Sunday, April 28, 2019
Traffic Signal Control by Leveraging Cooperative Adaptive Cruise Control (CACC) Vehicle Platooning Capabilities
Hao Liu, California PATH, UC Berkeley
Convective and Absolute Instability in Multiclass Car-Following Models
Hannes von Allwörden, Universität Hamburg
Vehicles equipped with Cooperative Adaptive Cruise Control (CACC) has the capability to broadcast the real-time speed and location information via wireless communications. They can also safely operate in multi-vehicle strings while keeping shorter than normal gaps among adjacent vehicles in the high-speed traffic stream. Such capabilities can greatly benefit the management of urban signalized intersections. In this study, we have developed a cooperative signal control algorithm that adopts the CACC datasets and the datasets collected by the traditional fixed traffic sensors to predict the future traffic conditions. The prediction allows the signal controller to assign the signal priority to the intersection approach that accommodates the most CACC strings. Such a control strategy can significantly enhance the CACC string operation, which ultimately improves the overall intersection performance. The effectiveness of the algorithm has been tested in a simulated 4-way signalized intersection. The algorithm substantially outperforms the traditional actuated controller as it perceives the traffic flow more comprehensively and assigns the green time resource more efficiently than the traditional controller. Particularly, the average vehicle speed and the average vehicle mile travelled per gallon fuel consumed (MPG) can be increased by more than 10% when the CACC market penetration is 100%. The impacts of the algorithm become greater in mixed traffic. The speed and MPG improvement exceed 30% when the CACC market penetration is 40%. The signal control algorithm can bring about significant benefit even when the CACC market penetration is 0%. In this case, it completely relies on the datasets obtained from the traditional traffic sensors. This finding demonstrates the robustness of the algorithm. It makes the proposed algorithm suitable to implement in real-world intersections under various CACC market penetrations and different levels of vehicle connectivity.
Restricted Lagrangians for Accelerated Model Predictive Control of Autonomous Vehicles
Anson Maitland, University of Waterloo
One of the interesting questions in connection with the introduction of connected and automated vehicles is how they influence the formation and behaviour of traffic congestion.Traffic flow behaviour is often described by microscopic follow-the-leader-type ODE models. The emergence of congestion is then associated with (linear) loss of stability of homogenuous flow solutions on variation of parameters such as average density, sensitivity, etc. In recent years, notions of up- and downstream convective instability and absolute instability have been introduced to further classify unstable parameter regimes. The idea is to be able to tell in which direction a jam will be moving from the perspective of an observer on the side of the road.The concepts of convective and absolute instability are well known from the PDE context. Using exponentially weighted norms, we can recover prior results on microscopic models and are also able to compare their properties with those of corresponding macroscopic models of different order.With a rising amount of connected and automated vehicles on our streets, it is very interesting to consider such questions for multiclass car-following models. The proportion of CAVs can be regarded as a new parameter influencing stability that can take any value in the unit interval for the infinite lane, but only discrete values in a circular road setting. It can be expected that the distribution pattern of CAVs will have an impact, too. For example, CAVs could be grouped together in platoons of some size, or appear individually between "normal" cars according to some pattern. While the arrangement makes no difference for linear stability analysis, periodic solutions that may emerge as a result of Hopf bifurcations will have a different shape.We present an analytical framework and illustrate our results with numerical examples. Our simulations are based on a variant of the widely used so-called optimal velocity model introduced by Bando et al. in 1995.
Design of Lane Change System for Collision Avoidance Using Time-to-Collision and Model Predictive Control
Bo-Chiuan Chen, National Taipei University of Technology
We present a new nonlinear programming (NLP) method tailored for the model predictive control (MPC) of autonomous vehicles (AVs). This method takes advantage of the succession of finite horizon optimal control problems (FHOCPs) characteristic of an MPC. In an offline stage, we use training data (either from simulations or controlled tests) to extract the important optimization search directions. Using these we then define a reduced NLP by restricting the FHOCP Lagrangian to an affine subspace spanned by the most important directions. At each iteration the new NLP method uses the restricted Lagrangian for Newton steps along the important subspace and gradient descent for steps orthogonal to the subspace. Because the reduced NLP is dimensionally smaller than the original (often by an order of magnitude or more) the resulting computational cost of the Newton steps can be reduced, leading to accelerated MPC turnaround times. Further, the combination of orthogonal steps at each iteration ensures there is no controller degradation.When this method is combined with symbolic computing, further computational gains can be made and new avenues of application are opened. In particular, when the dimensional reduction is great enough one can replace the linear solve found in Newton's method by an explicit expression. This realization has further advantages as it allows us to form an accurate explicit model of an MPC controlled vehicle. This can be immensely useful for higher-order tasks in the AV's computational stack, such as local path planning. Because the controlled vehicle model is explicit, one can use fast derivative-based optimization methods for the higher-order tasks, which require a model of the controlled vehicle.We will demonstrate the new method on models of the Autonomoose, the University of Waterloo's self-driving vehicle, in a variety of simulations. These will include reference tracking problems representative of moderately aggressive urban driving scenarios. We will also show how the controlled model can be used in a real-time local path planning algorithm that ensures the reference is controller-feasible.
Model-Based Fault Diagnosis for Lateral Control of Autonomous Vehicles
Chris van der Ploeg, TNO Netherlands (Department of Integrated Vehicle Safety), Delft University of Technology (Delft Center for Systems and Control)
For a static or decelerating object on the road, automatic emergency braking (AEB) can apply braking to avoid the collision if the last point to brake (LPB) is in front of the vehicle. However, if the vehicle has passed the LPB, the collision might be avoided using steering for lane change as long as the last point to steer (LPS) is still in front of the vehicle. Most automatic lane change systems are developed for mild maneuvers without any longitudinal decelerations. In this poster, the distance between the LPB and the front vehicle is calculated to determine if the host vehicle has enough stopping distance when the front vehicle starts emergency braking. If the stopping distance is not enough, the distance between the LPS and the front vehicle is calculated to determine if the host vehicle has passed the LPS. If the LPS is still in front of the host vehicle, time-to-collision (TTC) is used to design the desired lane change trajectory with a trapezoidal profile of lateral acceleration for collision avoidance. A linear time-varying model predictive control (MPC) based on a bicycle model is designed to track the lane change trajectory. If the driver applies the braking, the deceleration is assumed to be constant for the preview horizon. The constant deceleration is then used to update the vehicle speed of the prediction model and the longitudinal displacement of the desired lane change trajectory. CarSim is used to evaluate the control performance and robustness of the proposed algorithm in Simulink. Simulation results show that the proposed algorithm can achieve no collisions with the front vehicle for all the test maneuvers with different vehicle speeds, decelerations, road frictions, or load conditions. The lane change maneuvers for collision avoidance are accomplished with lateral accelerations smaller than the comfort regulation of the ISO 11270. Both collision avoidance and system robustness can be achieved simultaneously.
