NRI: Small: Collaborative Planning for Human-robot Science Teams

这个项目设想未来的科学探索是人类科学家和自主机器人系统之间的合作。实现这一愿景的关键挑战在于将科学家的专业知识与自主系统的优化能力相结合。科学家带来了专业知识和经验，而自主系统能够处理和评估大量数据。本研究利用这些互补优势，通过将科学家的输入整合到自主学习和规划框架中，开发一个能够指导科学探索和数据收集的协作系统。这是通过将概率规划与逆强化学习相结合，将人类输入和先验知识整合到科学探索背景下的统一优化框架中来实现的。项目团队正在验证自主水下海洋监测这一具有挑战性的领域的方法。由于水下可用的有限通信和必要的监督自治能力，该领域特别适合于人机协作的测试。通过将人类用户的反馈整合到算法规划框架中，目标是提高科学数据收集的效率，并收集以前科学调查范围之外的现象的数据。自动驾驶汽车在科学数据收集方面的应用日益突出；然而，研究界对科学家与自动驾驶汽车之间的相互作用缺乏基本的理解。这项工作的重点是将人类输入整合到算法优化技术中的原则方法，朝着机器人监督自治的目标迈进。通过开发人机科学协作的基础框架，该项目有可能改变收集科学数据的方式。这样的框架有望在整个人机交互和人工智能领域产生广泛的影响。拟议的研究已被纳入研究生和本科生的机器人和计算机科学课程。它还被用于洛杉矶的K-12机器人推广项目。通过与蒙特利湾水族馆研究所（MBARI）和南加州沿海海洋观测系统（SCCOOS）的持续合作，本研究中创建的算法将过渡到现场测试和操作。

This project envisions the future of scientific exploration as a collaborative endeavor between human scientists and autonomous robotic systems. The key challenge to materializing this vision lies in combining the expert knowledge of the scientist with the optimization capabilities of the autonomous system. The scientist brings specialized knowledge and experience to the table, while the autonomous system is capable of processing and evaluating large quantities of data. This research leverages these complementary strengths to develop a collaborative system capable of guiding scientific exploration and data collection by integrating input from scientists into an autonomous learning and planning framework. This is achieved by combining probabilistic planning with inverse reinforcement learning to integrate human input and prior knowledge into a unified optimization framework in the context of scientific exploration. The project team is validating the approach in the challenging domain of autonomous underwater ocean monitoring. This domain is particularly well suited for the testing of human-robot collaboration due to the limited communication available underwater and the necessary supervised autonomy capabilities. By integrating feedback from the human user into an algorithmic planning framework, the goal is to improve the efficiency of scientific data collection and gather data about phenomena that were previously outside the reach of scientific investigation. The use of autonomous vehicles for scientific data collection is becoming increasingly prominent; however, the research community lacks a foundational understanding of the interactions between scientists and autonomous vehicles. This work focuses on principled methods for integrating human input into algorithmic optimization techniques moving towards the goal of supervised autonomy for robots.This project has the potential to change the way scientific data are collected through the development of a foundational framework for human-robot scientific collaboration. Such a framework is expected to have broad implications throughout the fields of human-robot interaction and artificial intelligence. The proposed research is being integrated into the robotics and computer science curriculum at both the graduate and undergraduate levels. It is also being utilized for K-12 robotics outreach programs in Los Angeles. The algorithms created in this research are transitioned to field tests and operations via ongoing collaborations with the Monterey Bay Aquarium Research Institute (MBARI) and the Southern California Coastal Ocean Observing System (SCCOOS).