Search Results

While machine learning in autonomous vehicles development has increased significantly in the past few years, the use of reinforcement learning (RL) methods has only recently been applied. Convolutional Neural Networks (CNNs) became common for their powerful object detection and identification and even provided end-to-end control of an autonomous vehicle. However, one of the requirements of a CNN is a large amount of labeled data to inform and train the neural network. While data is becoming more accessible, these networks are still sensitive to the format and collection environment which makes the use of others’ data more difficult. In contrast, RL develops solutions in a simulation environment through trial and error without labeled data. Our research expands upon previous research in RL and Proximal Policy Optimization (PPO) and the application of these algorithms to 1/18th scale cars by expanding the application of this control strategy to a full-sized passenger vehicle.

Prediction of vehicle velocity is important since it can realize improvements in the fuel economy/energy efficiency, drivability, and safety. Velocity prediction has been addressed in many publications. Several references considered deterministic and stochastic approaches such as Markov chain, autoregressive models, and artificial neural networks. There are numerous new sensor and signal technologies like vehicle-to-vehicle and vehicle-to-infrastructure communication that can be used to obtain inclusive datasets. Using these inclusive datasets of sensors in deep neural networks, high accuracy velocity predictions can be achieved. This research builds upon previous findings that Long Short-Term Memory (LSTM) deep neural networks provide low error velocity prediction. We developed an LSTM deep neural network that uses different groups of datasets collected in Fort Collins, Colorado.

The EcoCAR3 competition challenges student teams to redesign a 2016 Chevrolet Camaro to reduce environmental impacts and increase energy efficiency while maintaining performance and safety that consumers expect from a Camaro. Energy management of the new hybrid powertrain is an integral component of the overall efficiency of the car and is a prime focus of Colorado State University’s (CSU) Vehicle Innovation Team. Previous research has shown that error-less predictions about future driving characteristics can be used to more efficiently manage hybrid powertrains. In this study, a novel, real-world implementable energy management strategy is investigated for use in the EcoCAR3 Hybrid Camaro. This strategy uses a Nonlinear Autoregressive Artificial Neural Network with Exogenous inputs (NARX Artificial Neural Network) trained with real-world driving data from a selected drive cycle to predict future vehicle speeds along that drive cycle.

Autonomous vehicle development has benefited from sanctioned competitions dating back to the original 2004 DARPA Grand Challenge. Since these competitions, fully autonomous vehicles have become much closer to significant real-world use with the majority of research focused on reliability, safety and cost reduction. Our research details the recent challenges experienced at the 2017 Self Racing Cars event where a team of international Udacity students worked together over a 6 week period, from team selection to race day. The team’s goal was to provide real-time vehicle control of steering, braking, and throttle through an end-to-end deep neural network. Multiple architectures were tested and used including convolutional neural networks (CNN) and recurrent neural networks (RNN). We began our work by modifying a Udacity driving simulator to collect data and develop training models which we implemented and trained on a laptop GPU.

Accurate perception of the driving environment and a highly accurate position of the vehicle are paramount to safe Autonomous Vehicle (AV) operation. AVs gather data about the environment using various sensors. For a robust perception and localization system, incoming data from multiple sensors is usually fused together using advanced computational algorithms, which historically requires a high-compute load. To reduce AV compute load and its negative effects on vehicle energy efficiency, we propose a new infrastructure information source (IIS) to provide environmental data to the AV. The new energy–efficient IIS, chip–enabled raised pavement markers are mounted along road lane lines and are able to communicate a unique identifier and their global navigation satellite system position to the AV. This new IIS is incorporated into an energy efficient sensor fusion strategy that combines its information with that from traditional sensor.

Connected and Automated Vehicles (CAV) provide new prospects for energy-efficient driving due to their improved information accessibility, enhanced processing capacity, and precise control. The idea of the Eco-Driving (ED) control problem is to perform energy-efficient speed planning for a connected and automated vehicle using data obtained from high-resolution maps and Vehicle-to-Everything (V2X) communication. With the recent goal of commercialization of autonomous vehicle technology, more research has been done to the investigation of autonomous eco-driving control. Previous research for autonomous eco-driving control has shown that energy efficiency improvements can be achieved by using optimization techniques. Most of these studies are conducted through simulations, but many more physical vehicle integrated test application studies are needed.