Lane Keeping and Longitudinal Control for a Semi Truck
Amir Darwesh, Texas A&M University
Over the past years, the automotive industry has seen a constantly increasing automation level for automotive vehicles. This increasing level of automation contributes to an increase of safety and a reduction of traffic congestions due to faster response times and higher reliability, with respect to human-operated vehicles. The automation of a vehicle over the longitudinal and lateral degree of freedom requires vehicle health monitoring to ensure safety of the passengers. In an autonomous highway lane-keeping or path-following scenario, the detection and isolation of faults occurring in the steering system belong to these set of safety measures. By isolating system critical faults, using available vehicle and camera measurements, a decision making process follows. This process results in, e.g., compensation for a fault or transitioning of the vehicle to a safe state.This research presents a set of novel methods for fault detection and isolation for a generalized set of linear time-invariant and parameter-varying systems (to which the linearized lateral dynamics of a vehicle belong). The faults under investigation comprise of an additive fault (e.g., a steering offset) and a multiplicative fault (e.g., a loss of tire pressure) acting on the vehicle's steering input. Additionally, the system is subjected to unknown natural disturbances (i.e., road banking and curvature). A challenge imposed is the lack of isolability of the additive and multiplicative faults using existing fault detection techniques, as well as the parameter-varying nature of the linearized lateral dynamics of a vehicle for a slowly time-varying longitudinal velocity.The first contribution of this research is a nonlinear fault isolation scheme that builds on a moving least-squares approach. The proposed diagnosis filter opts for estimating the additive and multiplicative faults separately. The second contribution is an extension to the first by taking into consideration one of the largest sources of error: the dynamical content of the detection filter. These contributions enjoy guaranteed performance bounds, providing an intuitive tool to analyze the estimation error. We further extend our theoretical results to the case of parameter-varying dynamics, a feature which increases filter robustness in the context of time-varying lateral dynamics.A contribution towards the more practical aspect of this research, is an experimental demonstration of the developed framework. The framework has been applied in a case study for lateral control of a modified Toyota Prius with autonomous driving capabilities. The experimental results validate the technical results, showing that the faults can be detected and isolated.
Spatio-temporal Anomaly Detection in Ego-centric Driving Videos
Yu Yao, Robotics Institute, University of Michigan
This project demonstrates the implementation and testing of an automatic lane-keeping algorithm and longitudinal controller in both simulation, and through actual experimentation for a 18 wheeler class 8 truck.The lane-keeping algorithm is based upon a nonlinear lateral control law developed by Hoffman et. al. for the Stanley vehicle that won the 2005 DARPA Grand Challenge. Previous works with this controller, to the best of the authors' knowledge, none have tested the use with a heavy-duty truck. In regards to the longitudinal controller, a black-box style approach of the truck dynamics is taken due to its nonlinearity, and the unavailability of the engine's rpm, gear, and torque during experimentation. Two separate PI controllers with feed-forward terms are used for the throttle and braking, respectively. The feedforward terms were generated by mapping pedal actuation with measured velocity and acceleration responses. A switching algorithm then chooses whether the truck should throttle or brake. For simulation, the American Truck Simulator is used. With the Stanley controller, calculated steering commands are sent through a virtual joystick which emulates control over the steering wheel and pedals. During experimentation, a retrofitted drive-by-wire semi-truck allows for drive-by-wire functionality. To obtain the lateral and heading errors, a MobilEye 560 sensor is used which provides fitted cubic polynomial functions describing the adjacent lane markers. Velocity and acceleration information are made available through a GPS/IMU system. Results in simulation indicated that the Stanley controller closely followed the pre-recorded path. Validating in actual experimentation, we found that the Stanley controller was able to keep the vehicle within the detected lanes; however, due to extreme dependence of the MobilEye, severe instabilities occurred with failed detection of the lanes, especially in turning. In the longitudinal controller, the switching logic was extensively tested in simulation. Moving to experimentation, we found that the longitudinal controller was successful in tracking speeds above 20 mph. Due to poor low speed performance, a gain scheduler was implemented on the throttle and braking controllers to help account for the nonlinearity of the system.
Hardware-in-the-Loop (HIL) Implementation and Validation of SAE Level 2 Automated Vehicle with Subsystem Fault Tolerant Fallback Performance for Takeover Scenarios
Adit Joshi, Ford Motor Company
Early anomaly detection, temporal anomaly localization, and anomalous objects localization are essential to autonomous driving systems. Existing video anomaly detection (VAD) approaches share three disadvantages: 1) They are either video-level or frame-level classification methods by assuming fixed-length context. However, length of the useful contextual information and anomaly duration can be different for different anomaly types; 2) They are applied to surveillance videos with fixed background while traffic videos are ego-centric with moving background and foreground; 3) Most existing traffic VAD approaches are one-class classification relying on human supervision. We argue that such supervision cannot include all traffic anomalies therefore the trained model can be biased to detect specific anomaly types. This poster proposes a joint learning algorithm for early anomaly detection and temporal anomaly localization in continuing ego-centric driving videos, called Spatio-Temporal Anomaly Detector (STAD). STAD is based on an encoder-decoder architecture for anomaly detection. The encoding module extracts deep representations of historic video frames and the decoding module (generator) generates future frames and reconstructs historic frames. We use a generative adversarial network (GAN) architecture to learn a better generator. A stochastic temporal segmentation module (segmentor) has been applied to sample start times and update the deep representation. The model is trained using an iterative training algorithm so that the learning of different modules benefit from each other. To evaluate the proposed method, we present a new ego-centric video dataset called the Traffic Anomaly (TA) dataset containing different types of driving/traffic anomalies collected from distributed dash cameras in China. Each video clip corresponds to human-annotated start and end times of the anomalies as well as pre-start times at which the anomalies are predictable by annotators. We also provide video-level annotation of the anomaly classes to support our research and make the open data more useful for future use. We have evaluated the performance of our STAD algorithm on the TA dataset as well as other published datasets. We show that the anomaly detection accuracy is increased by learning more accurate start times of anomalies and by leveraging information from the anomaly-related objects. We also present a benchmark comparison of our results with state-of-the-art methods using our TA dataset. The poster will include the following content: 1) Mathematics and architecture of the proposed STAD algorithm; 2) Illustrative samples from the TA dataset including data annotation and anomaly class statistics; 3) Evaluation of STAD on the TA dataset with comparative method benchmarking.
Provably Not-at-Fault Control of Autonomous Vehicles in Dynamic Environments
Hannah Larson, University of Michigan, ROAHM Lab
The advancement towards development of autonomy follows either the bottom-up approach of gradually improving and expanding existing Advanced Driver Assist Systems (ADAS) technology where the driver is present in the control loop or the top-down approach of directly developing autonomous vehicle hardware and software using alternative approaches without the driver present in the control loop. Most ADAS systems today fall under the classification of SAE Level 1 which is also referred to as the driver assistance level. The progression from SAE Level 1 to SAE Level 2 or partial automation involves the critical task of merging automated lateral control and automated longitudinal control such that the tasks of steering and acceleration/deceleration are not required to be handled by the driver under certain conditions . However, the driver is still required to monitor the driving environment and handle scenarios where control is handed over to the driver due to subsystem faults of the automated system. Due to the disadvantages of vehicle testing being expensive, time-consuming and hazardous for testing such scenarios, an alternative method of development and validation is required. Therefore, the objectives of this research are two-fold. The first objective focuses on a real-time powertrain-based Hardware-in-the-Loop (HIL) implementation and validation of an SAE Level 2 automated vehicle. The second objective focuses on studying the performance of SAE Level 2 automated vehicles during takeover scenarios due to subsystem faults. To accomplish these objectives, an acceleration-based Adaptive Cruise Control (ACC) was combined with a path-following lateral control along with supervisory control for system mode transitions due to system deactivations and faults. This research presents system modes in which longitudinal control only and lateral control only are engaged as fallback states to the automated system being faulted for lateral control and longitudinal control failures respectively. Simulations were conducted to evaluate the performance of the automated controls when subjected to these faults. A powertrain subsystem representative of the 2017 Ford Fusion Hybrid was used as the hardware simulation platform using a dSPACE® HIL simulator and CarSim® RT.
