Intelligent Vehicles and Systems Group

Investigators: Mohamed Wahba, Evan Pelletier, Sean Brennan

Sponsor: This work is supported by the Advanced Research Projects Agency-Energy’s (ARPA-E’s) NEXTCAR program.

This project is in collaboration with researchers from Penn State, UNC-Charlotte, MIT and Volvo Group North America. The goal of this project is to develop a model-predictive control system that will exploit vehicle connectivity to reduce fuel consumption for 2016-model internal combustion vehicles by 20% without compromising emissions, drivability, or mobility. This technology will be valuable for different vehicle sizes (light, medium, and heavy duty) and engines (gasoline and diesel). The project will bring together industry and academia to demonstrate the technology in simulation, on a chassis dyno, and on the road for the Volvo VNL 300 heavy diesel truck.

Our technology combines four main features. First, we will utilize vehicle-to-infrastructure/vehicle communications to anticipate traffic patterns and signals. Second, we will develop algorithms for coordinating vehicle activities such as platooning and departures/arrivals at intersections. Third, we will optimize chassis/vehicle control decisions, such as routing and speed setpoint trajectories. Finally, we will also optimize powertrain decisions such as driveline disengagement during coasting, gear shifting, accessory control, and (where available) decisions on engine start/stop and cylinder de-activation. These four features will operate in an integrated manner to achieve co-optimized and coordinated control. This will allow vehicle controllers to operate in a predictive manner that takes into account the collective impact of all the above decisions on vehicle fuel consumption, drivability, mobility, and emissions. The end product of our innovations will be a real-time software suite for fuel economy optimization, intended for vehicles with Level 1 automation or higher, and designed with cyber-security in mind, in the sense of using sensor fault diagnostics methods to determine if any agents are communicating malicious information.

The Intelligent Vehicles and Systems Group (IVSG) will be setting up validation testing hardware and running on-road tests to collect data off the instrumented truck (shown below) that will be used by the team at large. IVSG will also be creating, setting up and running experiments that will help mimic connected vehicle scenarios such as platooning, terrain look-ahead and traffic intersection control.

Investigators: Dr. Sean Brennan, Dr. Matt Parkinson, Bobby Leary, Kannan Anil Kumar, Nick Dow, Stefan Topper, Krishna Prasad, John Barczynski, Guangyu Zhou

This project is funded under the National Science Foundation (NSF) Grant No. 1635663.

Technology supporting visual and motion realism has progressed dramatically in the past two years, even allowing the commoditization of motion capture and head-mounted immersive displays (HMDs). Simultaneously, the ubiquity of low-cost data enables simulated visual environments to achieve near photo-realism. This has the potential to revolutionize both research practice and the designed artifact. However, there are fundamental behavioral differences between virtual reality (VR), augmented reality (AR), and realenvironment interactions. These behavioral differences are known to vary with user characteristics such as age, gender, and expertise.It is unclear how levels of user immersion affect risk-taking and other behaviors. The proposed effort will quantify and model this andrelated behaviors by identifying the relationship between risk compensation behavior and levels of immersion.

Investigators: Dr. Sean Brennan, Bobby Leary

With the rise in research in semi- and fully-autonomous vehicles over the past 30-40 years, environmental maps of the roads are becoming increasingly more important. These maps provide an extended lookahead distance for automated vehicles, as the field-of-view of on-board sensors may be occluded due to local traffic, buildings, or even sun flares. This mobile-mapping platform can createhigh-fidelity maps of the roads with the equipped array of HD cameras, including a 360 degree field-of-view Occam camera, SICK and Velodyne Laser Rangefinders, and a Novatel INS GPS System capable of achieving centimeter-level position accuracy.

Investigators: Kshitj Jerath, Dr. Sean Brennan

This project addresses the problem by identifying appropriate regions of the state space within which the control efforts exerted by a small set of agents can in�‚uence the self-organized dynamics of the ensemble. The problem is considered in the context of the ability of a small set of connected vehicles to affect the self-organized traffic jam dynamics. The methodologies adopted in this approach make use of the kinematic wave theory of traffic �‚ow and the notion of controllability to present the novel concepts of in�‚uential subspaces and the null and event horizons. Results indicate that there exists a strong spatial dependence that governs an agent�€™s ability to in�‚uence the self-organized dynamics of large-scale multi-agent systems. Specifically, under certain simplifying assumptions, the results indicate that connected vehicles must be able to communicate over a span of a few kilometers in order to impact traffic jam dynamics.

