Collaborative Research: Frameworks: Simulating Autonomous Agents and the Human-Autonomous Agent Interaction

该项目以变革性的方式增强了 Chrono 计算机模拟平台。 Chrono 的目的是通过仿真预测机电系统、其运行环境以及可能与之交互的人类之间的相互作用。该开源仿真平台预计将成为社区共享的虚拟调查工具，用于探索竞争性工程设计并测试通过物理实验验证过于危险、困难或昂贵的假设。 Chrono 已经并将继续应用于多个领域和学科，例如地形力学、天体物理学；软物质物理学；生物力学；机械工业;土木工程;工业工程；和计算机科学。具体来说，它被用来设计 2023 VIPER 月球车；美国陆军专家在评估其轮式和履带式车辆设计时依赖它；在美国和德国用于风力涡轮机行业；并参与欧洲波浪能转换解决方案的设计。项目完成后，Chrono将成为Gazebo的仿真引擎，广泛应用于机器人研究；在美国最大的驾驶模拟器上操作；增强生物机器人和现场机器人社区的研究能力；并协助汽车研究中心旗下的大学和公司联盟在广泛的汽车研究领域开展工作。该项目的教育影响有三方面：以多学科方式培训本科生、研究生和博士后学生，强调高级计算技能的发展；主讲两门关于自动驾驶车辆控制和机器人模拟的新课程；通过威斯康星大学麦迪逊分校校园内的住宿计划扩大对计算机的参与，该计划吸引了来自农村高中的教师和学生。高质量数据推动创新和发现。该项目的核心目标是增加以模拟为来源的数据的份额。在此背景下，由 40 名研究人员组成的多学科团队增强并验证了基于物理的模拟框架，该框架支持自主代理 (AA) 的研究。 AA 在复杂且非结构化的动态环境中运行，并且可能与人类或其他 AA 进行双向交互。该项目使 Chrono 能够快速且廉价地生成机器学习训练数据；便于比较竞争设计以评估权衡；并通过模拟极端情况场景的测试来衡量候选设计的稳健性。这些任务是通过升级和扩展 Chrono 以利用最新的计算动力学创新来完成的，例如更快的索引 3 微分代数方程求解器；解决摩擦接触问题的新方法；用于通过非线性有限元分析处理软机器人中的柔性体动力学的实时求解器；用于地形动力学应用的一流模拟器；依靠即时编译来生成针对问题和硬件进行优化的可执行文件；一种使用混合数据表示来节省状态信息存储的新方法；该奖项反映了 NSF 的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

This project augments the Chrono computer simulation platform in transformative ways. Chrono's purpose is to predict through simulation the interplay between mechatronic systems, the environment they operate in, and humans with whom they might interact. The open-source simulation platform is slated to become a community-shared virtual investigation tool used to probe competing engineering designs and test hypotheses that would be too dangerous, difficult, or costly to verify through physical experiments. Chrono has been and will continue to be used in multiple fields and disciplines, e.g., terramechanics, astrophysics; soft matter physics; biomechanics; mechanical engineering; civil engineering; industrial engineering; and computer science. Specifically, it is used to engineer the 2023 VIPER lunar rover; relied upon by US Army experts in evaluating its wheeled and tracked vehicle designs; used in the US and Germany in the wind turbine industry; and involved in designing wave energy conversion solutions in Europe. Upon project completion, Chrono will become a simulation engine in Gazebo, which is widely used in robotics research; operate on the largest driving simulator in the US; empower research in the bio-robotics and field-robotics communities; and assist efforts in the broad area of automotive research carried out by a consortium of universities and companies under the umbrella of the Automotive Research Center. The educational impact of this project is threefold: training undergraduate, graduate, and post-doctoral students in a multi-disciplinary fashion that emphasizes advanced computing skills development; anchoring two new courses in autonomous vehicle control and simulation in robotics; and broadening participation in computing through a residential program on the campus of the University of Wisconsin-Madison that engages teachers and students from rural high-schools. Innovation and discovery are fueled by quality data. At its core, this project seeks to increase the share of this data that has simulation as its provenance. In this context, a multi-disciplinary team of 40 researchers augments and validates a physics-based simulation framework that empowers research in autonomous agents (AAs). The AAs operate in complex and unstructured dynamic environments and might engage in two-way interaction with humans or other AAs. This project enables Chrono to generate machine learning training data quickly and inexpensively; facilitates comparison of competing designs for assessing trade-offs; and gauges candidate design robustness via testing in simulation of corner-case scenarios. These tasks are accomplished by upgrading and extending Chrono to leverage recent computational dynamics innovations, e.g., a faster index 3 differential algebraic equations solver; a new approach to solving frictional contact problems; a real-time solver for handling flexible-body dynamics in soft robotics via nonlinear finite element analysis; a best-in-class simulator for terradynamics applications; reliance on just-in-time compiling for producing executables that are both problem- and hardware-optimized; a novel way for using mixed data representations for parsimonious storing of state information; and a scalable multi-agent framework that enables geographically-distributed, over the Internet, real-time simulation of human-AA interaction.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.