently configured trucks.