Vehicles in region 1 are too far away to impact the traffic jam, and vehicles in region 3 are too close. A vehicle in region 2 is in its influential subspace, and its control actions can impact the jam evolution.

Investigators: Kshitij Jerath, Pramod Vemulapalli, Dr. Sean Brennan, Dr. Dave Bevly (Auburn University)
Sponsors: Federal Highway Administration (FHWA) Exploratory Advanced Research Program

The next generation of safety and vehicle automation will rely on precise positioning, yet GPS-based positioning is hampered by frequent blockages of the GPS signal, even in normal driving situations. This project develops prior work on terrain-based vehicle localization and tracking by extending the research to scenarios including low-cost sensors and real-time implementation. The work demonstrates that vehicle tracking is achievable in such scenarios. Specifically, the work uses an Unscented Kalman Filter with on-the-fly GPS-based initialization. Once initialized, the algorithm is able to maintain tracking in the complete absence of GPS signals. Tracking accuracies of 3-5 meters at low speeds and approximately 10 meters at high speeds have been demonstrated in real-time with simulated low-cost sensors. Efforts are currently underway to determine the theoretical lower bound on tracking accuracy and its dependence on system parameters. The terrain-based localization approach is merged with DSRC radio and vision-based technologies at Auburn University to create a robust alternative to GPS.

MEASURED vs MAP: Comparing an in-vehicle disturbance response with an on-board terrain map database, our algorithms can be used to localize a vehicle without the use of GPS.

Investigators: Kshitij Jerath, Dr. Sean Brennan

The universe around us shows numerous examples of emergent behavior, i.e. collective or macroscopic behavior arising from interactions between agents at the microscopic level. Examples include cascading failures in power grids due to interactions (load transfer) between failed substations, formation of swarms due to interaction (distance keeping) between individual agents (robots, birds, fish etc.) and cognitive functions arising due to interactions (firing) between individual neurons.This project analyzes the impact of modifications in agent interactions at the microscopic scale on the aggregate or emergent behavior at the macroscopic scale. Specifically, previously developed master equation and mean field theory approaches are used to analyze such behavior across various systems such as self-organizing traffic jams, seizures in the human brain and swarming behavior in mobile robots. Additionally, identification of the appropriate ‘aggregate scale’ for such analysis is performed using information-theoretic metrics developed from the concepts of entropy and statistical complexity.

INTERACTING AGENTS: Interactions among agents at a microscopic level may lead to macroscopic behavior that cannot be explained by the dynamics of a single agent studied in isolation. Here small disturbances in wall-following motion of robots lead to jams much like self-organizing traffic jams found on highways.

Investigators: Andrew Whalen, Dr. Sean Brennan, Dr. Steve Schiff
Sponsors: National Institute of Health (NIH)

Epileptic seizures have similarities to brain storms, yet we have no systematic way that reliably detects impending seizures. We propose the use of model based assimilation of seizure data to reconstruct and track EEG dynamics using metrics of generalized synchronization and innovation error to gauge tracking accuracy. Preliminary results demonstrate the ability of a single node Wilson-Cowan neural model to track and reconstruct the 2-layer dynamics for an electrode near a cortical seizure focus. This is the first demonstration of the use of an ensemble Kalman filter and a fundamental nonlinear model representing cortical dynamics, to be used to track human EEG. We are currently expanding tracking to incorporate 2-dimensional sheets of nodes, in the same geometry as the human EEG grid data, and analysis of the controllability and observability of such systems are being computed. Published results to appear in IEEE and PRL.

PREDICTING SEIZURES: The figure shows tracking and reconstruction of the Wilson-Cowan 2-layer dynamics for an electrode near a cortical seizure focus. Note in the reconstruction, and the expanded inset, that the reconstruction of the excitable node, in black, tracks the recorded voltage well. The reconstruction of the assumed inhibitory node, in red, tracks in a delayed fashion to the excitatory variable as we would expect from the Wilson-Cowan equations, the lower plot shows the innovation errors.

