“We are tantalizingly close,” a recent blog post by Pittsburgh-based Locomation asserted regarding true driverless vehicles navigating public roadways. “But the challenges that remain on the pathway to commercialization will keep the dream at bay for some time to come.” Instead, the tech company’s approach to autonomy keeps a human firmly seated in the cab — at least in the lead vehicle of a truck platoon.
Locomation’s proprietary Autonomous Relay Convoying (ARC) technology has an AI-controlled follower truck that uses its own perception and information exchanged over dedicated short-range inter-vehicle communication (DSRC) to follow the human-driven leader truck. The follower truck remains aware of its surroundings and is cognizant of potential threats, according to Locomation, but it leaves most of the high-level driving decisions to the human and is only concerned with mimicking the maneuvers of the lead truck.
Platoons employing this approach are logging miles and collecting data on different stretches of highway across the U.S. The company recently completed a pilot project operating on Interstate 84 between Nampa, Idaho, and Portland, Oregon. The challenging stretch, which includes steep grades, tricky curves and gusty winds, allowed engineers to observe how the system would have performed on its own. “We are focusing on the parts of full autonomy that can be commercialized successfully today, leaving the parts that require further development and validation for tomorrow,” the company said.
Due to aerodynamic-drag improvements and precise control of the vehicle by the autonomous driving system, ARC is expected to provide several benefits: an estimated 30% or more reduction in operating cost per mile, 8% reduction in fuel expense, increase in freight density of 2x per lane and annual CO2 emissions cuts of 41 metric tons per tractor. Locomation believes human-guided autonomous convoys provide a “stable development path” for higher levels of automation moving forward. Still, truck platooning is not without its own hazards.
Safety analysis of platooning-system concepts
A hazard is an event that poses a danger to people, the system, or the environment. “Generally, it’s caused by human error, a failure of hardware or a software defect. It may also be caused by limitations of the system design — for example, a scenario that’s unaccounted for,” said Alrik Svenson, a research engineer at the National Highway Traffic Safety Administration (NHTSA). Svenson presented at the 2021 SAE Government/Industry Digital Summit an overview of a recent study that assessed the hazards of heavy-truck platooning systems.
The team, which consisted of researchers from Battelle, the Volvo Group, WABCO (now part of ZF) and SAE International, performed a risk assessment for each hazard identified, conducted a Safety of the Intended Functionality (SOTIF) analysis (which addresses unintended system behaviors in the absence of electronic faults), and completed a fault tree analysis (FTA) for some selected hazards. Researchers developed two concept platooning systems, both of which had a driver in the lead vehicle who was fully responsible for driving and overseeing the platoon.
“These are not any particular system, but a generic version of concept systems that are in development for the North American market,” Svenson explained. One was a two-vehicle SAE Level 1 system (2VL1) and the second was a three-vehicle SAE Level 2 (3VL2) system that had two complicating features: the following vehicle had neither steering nor speed control by a human. The operational design domain (ODD) was freeways at nominally steady cruising speed.
A comprehensive list of 57 hazards was developed, based on the research team’s platooning expertise and with contributions from the American Trucking Associations, the Commercial Vehicle Safety Alliance, and SAE International standards committee members. The automotive standard ISO 26262 provided a framework for assessing each hazard on its severity, probability of exposure, and controllability. Potential design operations, maintenance, and safety mitigations were proposed to each hazard and their effect on the identified risks were assessed for both concept systems.
“Most of these [initial hazards] went to green, which is the lowest ASIL [Automotive Safety Integrity Level] rating, with appropriate safety mitigations,” said Svenson. “A few did not get reduced to the green level, so these are the hazards that were selected for further analysis to develop safety countermeasures.” For example, a cyber-attack on the following vehicle’s communication subsystem, which impacted both the 2VL1 and 3VL2 systems, was one such hazard that required SOTIF analysis.
“From this, risks that were unknown prior to the analysis can be identified. As functional hazard scenarios are identified, known unsafe scenarios are increased,” Svenson said. Due to limited published information available on platooning systems, the researchers made some assumptions about the functional system specifications. “The system integrator would need to further develop validation use-case scenarios to test the system in a real-world environment to identify unknown unsafe scenarios,” he said. “The iterative process of SOTIF can increase the safety and reliability of truck platooning systems.”
Based on the fault tree analysis, systems with a human-in-the-loop could benefit from safety mitigations, Svenson added. “Safety mitigation such as training and operating procedures to assess the workload of drivers in the vehicles, as well as how and what type of information is conveyed, could be beneficial for heavy-truck platooning systems.”
AI = real-time optimization
Optimization of fuel consumption is critical for truck platooning, but analysis of full-scale computational fluid dynamics (CFD) simulations can be expensive and time-consuming, according to Dr. Hadi Meidani, assistant professor of civil and environmental engineering at the University of Illinois at Urbana-Champaign. Meidani shared research during a recent Center for Connected and Automated Transportation (CCAT) webinar that proposes an AI-based surrogate model to enable near real-time optimization of platoon configurations.
“Platooning is really an optimization problem,” Meidani said, noting that it is important to run multiple simulations for various uncertainties. “To arrive at the best platoon configuration, which essentially is determining the longitudinal spacing between trucks and their lateral position, you need to run a CFD model to quantify the drag force,” he explained. “They’re not easy problems to solve using finite elements, so you typically need a full-scale high-fidelity simulation, and that may take hours depending on how many trucks you have and how many uncertain factors you want to consider in your problem.”
The researchers propose using a surrogate, which is an approximate model that replaces the full-scale stochastic simulation. “We could use deep neural networks (DNN) where, when you have a particular platoon configuration and there are a number of uncertain variables, instead of having to run an ANSYS simulation each time to calculate the drag force, for example, you just need to evaluate these deep neural networks to calculate the drag and also the gradient of the drag with respect to the design variables,” Meidani said. “The main difference is that deep neural networks are easy to evaluate — they take seconds or sometimes less than a second to evaluate these drag forces. Once this model is trained, it can be used inside an optimization and that optimization is going to be a lot faster.”
Researchers wanted to determine how well a deep neural net could predict the drag force for new scenarios that the training had not seen yet. “We ran 1,000 realizations of a two-truck platoon — this is a very time-consuming undertaking. We had to use supercomputers to facilitate the simulation of these trucks with ANSYS, a process called Design of Experiments,” Meidani said, noting that 85% of that dataset was used to train the neural network and the remaining 15% was used to test the DNN’s accuracy on previously unobserved data. “The accuracy is about 90%, which is pretty good for this very-complex nonlinear model,” he concluded.
For longer platoons, convolutional neural networks (CNN) can be used to create a modular single-truck surrogate when the input layer is an image. CNN is a machine-learning approach to analyze images and has been widely used in image classification, segmentation, detection, etc. “So, you could use image-based approaches in order to predict the performance or the response,” Meidani said, adding that using a physics constraint (partial differential equations) can better train image-based surrogates (CNNs).
“If you don’t include the physics, the training will be purely data driven,” he said, “and you wouldn’t know whether a candidate prediction makes sense physically.” Future work includes using CNNs and variational autoencoders to optimize the geometry of truck bodies in the platoon (using wind deflectors, for example), he said.
“As we increase the level automation for truck platooning, the need for using AI increases,” Meidani said. “Evaluation of costs needs further studies, but it is expected that the return of investment — as a result of fuel efficiency and higher safety — covers the costs of AI.”Continue reading »