Vision-Based Autonomous Driving: A Model Learning Approach
Ali Baheri, University of Michigan Ann Arbor
Autonomous vehicles (AVs) operating on public roads will frequently encounter dynamic obstacles such as pedestrians and other vehicles.It is intractable to guarantee that AVs will be safe (i.e., never crash), because these other moving actors may be malicious.Instead, based on how one plans trajectories, one can attempt to guarantee that an AV will not be at fault in a crash.However, since AVs are typically described by high-dimensional nonlinear dynamic models, performing trajectory planning with such guarantees in real time is challenging; it has thus far only been shown in static environments or structured dynamic environments such as lane keeping.To provide such guarantees in arbitrary dynamic environments, this work extends the existing Reachability-based Trajectory Design (RTD) method, which has been shown to provide provably collision-free trajectories in real time in arbitrary static environments.RTD begins by prescribing a low-dimensional model that generates parameterized trajectories for the AV to track; each trajectory contains a fail-safe maneuver (braking to a stop in this work).The method also requires measuring the error that the AV accumulates when tracking these trajectories.Then, a Forward Reachable Set (FRS) is computed offline that contains the trajectories and tracking error at each point in time.Online, the FRS is used to map predictions of dynamic obstacles from the state space to the space of parameterized trajectories; this produces a set of provably not-at-fault trajectories over which one can optimize at runtime.To enable planning in real time, RTD prescribes a method for discretizing predictions so that the online optimization only needs to evaluate point constraints.Importantly, this discretization provably maintains guarantees of not-at-fault planning.If online optimization cannot be completed within a user-specified planning timeout, then the AV has access to the fail-safe maneuver from its previous planning iteration.RTD is demonstrated over thousands of simulations and two hardware demonstrations on a Segway RMP robot and a small Electric Vehicle (EV).
A Hybrid Control Design for Autonomous Vehicles at Uncontrolled Intersections
Nitin Kapania, Stanford University
We jointly take into account the perception and control problems for an autonomous vehicle in a high-fidelity urban driving simulator. Specifically, we first build a model for the environment, then train a policy inside the learned model to identify the action to take at each time-step. To build a model for the environment we leverage several deep learning algorithms. To that end, first we train a variational auto-encoder to encode the input image into an abstract latent representation. We then utilize a recurrent neural network to predict the latent representation of the next frame and handle temporal information. Finally, we utilize evolutionary-based reinforcement algorithm to train a controller based on these latent representations to identify actions. We evaluate our approach in CARLA, which is a high-fidelity urban driving simulator and conduct an extensive generalization study. Our results demonstrate that our approach outperforms baseline scenarios in terms of the percentage of successfully completed episodes for a lane-keeping task.
Trajectory Tracking for Autonomous Vehicles on Varying Road Surfaces by Friction-Adaptive Nonlinear Model Predictive Control
Rien Quirynen, Mitsubishi Electric Research Laboratories
While autonomous vehicles have the potential to save thousands of lives every year and create significant societal benefits, widespread adoption is unlikely until AVs gain the broad trust of society. Given that every human is a pedestrian at some point during the day, one of the central ways that autonomous vehicles will be evaluated is through their interactions with pedestrians. Interactions with pedestrians can be complex, even for experienced human drivers. From 2015-2016, pedestrian fatalities increased by 9% to 5987, representing the highest number since 1990, and also representing 16% of all automotive fatalities. As autonomous vehicles inch closer to widespread adoption, they must have a clear control strategy for pedestrian interaction that can handle a wide variety of pedestrian behaviors while maintaining a reasonable flow of traffic.Significant research has been conducted to understand the likelihood of pedestrian crossing given certain traffic conditions and pedestrian demographics, and a small but growing body of literature develops control strategies for pedestrian avoidance. However, the literature lacks a contribution that combines the two and explicitly tests whether a proposed control strategy is robust to the variety of pedestrian behaviors that have been observed from experimental studies on real roads. What is also lacking is analysis of how this controller should behave across multiple traffic scenarios - for example, the navigation problem is different for the pedestrian crossing on the opposing side of traffic. To the authors knowledge, there are also no studies that explicitly compare different approaches in order to provide a comparison of alternative solution methods.This poster aims to address the above issues by developing a hybrid control architecture that accounts for several distinct pedestrian modes of behavior at an unsignalized crosswalk. Simulations show that the proposed hybrid controller is able to handle a continuous spectrum of pedestrian gap acceptance behavior, tolerating a range of highly conservative to highly aggressive pedestrians. Additionally, this poster provides a simulated comparison between the hybrid controller and an alternative POMDP approach, citing advantages and disadvantages of each method. Finally, experimental results are shown on a real vehicle to demonstrate the feasibility of the proposed controller.
Using control synthesis to generate corner cases: A case study on autonomous driving
Glen Chou, University of Michigan
We propose an adaptive nonlinear model predictive control (NMPC) approach for combined longitudinal and lateral vehicle control that enables tracking a time-dependent reference trajectory. A key challenge for the control system is to adapt to environmental changes that cannot be predicted but rather are observed from real-time measurements. In particular, we consider the case of a varying road surface that affects the tire forces acting on the vehicle. Such forces are not directly measured, but their effect on the vehicle dynamics can be observed and this can be used to learn the tire model. As the full tire model is nonlinear, data would need to be collected on the entire tire force curve, which is challenging because it requires driving close to the unstable region of the vehicle dynamics. This operating region is not typically visited during normal vehicle driving, and it may be dangerous to visit this part of the dynamics with a controller that has not yet acquired a good prediction model for the vehicle behavior, as closed-loop instability may occur.Our approach relies on NMPC based on real-time iterations of a sequential quadratic programming technique and a fast active-set optimization algorithm. This is combined with a vehicle state and tire stiffness estimator based on a noise-adaptive particle filter and a pre-determined library of tire friction models. The stiffness estimator determines the linear component of the tire model during normal vehicle driving, and a relation between the tire stiffness and the nonlinear part of the tire force is exploited to select the appropriate full tire model from the library. The selected tire model is used to adjust in real time the NMPC prediction model to ensure reliable operation in both the linear and nonlinear region of the tire force curve. We validate the approach in simulation using real vehicle parameters and demonstrate the real-time feasibility of the approach in automotive hardware by assessing the computational load in a dSPACE MicroAutoBox-II rapid prototyping unit.
Intention-Aware Supervisory Control with Driving Safety Applications
Zexiang Liu, University of Michigan
This paper employs correct-by-construction control synthesis, in particular controlled invariant set computations, for falsification. Our hypothesis is that if it is possible to compute a "large enough" controlled invariant set either for the actual system model or some simplification of the system model, interesting corner cases for other control designs can be generated by sampling initial conditions from the boundary of this controlled invariant set. Moreover, if falsifying trajectories for a given control design can be found through such sampling, then the controlled invariant set can be used as a supervisor to ensure safe operation of the control design under consideration. In addition to interesting initial conditions, which are mostly related to safety violations in transients, we use solutions from a dual game, a reachability game for the safety specification, to find falsifying inputs. We also propose optimization-based heuristics for input generation for cases when the state is outside the winning set of the dual game. To demonstrate the proposed ideas, we consider case studies from basic autonomous driving functionality, in particular, adaptive cruise control and lane keeping. We show how the proposed technique can be used to find interesting falsifying trajectories for classical control designs like proportional controllers, proportional integral controllers and model predictive controllers, as well as an open source real-world autonomous driving package.
Effects of driver's emotion on takeover readiness and performance in highly automated driving
Na Du, University of Michigan
This paper proposes a guardian architecture, consisting of an estimation and a supervisor module providing a set of inputs that guarantees safety, in driving scenarios. The main idea is to offline compute a library of robust controlled invariant sets (RCIS), for each possible driver intention model of the other vehicles, together with an intention-agnostic albeit conservative RCIS. At run-time, when the intention estimation module determines which driver model the other vehicles are following, the appropriate RCIS is chosen to provide the safe and less conservative input set for supervision. We show that the composition of the intention estimation module with the proposed intention-aware supervisor module is safe. Moreover, we show how to compute intention-agnostic and intention-specific RCIS by growing an analytically found simple invariant safe set. The results are demonstrated on a case study on how to safely interact with a human-driven car on a highway scenario, using data collected from a driving simulator.