Investigators: Andrew Whalen, Dr. Sean Brennan, Dr. Steve Schiff
Sponsors: National Institute of Health (NIH)

Spreading depression (SD) is a dramatic depolarization of brain that propagates slowly and is the physiological underpinning of the initial aura in migraines. This study proposes to represent SD in computational models of the underlying neuronal biophysics, and apply mitigating electrical stimulation using model-based control strategies. The project starts by developing an experimental preparation using a tangential 2-dimensional visual cortex rodent brain slice. SD is triggered with a perfusate potassium perturbation, and SD is imaged using a sensitive CCD camera that detects the intrinsic optical imaging signal associated with index of refraction changes from cellular swelling. A model-based strategy similar to that used in autonomous robotics such as airframe autolanders is employed. A hardware and software control system takes the optical image in real-time, fuses it with a model of SD, reconstructs the underlying physiological processes, calculates needed control, and modulates an electrical field to modulate SD. This will be the first experimental demonstration of model-based control of a neuronal network.

A transdisciplinary German-American educational collaboration has been formed where the researchers (and PIs) synergistically work together within the interface between computational neuroscience, control theory, experimental neurophysiology, and control system engineering. As a collaborative partnership, we anticipate that what is learned in controlling SD may provide a set of testable strategies for electrical control of migraines in people who suffer from severe migraine attacks and are pharmacologically intractable.

EVOLUTION OF SPREADING DEPRESSION: Control chamber (top) and slow wave progression of spreading depression depolarization phenomenon (bottom) imaged with voltage sensitive dye.

Investigators: Mike Robinson, Dr. Sean Brennan

Imagine you are driving down a two-lane one-way city street when suddenly a car parked along the curb opens a door out into the street and the driver steps out into the road. What should you do? Do you have enough time to stop before hitting the driver and the car door? Is the lane next to you open, and if so, should you turn the wheel as fast as possible? Perhaps you should steer and brake at the same time, but how much of each? Now imagine that you are a computer and that you have access to information about the vehicle and your surroundings. What is the best possible maneuver?

We are developing a rapid method for obstacle avoidance path planning that uses optimal control theory to generate trajectories. Compared to many current methods, this research uses analytical solutions to optimization problems as much as possible to reduce the computational complexity of the final algorithm. As a first step, we are developing a quick method to estimate the maneuver envelope of a vehicle for a simplified vehicle model.

Calculating Vehicle Maneuver Envelopes: A rapid method of calculating a vehicle's maneuver envelope would allow autonomous vehicles to quickly evaluate possible avoidance strategies in the event of a potential collision.

Investigators: Michael Petersheim, Dr. Joel Anstrom, Dr. Michael Lanagan, Dr. Andre Boehman, Jamie Clark, Kandler Smith, Dr. Dan Haworth, Ben Zile
Sponsors: Department of Energy, Advantech Inc. EAutomation Group, The MathWorks

This project seeks to develop a first-of-its-kind campus-wide Hardware-in-the-Loop network at Penn State University. This network uses internet-enabled data-acquisition systems and local power-processing interfaces to connect vehicle subsystems - fuel cells, novel batteries, ultracapacitors, flywheel energy storage systems, high efficiency electric drive motors, advanced combustion engine test stands, and chassis dynamometers - in different laboratories distributed over the wide geographic area of the Penn State campus. These components can "talk" with each other just as they would interact in a vehicle, and thus enable the study of the coupled interaction of these components. Hence, researchers can better understand advanced Hybrid Electric Vehicle (HEV) architectures, components, simulation models, and energy management controllers without requiring construction of a new vehicle for each experiment and thus advance understanding of such architectures that are years or decades from integration.

The core challenge in developing a new HIL system is to solve issues of forming and operating a networked, embedded, distributed data-acquisition and control system where each node consists of a hardware or software representation of physical hardware. For most engineering systems, vehicles included, these nodes will often interact in a closed-loop manner such that unsupervised connectivity loss at each node may have catastrophic safety, cost, and data-quality implications. While other faculty are addressing the hardware platform deployments, our interest is to understand the influence of the node-to-node HIL communication interface including development and validation of low-order models of the hardware at each node. The HIL-specific challenges include finding communication sample rates and data lengths sufficient to maintain equivalence to true systems (Speed), finding scaling factors on the data such that one system can be correctly resized to interact with another (Scaling), and finding means to switch seamlessly between hardware and software representations such that there is no perceptible change in data (Switchover).