Practical applications of recently developed sensor fusion algorithms perform poorly in the real world due to a lack of proper evaluation during development. Existing evaluation metrics do not properly address a wide variety of testing scenarios. This issue can be addressed using proactive performance measurements such as the tools of resilience engineering theory rather than reactive performance measurements such as root mean square error. Resilience engineering is an established discipline for evaluating proactive performance on complex socio-technical systems which has been underutilized for automated vehicle development and evaluation. In this study, we use resilience engineering metrics to assess the performance of a sensor fusion algorithm for vehicle localization. A Kalman Filter is used to fuse GPS, IMU and LiDAR data for vehicle localization in the CARLA simulator.

Development of autonomous vehicle technology is expensive and perhaps more complicated than initially thought, as evidenced by the recent rollback of anticipated delivery dates from companies such as Tesla, Waymo, GM, and more. One of the most effective techniques to reduce research and development costs and speed up implementation is rigorous analysis through simulation. In this paper, we present multiple autonomous vehicle perception and control strategies that are rigorously investigated in the user friendly, free, and open-source simulation environment, CARLA. Overall, we successfully formulated potential solutions to the autonomous navigation problem and assessed their advantages and disadvantages in simulation at no cost. First, a lane finding method utilizing polynomial fitting and machine learning is proposed. Then, the waypoint navigation strategy is described, along with route planning. Object detection is then implemented using pre-trained convolutional neural networks.

The transportation sector contributes significantly to emissions and air pollution globally. Emission models of modern vehicles are important tools to estimate the impact of technologies or controls on vehicle emission reductions, but developing a simple and high-fidelity model is challenging due to the variety of vehicle classes, driving conditions, driver behaviors, and other physical and operational constraints. Recent literature indicates that neural network-based models may be able to address these concerns due to their high computation speed and high-accuracy of predicted emissions. In this study, we seek to expand upon this initial research by utilizing several deep neural networks (DNN) architectures such as a recurrent neural network (RNN) and a convolutional neural network (CNN). These DNN algorithms are developed specific to the vehicle-out emissions prediction application, and a comprehensive assessment of their performances is done.

There is a pressing need to develop accurate and robust approaches for predicting vehicle speed to enhance fuel economy/energy efficiency, drivability and safety of automotive vehicles. This paper details outcomes of research into various methods for the prediction of vehicle velocity. The focus is on short-term predictions over 1 to 10 second prediction horizon. Such short-term predictions can be integrated into a hybrid electric vehicle energy management strategy and have the potential to improve HEV energy efficiency. Several deterministic and stochastic models are considered in this paper for prediction of future vehicle velocity. Deterministic models include an Auto-Regressive Moving Average (ARMA) model, a Nonlinear Auto-Regressive with eXternal input (NARX) shallow neural network and a Long Short-Term Memory (LSTM) deep neural network. Stochastic models include a Markov Chain (MC) model and a Conditional Linear Gaussian (CLG) model.

Heavy-duty commercial vehicles consume a significant amount of energy due to their large size and mass, directly leading to vehicle operators prioritizing energy efficiency to reduce operational costs and comply with environmental regulations. One tool that can be used for the evaluation of energy efficiency in heavy-duty vehicles is the evaluation of energy efficiency using vehicle modeling and simulation. Simulation provides a path for energy efficiency improvement by allowing rapid experimentation of different vehicle characteristics on fuel consumption without the need for costly physical prototyping. The research presented in this paper focuses on using real-world, sparsely sampled telematics data from a large fleet of heavy-duty vehicles to create high-fidelity models for simulation. Samples in the telematics dataset are collected sporadically, resulting in sparse data with an infrequent and irregular sampling rate.

Traditional autonomous vehicle perception subsystems that use onboard sensors have the drawbacks of high computational load and data duplication. Infrastructure-based sensors, which can provide high quality information without the computational burden and data duplication, are an alternative to traditional autonomous vehicle perception subsystems. However, these technologies are still in the early stages of development and have not been extensively evaluated for lane detection system performance. Therefore, there is a lack of quantitative data on their performance relative to traditional perception methods, especially during hazardous scenarios, such as lane line occlusion, sensor failure, and environmental obstructions.