Communication Modality and Automated Vehicle Acceptance
Qiaoning Zhang, University of Michigan
Highly automated vehicles (AVs) are a major focus for automobile manufacturers in part because they have the potential to provide safer and more efficient driving. However, if the AV reaches its system limit, the driver will be required to resume control of the vehicle in a short amount of time. Previous studies have shown that drivers have difficulty taking over because they have been decoupled from the operational level of control and do not have enough situational awareness to deal with such an urgent event. To tackle this problem, researchers have investigated the effects of different factors on the driver's takeover performance including takeover lead time, secondary task workload, traffic density, etc. Nevertheless, to our knowledge, few studies have paid sufficient attention to the influence of emotion on takeover performance. Given the gap in the literature and the importance of drivers' emotion in driving, a user study was conducted. The study's goal was to examine the effects of driver's emotions on takeover readiness and performance. The study employed a within-subject design with 32 participants in a driving simulator. Four types of emotions were examined: angry, sad, happy and calm. Participants had each emotion induced through watching movie clips, and were required to take control once the takeover warning displayed. Takeover driving behavior and subjective ratings of takeover readiness and performance were recorded for each takeover event. Participants had the highest takeover readiness and performance while they were calm and the lowest takeover readiness and performance when they were angry. These results are critical in helping us understand the vital role emotions play in takeover situations. Consequently, these results have important implications for the design of effective and safe, in-vehicle alert systems for highly automated vehicles.
Utilization of Connectivity in Traffic Flow Forecasting
Tamas G. Molnar, University of Michigan, Ann Arbor
Despite significant technological progress and enthusiasm from the popular press, there are still many barriers to widespread adoption of automated vehicles (AVs). One such challenge pertains to the communication modality between the AV and its passengers. Communication between the AV and a user is vital to ensuring acceptance of AVs. However, we know very little about the impacts of communication modality on AV acceptance. To address this, this study examines the impacts of communication modality and cognitive workload on several measures of AV acceptance: trust, preference, and anxiety. The study is a 2X2 factorial design which varies communication modality: audio or visual and cognitive workload: high or low. The audio & the visual alerts will be delivered in the vehicle via the vehicle's audio system or via the vehicle's instrument panel, respectively. Cognitive workload will be manipulated via a secondary task. The secondary task will involve participants watching two videos on an tablet and answering questions about the contents of the videos. 32 participants will be recruited, and each will experience four drives with various events pre-scripted within each drive. Events types will be balanced for urban and highway driving. Participants will sit inside an immersive level-4 driving simulator for four eight-minute drives. The roadway environment will be projected on three screens about 16 feet in front of the driver (120-degree field of view) and a rear screen 12 feet away (40-degree field of view). Once the participants have finished each drive, they will fill out surveys that measure their perceptions about trust, preference, anxiety, and workload. Physiological metrics including gaze and fixation will be recorded and analyzed as well. The results will be important for helping us understand the effects of communication modality on the acceptance of AVs. Study findings will have implications for design of AVs that customers are likely to accept, especially in terms of communication, user interface, and other human-machine user experience considerations.
A Scalable Framework For Real-Time Multi-Robot, Multi-Human Collision Avoidance
Sylvia Herbert, UC Berkeley
The potential of vehicle-to-vehicle (V2V) connectivity in traffic flow forecasting is investigated experimentally. Experiments were performed by participation in congested highway traffic with a small group of connected vehicles. Each vehicle was equipped with GPS sensors to measure their motion and with cameras to record the ground truth about their environment. In addition, connectivity was established between each pair of vehicles: dedicated short range communication (DSRC) devices were utilized to share position, velocity and heading angle data amongst the vehicles with an update rate of 10 Hz. The recorded experimental data is analyzed from two perspectives. On the one hand, the reliability of communication is evaluated in terms of its range and packet delivery ratio. On the other hand, the formation and propagation of traffic congestions is studied. Velocity fluctuations as a function of time and space are inspected to extract the propagation speed of traffic jams and to examine the amplification or decay of congestion waves. The experimental data is also used to drive analytical traffic models in order to predict the motion of vehicles upstream. Based on the comparison of predictions and data, it is demonstrated that traffic flow models supplemented with data provided by V2V communication are capable of forecasting traffic flow in the timescale of minutes.
As robotic systems are increasingly used in crowded environments, safe multi-robot multi-human navigation is crucial. Achieving this has three main challenges, explored below: Safe and Efficient Human Motion Prediction:For probabilistically safe human motion prediction we developed a confidence-aware prediction framework that allows us to employ simple models of human motion while reasoning about the mismatch between these models and the observed human behavior. The framework employs a probabilistic Boltzmann model of human behavior, where the reward function and dynamics can be learned or encoded. By maintaining a Bayesian belief over the scalar temperature parameter, our framework automatically adjusts the variance in the distribution based on measurements. When the observed human behavior is well-explained by the model, our framework provides a tight "confident" distribution over future human motion. When our model does not explain human behavior well, the probability distribution over the next human states increases in variance, and in the limit becomes the human's forward reachable tube. This allows us to take advantage of known structure in human motion when it exists, while maintaining safety when this structure is incorrect. Safe and Efficient Robot Planning and Control:We couple these probabilistic human predictions with a method we developed for guaranteed safe online motion planning for robotic systems called FaSTrack. This technique allows an autonomous system to plan around moving obstacles in real time using a low-dimensional model with any motion planning algorithm. Through a precomputation using Hamilton-Jacobi reachability analysis, a feedback controller is precomputed to ensure that tracking error due to model mismatch and external disturbances remains bounded. This allows us to plan quickly using the low-dimensional model while preserving safety for the high-dimensional robot. Synthesis into Multi-Robot Multi-Human Scenarios:Scaling planning, prediction, and safety guarantees to multiple robots and humans is challenging, primarily due to the difficulty of joint planning and prediction for multiple robots and humans. To ensure real-time feasibility around multiple humans, robots predict multi-human motion using a simple model neglecting future interaction effects. Because this model will be a simplification of true human motion, we use our confidence-aware predictions that become more conservative whenever humans deviate from the assumed model. Our connected robots plan sequentially according to a pre-specified priority ordering, which serves to reduce the complexity of the joint planning problem while maintaining safety with respect to each other. We demonstrate our framework in hardware with real humans and multiple quadcopters, and provide a large-scale simulation to showcase scalability.
Monday, April 29, 2019
Adaptive Cruise Control for a Class 8 truck
Timothy Overbye, Texas A&M University
Large Scale Simulation for Autonomous Driving
Akhil Nagariya, Unmanned Systems Lab, Department of Mechanical Engineering, Texas A&M University
Adaptive cruise control (ACC) is quickly becoming a common feature on modern cars and is one of the first steps towards total autonomy. However, comparatively little work has been done on ACC for semi trucks and other large vehicles. In this project we investigate the adaptation of an existing car ACC module for use on a semi truck. The basic algorithm works in three stages. First, a desired lead vehicle distance is calculated based of a constant time gap plus offset. Next, the difference between desired and actual lead vehicle distances is used in a lookup table to find our desired speed. Finally, our actual speed, desired speed, and lead vehicle acceleration are used to set maximum and minimum acceleration limits.To validate performance we used several common road scenarios as test cases. These include stop and go traffic, approaching a stationary vehicle at high speed, lead vehicle slowing down, lead vehicle speeding up, and lead vehicle accelerating from stop to high speed. Initial testing was done using American Truck Simulator. For further testing we used a virtual lead vehicle combined with a real semi truck on a flat level road.We found that the unmodified ACC frequently requested target speeds that could not be achieved by the truck. Similarly, acceleration limits were also out of range. The ACC was then modified by increasing both the time gap and constant offset. Additionally, the distance to speed lookup table was changed to accommodate larger following distances. After making these changes the truck was better able to achieve requested target speeds and performance in test scenarios was acceptable. However, acceleration limits were still out of range and could be removed with no performance change.In the future we intend to start testing with a real lead vehicle. We expect new challenges such as noisy data and tracking loss in turns to change our results. Finally, we expect varying road grades, especially highway ramps, to be a larger factor for trucks than cars.