HARDWARE-IN-THE-LOOP - THE NEXT GENERATION OF VEHICLE PROTOYPING: PSU's hydrogen fuel cell vehicle is seen above on the small vehicle chassis dynamometer. Data from this vehicle is streamed across PSU's HIL network, and the vehicle serves as one of the "nodes" of the campus-wide HIL system.

Investigators: Kshitij Jerath, Dr. Sean Brennan

As adaptive cruise control (ACC) technologies improve and make way into the mainstream vehicle market, it becomes necessary to study the impact of the interaction between ACC and human-driven vehicles on traffic flow. This project studies the impact of adaptive cruise control (ACC) algorithms on the formation of self-organizing traffic jams in medium-to-high density highway traffic. Human and ACC driving behaviors are modeled using the GM fourth model with appropriate model parameters. A master equation approach is used to model the traffic jams at mesoscopic scale. Further, mean field theory is used to reduce the probabilistic traffic jam dynamics to deterministic vehicle cluster or aggregate dynamics. Current work suggests that ACC results in higher traffic flows at the cost of increased susceptibility to congestion.

EMERGENT TRAFFIC JAMS: Analysis and simulations of traffic with mixed human and automated drivers indicate that traffic can operate at significantly higher densities when a large proportion of the traffic stream consists of ACC-enabled vehicles. Contour lines indicate number of simulation runs that resulted in a jam. Dashed red line indicates analytical results.

Investigators: Stephen Chaves, Sanket Amin, Dr. Joel Anstrom - LTI, Dr. S. Brennan - MNE/LTI, Ed Crow - Applied Research Lab, Karl Reichard - Applied Research Lab
Sponsors: Department of Defense, Army TACOM

Military convoys often operate in unpredictable and dangerous situations, including inclement weather conditions and combat zones subject to unexpected attack. This research initiative focused on mitigating rear-end collisions by improving driver-assist systems and vehicle-to-vehicle communications. This study examined commercial-off-the-shelf technologies and their compatibility with sensor/algorithm combinations to be installed on existing military vehicles.

The goal of the study was to produce an inexpensive, unobtrusive, and easily-incorporated solution to warn drivers of impending collisions or unsafe driving situations. This project also examined other convoy-assist research including individual vehicle status monitoring within a convoy, GPS breadcrumb-based platooning, and vehicle string stabilization.

CONVOY COLLISION AVIDANCE: The HEMTT, shown on the left, and other heavy tactical vehicles were tested for the implementation of GPS-based warning systems and vehicle-to-vehicle networks.

Investigators: Sneha Kadetotad, Pramod Vemulapalli and Dr. Sean Brennan

There has been tremendous interest in recent years to develop techniques to locate and track a vehicle on the roadway to enhance the safety and security of the driver as well as develop collision-avoidance systems, driver-assist systems and autonomous vehicle control. The Global Positioning System (GPS) suffers from many issues including poor signal reception and slow update rate, which have prompted researchers to explore alternate map-based techniques of localization. In this approach the roadways are initially mapped by collecting sensor data and processing it to extract parameters of interest and their corresponding locations. The vehicle is then outfitted with the sensor and supplied the map. Map-based localization can be broken down into the global localization problem and the local tracking problem. In the case of global localization, the vehicle has no prior information about its location while in local tracking the vehicle knows its initial position and has to track itself in the map by locally minimizing the error between measured and mapped data. Typically the global localization methods utilize feature-matching techniques that perform matches using data structures like KD-trees or vocabulary trees. In contrast, local tracking algorithms utilize model-based tracking methods such as Kalman filtering or particle filtering.

This work demonstrated a unifying approach that combined the feature-based robustness of global search with the local tracking capabilities of a particle filter. A particle filtering algorithm was implemented that uses the 'uniqueness' of features to localize and track the vehicle. Optimization of the classical particle filtering approach in terms of computational effort, memory requirements and accuracy of localization was performed. Using feature vectors produced from pitch measurements from Interstate I-80 and US Route 220 in Pennsylvania, this work demonstrated wide area localization of a vehicle with the computational efficiency of local tracking.