S&AS: INT: Inference, Reasoning, and Learning for Robust Autonomous Driving

点击翻译按钮获取中文摘要

While research in autonomous driving has made great strides in recent years, fully autonomous cars are still a distant goal, primarily because of a lack of robustness. Current autonomous cars cannot drive on new roads, or roads that have changed substantially (such as after an earthquake), or when there is a GPS or data outage such as in parking garages, urban cities and tunnels. Importantly, humans are good at all of this: Humans can drive without detailed maps or high precision GPS/IMU sensors, and typically require only a small amount of sparse information for guidance, and their performance typically gets better over time through learning. Using the "intelligent" human driver as a guide, the planned research will develop algorithms that can perceive and make predictions about a scene in real time with measurable confidence, particularly as the scene is closer to the car. New robustness characteristics will be achieved through the ability to detect and overcome mistakes, both in the near term (real time) and long term (learning). The planned algorithms will be designed and validated in a way to enable an inherent robustness not currently available in autonomous driving, and fast adoption by the community. This project is aligned with NSF's Intelligent Physical Systems (IPS) because the algorithms will require cognizant and reflective capabilities in a knowledge-rich environment. Additionally, outputs of this project will impact robotics, machine learning and cyber-physical systems. Educationally, data logs will be disseminated to enable open ended student projects in the community, and undergrad and high school students will collaborate with the research team to integrate sensors, perform experiments and data collection, and disseminate data logs to the community. Led by researchers in Mechanical and Aerospace Engineering, and Computer Science at Cornell University, the goal of this research is to develop, integrate and validate theory and algorithms to enable robust and persistent autonomous driving. This project is aligned with NSF's Intelligent Physical Systems (IPS) because the algorithms will require cognizant and reflective capabilities in a knowledge-rich environment. The technical approach will develop a robust perceptual pipeline for detection, scene estimation, prediction, and anomaly/mistake detection and learning; integrate the algorithms into Cornell's autonomous car software framework and validate the components and system in a series of experimental scenarios to enable their faster adoption by the community. Key component level algorithms to be developed include anytime deep learning detectors with quantifiable performance; multiple hypothesis reasoning with memory attributes; generalized probabilistic anticipation algorithms to mimic a human's mental model of a dynamic scene; and anomaly/mistake detection coupled with online learning. Outcomes will include open source algorithms and data logs; publications, conferences, workshops; data logs for open ended projects in courses and across the community; and undergrad and high school education and diversity programs in the interdisciplinary area of autonomous driving.