CONNECTING CARS OF TODAY: VEHICLE FLEET BASED TRAFFIC AND ROAD WEATHER MONITORING
Ari Tuononen, RoadCloud
The simulation platforms available today for autonomous driving provide traffic simulation and behaviour of other agents which are rule based and hence limited in representing real traffic situations. To provide realistic simulation of complex urban traffic situations like intersections, merging, left turns etc and test the behaviour of an autonomous vehicle in presence of other autonomous vehicles, the simulation environment must be able to simulate large number of vehicles each of them equipped with multiple sensors. Each simulated vehicle in turn can be used to test and validate different autonomous driving solutions. To effectively test the perception and control modules of an autonomous driving solution, it's also necessary for a simulation platform to provide visually rich photorealistic scene rendering and accurate physics simulation which comes at a price of high computation cost. Given the computational resources it is not feasible to simulate more than 3-4 vehicles each equipped with multiple cameras without significant performance issues. Most of the simulation environments available today support simulation of a single vehicle, often due to these hardware limitations.In this work we present a large scale simulation environment for simulating multiple vehicles aimed at autonomous driving scenarios. Game engines like Unity provide excellent realistic scene rendering and vehicle dynamics simulation exploiting the full power of current generation GPUs. These game engines also provide support for multiplayer gaming in which multiple machines connected over a network share the same scene and interact witheach other. We leverage the features provided by these powerful game engines to build a simulation environment for autonomous driving which is capable of simulating detailed realistic scenes. This environment can be shared by multiple machines connected over a network, with each machine simulating multiple vehicles. Our simulator can be used to simulate complex urban traffic where each vehicle can be independently controlled. The simulation environment uses Rosbridge server for communicating with ROS, thus enabling a tight integration with a widely used robotics framework. ROS is used in the prototyping and development of many autonomous vehicles hence this simulation environment can be used to effectively test algorithms developed by the robotics community. We have used this platform to successfully simulate 15 vehicles on 5 different machine connected over network each simulating a camera(30fps) and a Lidar(10Hz) sensor.
Hierarchical Optimization for Eco-Cooling of Connected and Automated Vehicles (CAVs)
Hao Wang, University of Michigan
Accurate wintertime road weather information is essential for road maintenance operations and for more autonomous driving functions of vehicles. Traditional static road weather stations typically provide basis for road weather monitoring. However, on-board sensors of cars have become more popular and recently wireless data transmission allows that commercial vehicles can be used as probes instead of road maintenance vehicles itself. In other words, cars are turning into mobile weather stations producing massive amount of different type of data about continuously alternating surroundings. It is not mandatory to have all the cars equipped with special instrumentation. We take advantage of, carefully selected and highly utilized vehicles which produce data around the clock from the road network of interest.The data acquisition system of the fleet vehicles consists of optical road probe, 6 degree-of-freedom Inertial Measurement Unit and CAN-bus access. The position is available with GNSS and data is transmitted via cellular network to the cloud where it is processed. This enables for example automatic calibration functions of sensors removing the repeated need for maintenance. The example results to be shown on the poster were analyzed from over 3 million kilometers of measurements in Europe. We show how static road weather stations can give dangerously misleading information about the road condition: information is too sparse to be used in active safety systems or autonomous cars. We also show local road weather evolution including chemically wet road surfaces together with resulting vehicle active systems interventions (ASR, ABS & ESC). Indeed, we observe thousands of active systems interventions during a blizzard. As a conclusion, even a dense network of static road weather stations cannot provide accurate local road condition information, and overall road weather can be estimated incorrectly. Meanwhile, mobile measurements by using highly utilized commercial vehicles is cost effective method to produce current road condition information and is undoubtedly highly valuable for now-casting and forecasting road weather models. By utilizing a fleet of measurement vehicles and cloud-based analysis, both local and large-scale road weather phenomena can be observed in real-time.
Game Theoretic Modeling of Interactive Traffic for Verification and Validation of Autonomous Vehicle Control Systems
Nan Li, Department of Aerospace Engineering, University of Michigan, Ann Arbor
With the emergence of connected and automated vehicles (CAVs) equipped with advanced sensors for perception and localization, communications and feedback have become more integrated and ubiquitous, leading to tangible benefits with enhanced safety and improved energy efficiency. At the meantime, connectivity and autonomous driving technologies open up new dimensions for control and optimization of vehicle dynamics and powertrain systems. While most of the recent CAV-related research, such as eco-driving and platooning, has been focused on reducing traction power related losses, little has been reported on efficient thermal management of connected and electrified vehicles using external information made accessible through vehicle to vehicle (V2V) and vehicle to infrastructure (V2I) communications. One key challenge in integrating the vehicle thermal management (VTM) system with the powertrain control system is to deal with the time-scale difference between these two systems, which is mainly due to the relatively large thermal inertia. Thus, to capitalize on the valuable V2V/V2I (V2X) information for improving the vehicle energy efficiency, optimization of the VTM process over a long time-horizon is required to determine the optimal energy flows and power split strategies. However, the vehicle speed profile can be accurately predicted using V2X information only over limited time horizon, and extending this horizon is subject to uncertainties. In this poster, we will present a hierarchical optimization framework based on model predictive control (MPC) scheme for efficient cabin thermal management (i.e., eco-cooling) of connected and automated HEVs. The proposed hierarchical MPC (H-MPC) allows for exploiting traffic events forecast with different prediction accuracies over short and long prediction horizons to schedule optimal vehicle thermal load trajectories. These trajectories will be next incorporated into a vehicle-level power-split controller to perform fuel economy evaluations. Simulation results over different driving cycles with the proposed eco-cooling strategy demonstrate up to 5.3% improvement in the fuel economy when compared with a conventional cabin thermal management strategy.
An Autonomous Shuttle Platform for Evaluating Perception and Control Strategies
Garrison Neel, Unmanned Systems Lab, Department of Mechanical Engineering, Texas A&M University
In the near to medium term, autonomous vehicles will operate in traffic scenarios together with human-driven vehicles. The complexity of autonomous vehicle control emanates from the need to deal with the extensive interactions among different traffic participants, such as the interactions between autonomous vehicles and human-driven vehicles. To verify the correctness of an autonomous vehicle control system in terms of safety and performance in such interactive scenarios, traffic simulators capable of representing vehicle interactions with reasonable fidelity can be used for safe and quick virtual tests before actual road tests.In this poster, we will describe a game theoretic framework to model the interactive behavior of vehicles in various traffic scenarios involving multiple vehicles, including on highways and at uncontrolled intersections. The framework is computationally scalable in terms of being capable of modeling and simulating traffic scenarios with a possibly large number of, e.g. 30, interacting vehicles with manageable computational effort.The model for highway traffic scenarios developed based on the proposed game theoretic framework has been integrated with a high-fidelity car-driving simulator and validated using human driving data. The model has been used as a platform for simulation-based testing, verification and validation of autonomous vehicle control systems, where simulation results are used to statistically evaluate and compare the performance of these systems, identify challenging scenarios for and faults in them, and calibrate their parameters. These results will be described in this poster.In addition, recent results for modeling traffic scenarios at uncontrolled intersections will also be presented in this poster.
