Uncertainty Modeling of Learning to Enable Probabilistic Perception

现代预测模型（例如，神经网络）已经彻底改变了从商业分析到机器人技术的应用。大部分的发展都集中在通过广泛的数据收集和架构来提高平均预测精度上。然而，许多这些模型的一个关键弱点是，它们只是提供一个输出，没有信心或结果的准确性。这些模型的预测精度可以根据训练数据的数量和多样性、模型体系结构和测试环境而变化。例如，视觉定位模型在恶劣天气下会退化，而物体探测器在物体可能被遮挡的环境中表现更差。然而，在这两种情况下，通常都无法衡量预测的准确性。然而，其他类型的预测模型确实在其输出中包含对准确性的度量。一种成功的方法已经在概率感知算法中显示出来，该算法已经处理不确定的输出，包括异常值，很多年了。GPS导航系统必须在来自建筑物的多路径误差存在的情况下工作，雷达跟踪传感器必须在杂波导致的许多误报返回的情况下工作。由于存在捕获大多数误差的不确定性模型，使用这些传感器的感知算法已经开发成功。该项目将开发新的预测模型，将深度学习输出与概率感知/推理算法相结合，以提高机器人在不确定环境中导航的能力。这项研究将开发深度学习输出的正式不确定性模型，结合概率感知/推理算法，为关键的机器人功能（如定位、跟踪和预测）创建整体的、高完整性的框架。该项目充分利用了深度学习（快速准确地处理大量数据）和感知（对错误和错误进行推理）的优势。通过考虑三种不同类别的学习和感知问题（定位、跟踪、预测），这项工作将更容易过渡到其他应用领域，如家庭、仓库、繁忙的酒店/博物馆/火车站和仓库中的机器人；以及空中和水下交通工具。该项目将向社区提供数据集和软件，研究结果将通过出版物、会议、课程、针对行业的讲习班和项目负责人会议传播。将实施扩大参与的综合计划，包括接待代表性不足的学生，与高中生和代表性不足的学生合作，以及对本科生和研究生进行培训和指导。该项目由跨部门机器人基础研究项目支持，由工程（ENG）和计算机与信息科学与工程（CISE）联合管理和资助。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Modern prediction models (e.g., neural networks) have revolutionized applications ranging from business analysis to robotics. Much of the development has focused on increasing average prediction accuracy via extensive data collections and architectures. However, a key weakness of many of these models is that they simply provide an output, with no sense of the confidence or the accuracy of the result. Prediction accuracy of these models can vary based on the amount and diversity of the training data, the model architecture, and the test environment. For example, visual localization models degrades in poor weather, and object detectors do more poorly in environment in which objects can be obscured. Yet in either example, there is typically no measure of accuracy for the predictions. However, other types of prediction models do include measures of the accuracy in their outputs. A success approach has been shown in the probabilistic perception algorithms, which have been handling uncertain outputs, including outliers, for many years. GPS navigation systems must function in the presence of multi-path errors from buildings, and radar tracking sensors must function even with the return of many false positives due to clutter. Successful perception algorithms using these sensors have been developed because there exists uncertainty models capturing most errors. This project will develop new predictions models that merge deep learning outputs with probabilistic perception/reasoning algorithms to improve the ability of robots to navigate in uncertain environments. This research will develop formal uncertainty models of deep learning outputs in combination with probabilistic perception/reasoning algorithms to create holistic, high-integrity frameworks for key robotics functions such as localization, tracking and forecasting. This project leverages the best of both deep learning (processing large amounts of data quickly and accurately) and perception (reasoning over errors and mistakes). By considering three different classes of learning and perception problems (localization, tracking, forecasting), the work will more easily transition to other application domains such as robots in the home, warehouse, busy hotels/museums/train stations, and warehouses; and aerial and underwater vehicles. The project will make datasets and software available to the community, and research results will be disseminated through publications, conferences, courses, industry-targeted workshops and PI meetings. A comprehensive plan for broadening participation will be implemented including hosting under-represented students, working with high school and under-represented students, and training and mentoring undergraduate and graduate students.This project is supported by the cross-directorate Foundational Research in Robotics program, jointly managed and funded by the Directorates for Engineering (ENG) and Computer and Information Science and Engineering (CISE).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.