Self driving trucks

Background

Modern ground freight systems support complex supply chains and logistics helping cities and economies thrive and grow. More weight and more value is being

transported everyday. A huge portion of those goods are transported by trucks on existing highly connected shared road infrastructure. Those trucks represent a significant fraction of road and mile users, and of energy consumers in many countries. Among those countries are the United States at its continental scale, and Saudi Arabia, where rail-road network is sparse putting high pressure on the road network to connect its distant ports and cities. This project is funded by King Abdulaziz City for Science and Technology (KACST), the national research institute and the national research funding agency in Saudi Arabia, and is being conducted in collaboration with UC Berkeley. The project researches technologies that make the movement of goods on the road safer and more efficient. It builds on top of the California PATH’s decades long experience in highway automation and heavy duty vehicle automation, and on recent advances in robotics and control, artificial intelligence, and connected automation.

Problem statement and objectives

The current area of focus in this project is connected automated heavy duty trucks with emphasis on platooning. Platoons utilize automation and vehicle-to-

vehicle connectivity to enable safe and stable short vehicle following distances. Automation and short following distances increase road capacity, improve road safety, contributes to a smoother traffic, and allows for a more efficient and aerodynamic driving for groups of vehicles. Safe driving and string stability of vehicles in a platoon are priorities for system design and implementation. String stability refers to the ability of the vehicles to safely attenuate oscillations and maintain stability downstream the platoon. Actuation delays, actuation saturation, power to mass ratio, and mass dominance among other factors pose unique challenges to the design of dynamical control for heavy duty vehicles. Keeping those challenges in mind, this project attempts to advance heavy duty truck platooning technology and brings it a step closer to becoming a widely adopted reality. 

Progress to date and plans

Design and implementation of longitudinal control for heavy duty trucks, namely, Cooperative Adaptive Cruise Control (CACC), is the the first building block for the project. End-to-end data-based design approaches are being developed and being benchmarked against classical model-based design approaches. Classical design approaches emphasize highly accurate physics modeling of vehicle dynamics, and more importantly, of the powertrains of trucks. On the other hand, data-based approaches depend on computational methods; they promise performance improvement and design process simplicity. Prototypes for both approaches has been validated against a high fidelity truck simulation environment. Those approaches will be tested on real physical heavy duty trucks, for which data and engineering details have been collected. Moreover, real-time embedded controller systems are being developed to interface with the truck and the other subsystems necessary to carry out the feedback control laws.

Specific deep-RL approaches to truck platooning

Deep-reinforcement-learning (deep-RL) is a computational framework rooted in artificial intelligence and depends on advanced optimization techniques to solve complex problems. It interacts with simulations, or the real-physical world when possible, to achieve the learning (design) objective. This makes it particularly promising for problems difficult to model analytically, and for previously intractable problems. Heavy duty truck platooning is one of such applications that can potentially benefit from more powerful control design techniques. More importantly, driving maneuvers and operations involving such platoons grow quickly in complexity posing tractability challenges to classical design approaches. Those challenges contributed to a limited literature beyond the longitudinal string stability analysis and design. This project explores and improves upon recent advances in deep learning and deep-RL for continuous and high-dimensional control and aims to demonstrate the applicability of these techniques to the future of mobile actuation.