Predictive Data-driven Vehicle Dynamics and Powertrain Control with Connectivity
Yeojun Kim, University of California, Berkeley
This poster describes an autonomous shuttle platform which targets providing last-mile transportation. The platform consists of a Polaris drive by wire unit equipped with Velodyne VLP-32C LIDAR, Perceptin DragonFly stereo vision system, and a VectorNav VN300 GPS/INS. Mounting space is available for many additional other sensors. Our aim is to create a functional shuttle whose software and hardware components are interchangeable and upgradeable as new approaches are developed in each component area. Interchangeability allows direct comparison of the performance of different approaches, such as localization methods, or path following or obstacle detection algorithms. We have the ability to evaluate performance in a wide-open obstacle-free testing environment, as well as in highly pedestrian areas such as a college campus, and downtown areas with dense vehicle and pedestrian traffic. An initial implementation of such a shuttle is presented, detailing the overall architecture, controller structure, waypoint following, obstacle detection and avoidance, LiDAR based sign detection, and pedestrian communication. Our software is built on ROS, the Robot Operating System. The lateral and longitudinal controller is a PID controller that uses velocity and yaw rate input for controlling steering, throttle, and brake. The waypoint following controller is a simple proportional controller using heading error. We have evaluated the accuracy of path following for our controller as well as a pure-pursuit controller using CTE (cross-track error) as the metric. Our current shuttle implementation is able to follow a predefined path with marked stop positions, and stop for obstacles detected in LiDAR. Deployment in downtown Bryan, TX presented challenges such as poor GPS accuracy, passenger interaction, navigating 4-way intersections, obstacle motion, and static obstacles in the path. Our short-term workarounds to these challenges are presented, as well as long-term solutions for future deployments. Current projects involving the shuttles include drivable area segmentation, visual odometry, local path planning, LiDAR based curb detection, testing of various SLAM algorithms, and implementation of neural net based steering models.
Impact of connected vehicles on traffic dynamics by experiment and simulation
Sergei Avedisov, PhD student
We design, implement, and test a novel predictive and data-driven vehicle dynamics and powertrain control technology on a plug-in hybrid electric vehicle (PHEV). In our approach, historical and real-time data-feeds are used in a systematic and integrated way to co-optimize powertrain and vehicle dynamics control at different time-scales. In particular, the developed technology coordinates real-time predictive control on the ECU with connectivity to other vehicles, traffic lights, and the cloud. Because we have multi-states and inputs with their respective constraints in our system and want to exploit predictions and optimization, Model Predictive Control design is adopted in our approach.Our goal is to demonstrate a minimum of 20% reduction in total energy consumption (MPGe) for PHEV. We show the performance of the developed technology in representative cases, considering both urban and highway contexts. In particular we will focus on two scenarios: (1) eco-approach/departure at signalized intersections and (2) eco-CACC (Cooperative Adaptive Cruise Control) and SPD-HARM (speed harmonization.Our approach will be validated in two different testing platforms: Hardware-In-the-Loop (HIL) simulation and real-world testing. In HIL simulation, we use high-fidelity traffic simulators in our test desktop, the actual PHEV sitting on a Dynomameter for accurate vehicle dynamics and powertrain dynamics, and the actual on-board ECUs. Especially, our simulation reproduces the realistic traffic flow, traffic lights, and road grades based on real-world data for fair assessment of our approach in diverse and reproducible settings. The real-world testing will be conducted after our developed technology is validated for both safety and energy efficiency in our HIL simulation. The poster will be structured into the following sections: (i) the system and control hardware architecture for our test vehicle, (ii) the overview of our HIL setup, (iii) our model predictive control design, and (iv) the latest HIL simulation results and analysis for scenarios (1) and (2).
A Hybrid Framework for Improved AV-Pedestrian Predictions
Suresh Kumaar Jayaraman, Mechanical Engineering
Recent developments in vehicle automation and advanced driver assistance systems (ADAS) coincide with improvements in driver safety, however mobility of roadways has not witnessed a similar trend. The limiting effect of vehicle automation on mobility can be attributed to the limited sensory range, as traffic patterns such as congestion waves or jams occur on a much larger scale than the ranges of cameras, lidars, and radars. Moreover, obstructions such as blind turns or other vehicles may further limit the range of these sensors. The range shortcomings of sensors seen can be addressed using wireless vehicle-to-everything (V2X) communication. Vehicles equipped with V2X communication, called connected vehicles, can send and receive messages over hundreds of meters and through obstructions such as blind corners or large vehicles. An automated vehicle equipped with V2X communications, called a connected automated vehicle, can then respond to large-scale traffic patterns to mitigate congestion. The resulting traffic network, called a connected vehicle network, and composed of conventional human driven vehicle, connected human driven vehicles, and connected automated vehicles would then have a higher throughput than a conventional traffic network. To evaluate the effects of beyond-line-of-sight communication on traffic patterns in a realistic environment, we develop an experimental setup for a connected vehicle network consisting of two human driven connected vehicles and a connected automated vehicle. Our setup uses a boundary condition, which allows traffic patterns to propagate for a small number of cars and at realistic vehicle speeds. These experiments show that by using beyond-line-of-sight information, the connected vehicle damps traffic congestion in the network much more effectively than when it only uses information from the immediate predecessor. We also demonstrate that these experimental results are reproduced by simulations of car-following models with model-matched parameters and delays.Since the car-following models replicate the traffic patterns observed in the experiments, we develop large-scale simulations of connected vehicle networks to asses the impact of connected automated vehicles in mixed traffic, which features different penetrations of connected and connected automated vehicles. These simulations show that connected automated vehicles using beyond-line-of-sight information to control their longitudinal motion are significantly more effective at improving the throughput of the traffic network than connected automated vehicles using information solely from their immediate predecessor. This demonstrates that the cognizance connected automated vehicles gain from using V2X communication is key to benefiting mobility of future roadways.
Motion Prediction of Human-Driven Vehicles by Exploiting Vehicle-to-Vehicle Communication
Linjun Zhang, Ford Motor Company
Automated vehicles (AVs) have the potential to make transportation safer and more efficient. However, for safe and efficient operations, the AVs must be able to safely interact with pedestrians. This requires accurate long-term (~3-5 s) predictions of pedestrians' trajectories. This topic is particularly critical as pedestrians can change their trajectories instantaneously leading to fatal accidents. Thus, the research problem is to develop an accurate prediction framework for long-term (~5 s) pedestrian trajectory prediction. Several models have been developed for predicting pedestrian trajectory. However, these models (1) predominantly use only pedestrian parameters such as dynamics, pose, and gaze and do not consider the social factors affecting the pedestrian behavior, and (2) generalize the crossing activity as having a single goal to reach a destination and fit a single model for prediction. However, the actual crossing task may include various sub-goals (stages) such as walking up to the crosswalk, waiting to make the decision of crossing, and starting to cross and reaching the destination on the other side of the road. At each stage, different factors can influence the pedestrian behavior and a single model may not be able to capture these dynamics. We propose a hybrid modelling framework, where the pedestrian state is identified by their sub-goal (approaching the crosswalk, deciding to cross or crossing the road) and each sub-goal has a separate model to predict pedestrian trajectory conditioned on different sets of factors. State (sub-goal) transitions would be a function of pedestrian parameters (position, gaze) and the prediction likelihood of the models. An initial set of parameters that influences pedestrian's crossing behavior were identified: pedestrian position, speed, distance to curb, waiting time, gaze, vehicle speed, and distance to collision. Preliminary data from a virtual reality study was analyzed. Results from a non-linear symbolic regression between pedestrian trajectory and the collected features indicate that vehicle acceleration, distance to curb, and position of adjacent vehicle are good predictors of pedestrian trajectory during the 'Starting to cross' phase whereas 'Gaze angle' is a good predictor during the 'Crossing' phase. Ultimately, by completing this framework we expect to achieve a more accurate long-term pedestrian trajectory prediction leading to safer interactions between AVs and pedestrians.
Longitudinal and lateral car following with connected automated vehicles
Sandor Beregi, MTA-BME Research Group on Dynamics of Machines and Vehicles, Budapest University of Technology and Economics
Accurately predicting the future motion of surrounding vehicles plays a significant role in the safety and the comfort of connected and automated vehicles (CAVs). In near future when the market penetration of CAVs is low, it leads to mixed traffic that is comprised of both human-driven vehicles and CAVs. The uncertainties in the dynamics of human-driven vehicles lead to challenges in the prediction of their future motion.Conventional methods for predicting the motion of human-driven vehicles usually assume constant acceleration to predict their future positions and velocities. However, such methods may result in large errors, especially when the velocities of human-driven vehicles significantly fluctuates. Deep neural network methods are potential alternatives for predicting the motion of human-driven vehicles. However, deep neural networks usually contain a tremendous number of parameters and hence may not satisfy the real-time performance required by the implementation of CAVs. Moreover, the results of deep neural networks may be accurate in the statistical sense, and may not lead to accurate predictions for individual vehicles.In this research, we propose a data-driven approach for predicting the motion of human-driven vehicles immediately in front of CAVs by exploiting the motion data received from distant CAVs via vehicle-to-vehicle (V2V) communication. First, we use a linear discrete-time model to approximate the dynamics of human-driven vehicles, where the reaction delays of human drivers are taken into account. Then, we collect motion data from distant CAVs and the human-driven vehicle immediately ahead, and identify the model parameters by applying the recursive least square approach to the collected data. Then, assuming that distant CAVs broadcast their planned future motion, we predict the future motion of the human-driven vehicle immediately ahead. To improve the prediction accuracy, we include triggering time and physical constraints. Numerical simulations show that our presented method has a significant improvement on the prediction accuracy compared to the conventional constant acceleration prediction. The results also show that physical effects such as uncertain human delays and packet drops in V2V communication have little effects on the prediction accuracy; however, the varying parameters of human-driven vehicles may degrade the prediction performance.
Differential Dynamic Logic Model for Verifying Correctness of Collision Avoidance System in a Car
Aakash Abhishek, University of Michigan Student - (working under Prof. Jean Baptiste Jeannin)
Recent years have seen an increased focus on automation and advanced driver assistance systems (ADAS) in academia and industry. The goal of vehicle automation is to increase safety, driving comfort and fuel economy compared to human drivers. Moreover, communication between the automated vehicles may be used to coordinate the traffic flow on the roads and to avoid traffic jams. Nowadays, most of the automated vehicle control systems are realised primarily by optical sensors. In some cases however, these sensors may not be reliable for vehicle control e.g. in bad weather conditions or in case of missing lane markings. On the contrary, satellite navigation may be used to vehicle localisation even in poor visibility and to improve the resilience of automated vehicles in inclement weather.In our study, we aim to carry out lane-changes with an automated vehicle following a human driven car relying on GPS data and vehicle-to-vehicle communication. From the computational point of view, this might be a complex task as it requires simultaneous control of the vehicle speed and lateral dynamics. It can be shown though that – at least from the point of view of linear stability – the longitudinal and lateral controls can be studied independently from each other. To guarantee the linear stability, these control algorithms need to handle the time delay brought by communication and actuation. We show that although the combination of these can be relatively large (around 1 s), stable control can be achieved both in the lateral and the longitudinal directions.Experiments were conducted on a test track to explore the stable parameter domain of the control gains and to identify the unknown system parameters. These results were also supported by the numerical stability analysis of the longitudinal and lateral controls of the single-track model of the vehicle. Based on both the experimental results and the numerical analysis, we chose the lateral and the longitudinal control gains and successfully carried out connected lane changes on the test track.
Trust and risk in autonomous ground systems
Connor Esterwood, UM School of Information
The task of guiding an automobile while avoiding any collision with the surrounding objects, is of immense practical value. This requires predicting the trajectory of an automobile, during on road maneuvers and using this knowledge to autonomously guide the vehicle on a given track. At present, a lot of literature exists on predicting vehicle behavior through various models of dynamics. Similarly, the area of the trajectory planning for obstacle avoidance is also very developed. However, the implementation all these existing systems (path planning and obstacle avoidance controllers) involve some sort of interaction between the cyber systems (controllers, processing units) and physical system (actuators, final behavior and path of car on the track). Due to the numerous subtleties' lying underneath the surface of any cyber-physical system's operation as well as the rise in complexity of such systems in the current era, its difficult to say whether the designed system will be able to avoid collisions under all possible scenarios on the road. However, such guarantee of obstacle avoidance is of utmost necessity before the system can be deployed in practice. Hence, there's a need for formally verifying the safety and performance of these said systems, in order to prevent any mishaps in practice. In our present work, we have addressed the task of formally verifying the correctness of a simple collision avoidance system for an automobile. This formal verification provides a mathematical guarantee that the given system can prevent the car from collisions under any possible scenarios as long as certain (well defined) conditions are satisfied. We have used a basic kinematic model of vehicle based upon the Unicycle Model. This model's prediction of the motion of vehicle in a given maneuver, has been used as a basis in formally verifying the correctness of a simplistic collision avoidance system for a car. At present we consider the car's motion to be limited to a few basic maneuvers viz. turning in a constant radius, with or without simultaneous braking. Under these simplifications, the formal proof of collision avoidance for the employed safety system has been developed. The logical model employed to formally verify the correctness, is developed in Differential Dynamic Logic. This DDL model has been used along with an automatic theorem prover KeYmaeraX, to generate the formal proofs of safety.
Effects of limited acceleration capabilities in connected automated vehicles
Adam Kiss, PhD Student, Department of Applied Mechanics, Budapest University of Technology and Economics
Trust is an essential ingredient for the successful adoption of any new technology. Trust's impact is especially felt in relation to technologies that require an effective level of human-automation teaming. Autonomous ground systems (AGS) are just one example of such a technology and have become quite popular in recent years. Perhaps the most significant benefit these systems offer is that of flexibility. AGS are projected to allow a user to work alongside the AGS so that they may then perform a second task simultaneously. This study contributes to the literature on human-automation trust by investigating different types of risk and their impact on trust and a secondary task performance. Utilizing a 2x2 within subject design, we manipulated internal and external risk in a simulated driving environment. We measured secondary task performance by calculating participant's overall scores on a touchscreen search task where users had to locate the "Q" among a series of "O"s. We recruited thirty seven licensed drivers and ran each through four different experimental conditions. Our results indicated that both types of risk moderated the effects of trust on performance. Specifically, we first found that under high external risk conditions, higher trust resulted in lower performance while under low external risk conditions, higher trust resulted in higher performance. Second we found that, under high internal risk conditions, and low internal risk conditions, higher trust resulted in higher performance. In relation to this second finding, the increase in performance was greater for the high internal risk conditions than it was for low internal risk conditions. Together, these findings highlight the impact of trust on performance moderated by risk. This study's findings are applicable to the design and development of AGS as they provide a basis for understanding to what degree the secondary task performance will be impacted by trust as moderated by risk.
Autonomous Freeway On-ramp Merging Using Deep Reinforcement Learning
Yuan Lin, University of Waterloo
In this contribution, heterogeneous connected vehicle systems, which include connected human-driven vehicles as well as connected automated vehicles, are investigated. Delays due to human driver reaction, vehicle-to-vehicle (V2V) communication, and throttle/brake actuation are incorporated into the car-following model. Saturations due to limits of acceleration and breaking of the vehicles are also taken into account. Stability analysis is used to identify regions in parameter space where oscillations arise due to loss of linear stability of the uniform flow. Moreover, with the help of numerical continuation, the skeleton of the traffic dynamics is drawn as a three-dimensional wire frame representation. It is demonstrated that acceleration saturation leads to new dynamic behavior, where bistability between uniform flow and traffic waves appears through the presence of isolas. In this scenario, no linear stability loss occurs, but larger perturbations may still trigger traffic waves. Then the effects of connected automated vehicles are investigated at the nonlinear level. It is demonstrated that utilizing long-range wireless vehicle-to-vehicle (V2V) communication the connected automated vehicle is able to eliminate the oscillations and make the uniform flow globally stable.
Improving Fuel Economy of Heavy Duty Vehicle in Traffic using Connectivity and Automation
Chaozhe He, Navistar, Inc.
Freeway on-ramp merging is a challenging task for both human drivers and autonomous driving systems. Freeway traffic flow is usually moving at high speeds and a merging vehicle needs to maintain a high speed during merging. The decision-making process for the merging vehicle involves selecting a gap between two neighboring vehicles at the merging zone on the freeway. During merging, the merging vehicle may need to slow down or speed up appropriately to merge into the selected gap so as to keep safe distances between itself and the preceding and following vehicles. Current literature has suggested heuristic and optimization and control methods for the merging decision making for autonomous vehicles . These methods may suffer from laborious parameter tuning due to different environmental conditions. Recent advances in machine learning, deep reinforcement learning in particular, have demonstrated that learning-based systems could out-perform traditional rule-based systems through trial and error in playing board games . Here we study freeway on-ramp merging control for an autonomous vehicle using deep reinforcement learning. We only consider longitudinal control of the autonomous vehicle during merging at a taper entrance ramp. We formalize the on-ramp merging as a reinforcement learning problem with a Markov Decision Process. The states of the Markov Decision Process include the position and velocity information of a fixed number of nearest vehicles at the ramp merging zone. The action is the acceleration value of the autonomous vehicle. The immediate reward includes safe gap-keeping with the preceding and following vehicles after merging, and adds a large negative penalty on collisions related to the merging vehicle. The autonomous vehicle agent is trained in the SUMO (Simulation of Urban Mobility) driving simulator with autonomous traffic running on a freeway with an entrance ramp. We use Deep Deterministic Actor Critic , a state-of-the-art deep reinforcement learning algorithm, to train the autonomous agent. After sufficient training, we found that the trained agent is able to merge onto the freeway with a low collision rate.References: Rios-Torres, Jackeline, and Andreas A. Malikopoulos. "A survey on the coordination of connected and automated vehicles at intersections and merging at highway on-ramps." IEEE Transactions on Intelligent Transportation Systems 18.5 (2017): 1066-1077. Silver, David, et al. "A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play." Science 362.6419 (2018): 1140-1144. Lillicrap, Timothy P., et al. "Continuous control with deep reinforcement learning." arXiv preprint arXiv:1509.02971(2015).
Connected Testbeds for Connected Vehicles
Tulga Ersal, University of Michigan
In this work, we aimed at unifying two once separate concepts for heavy duty vehicles: be reactive to preceding vehicles for safety; and be proactive to elevations for better fuel economy. The two concepts are unified to provide a safe yet fuel efficient connected and automated technology for heavy duty vehicles.We first enhanced the reactive part by establishing a fuel-efficient control algorithm for reacting to preceding vehicles using vehicle-to-vehicle communication, which uses beyond light of site information. This algorithm captures the pattern of local traffic in frequency domain and tune the control parameters for better fuel economy in real-time.Next, we proposed a safety monitor mechanism to determine whether a certain control action would be safe with respect to the preceding vehicles. With this monitor, not only the reactive decision to the preceding vehicles but also the proactive decisions would be verified to ensure the safety, and thus, be allocated for the longitudinal control.Finally, we built a longitudinally by-wire heavy duty vehicle with the fuel-efficient reactive algorithm, the in-production proactive algorithm by previewing elevation, and the safety monitor together, to test and evaluate the overall performance.This project is supported through the DOE Super Truck II program and Navistar, Inc.
Increasing GPS Localization Accuracy with Reinforcement Learning
Ethan Zhang, University of Michigan
We envision that connected testbeds, i.e., remotely accessible testbeds integrated over a network in closed loop, will provide an affordable, repeatable, scalable, and high-fidelity solution for early cyber-physical evaluation of connected automated vehicle (CAV) technologies. Engineering testbeds are critical for empirical validation of new concepts and transitioning new theory to practice. However, the high cost of establishing new testbeds or scaling the existing ones up hinders their wide utilization. This project aims to develop a scientific foundation to support this vision and demonstrate its utility for developing CAV technologies. This application is significant, because a synergistic combination of connected vehicles and automated driving technologies is poised to transform the sustainability of our transportation system; automated driving technologies can leverage the information available from vehicle-to-vehicle (V2V) connectivity in optimal ways to dramatically reduce fuel consumption and emissions. However, state-of-the-art simulation and experimental capabilities fall short of addressing the need for realistic, repeatable, scalable, and affordable means to evaluate new CAV concepts and technologies. The goal of this project is to enable a high-fidelity integration of geographically dispersed powertrain testbeds and use this novel experimental capability to develop and test powertrain-level strategies to increase sustainability benefits of CAVs.This poster will summarize the following accomplishments to date. (i) A predictor framework is under development to compensate for the network delays to increase the fidelity in closed-loop integration of remotely accessible testbeds over the network. Stability boundaries of the predictor have been derived analytically as a function of its design parameters and the network delay, and performance of the predictor has been characterized in the frequency domain. (ii) An optimal vehicle speed management strategy has been created to reduce fuel consumption without violating emissions performance in CAV platooning. Choice of optimization objective is studied to achieve good performance with short prediction horizon. (iii) A method has been developed to release the V2V sequential data in real time under a differential privacy constraint. Fuel consumption and emissions performance degradation due to differential privacy has been quantified in a scenario where the lead vehicle is following the standard FTP-75 drive cycle. (iv) A connected testbed has been created and the proposed predictor framework has been tested using a medium-duty engine with simulated network round-trip delays of 350ms. The fidelity of the connected testbed has been observed to increase significantly when predictors are used compared to the benchmark delayed case without predictors. In particular, the accuracy of the experimental results with predictors are improved by up to 93% compared to the delayed case.
Real-Time Sensor Anomaly Detection and Recovery in Connected Automated Vehicle Sensors
Yiyang Wang, University of Michigan
Automated vehicles are envisioned to be an integral part of the next generation of transportation systems. Whether it is striving for full autonomy or incorporating more advanced driver assistance systems, high-accuracy vehicle localization is essential for automated vehicles to navigate the transportation network safely. In this poster, we propose a reinforcement learning framework to increase GPS localization accuracy. The framework does not make rigid assumptions on the GPS device hardware parameters or motion models, nor does it require infrastructure-based reference locations. The proposed reinforcement learning model learns an optimal strategy to make "corrections'' on raw GPS observations. The model uses an efficient confidence-based reward mechanism, which is independent of geolocation, thereby enabling the model be generalized. We employ a map matching-based regularization term to reduce the variance of the reward return. We construct the reinforcement learning model using the asynchronous advantage actor-critic (A3C) algorithm. A3C provides a parallel training protocol to train the proposed model. The asynchronous reinforcement learning strategy facilitates short training sessions and provides more robust performance. We assess the performance of our model by comparing it with an extended Kalman filter algorithm as a baseline model. Our experiments indicate that the proposed reinforcement learning model converges faster, has less predication variance, and can localize vehicles with 50% less error compared to the baseline Extended Kalman Filter model.
Anomaly detection in CAV sensors is an important but also challenging task. In this study we propose a novel observer-based method that combines signal filtering and anomaly detection together in order to improve the safety of connected and automated vehicles (CAVs). Specifically, we use extended Kalman filter (EKF) to smooth sensor readings based on a nonlinear car-following model, in which we utilize the leading vehicle's information to detect sensor anomalies by using trained One Class Support Vector Machine (OCSVM) models. This approach allows the EKF to estimate the state of a vehicle not only based on the vehicle's location and speed, but also by taking into account the state of the surrounding traffic. A communication time delay factor is considered in the car-following model. Our experiments show that compared with the EKF with a traditional 2-detector, our proposed method achieves a better anomaly detection performance. They also demonstrate that a larger time delay factor has a negative impact on the overall detection performance.