EAGER: Uncertainty-aware Planning for Robot Navigation in Human Environments

从服务机器人到自动驾驶汽车，机器人正越来越多地融入社会。然而，为了既安全又高效，它们需要更好地了解人们如何在非结构化环境中导航，并规划考虑到人们自己的路径的路径。本项目中开发的规划方法将生成满足双重目标的机器人路径，即最小化机器人与人碰撞的风险，同时最小化其预期任务完成时间。这项工作的独特之处在于，机器人动态地估计可预测的行人在任何时间点的行为，并相应地调整其计划，在不可预测的情况下自动选择更安全、更保守的路径，当情况更可预测时强调效率。机器人在人类环境中自主规划和执行安全、高效的路径的能力将提高它们与人类合作的能力，特别是在拥挤的情况下，如在工厂、学校、医院或商场。开发的导航方法遵循感觉-计划-预测-动作循环，在该循环中，通过找到避免与预期未来可能的障碍物状态分布相撞的中长期轨迹来计算机器人计划。这项工作的目标分为两个任务：第一个任务是通过使用不确定性估计来增加预测，从而改进现有的行人路径建模技术。这是通过迭代最大似然框架实现的。第二个任务是在考虑机器人动力学的同时，寻找路径成本函数最小的机器人轨迹，同时考虑控制成本和碰撞成本。这是使用迭代的、基于梯度的轨迹优化方法来完成的。通过仿真和实际行人数据对所提出的导航方法进行了评估。

Robots are increasingly being integrated into society, from service robots to autonomous cars. To be both safe and efficient, however, they need to better understand how people navigate in unstructured environments and plan paths that take people's own paths into consideration. The planning methods developed in this project will generate robot paths that satisfy the dual goals of minimizing the robot's risks of collision with people while simultaneously minimizing its expected task completion time. Unique to this work, the robot dynamically estimates how predictable pedestrians are behaving at any point in time and adapts its plans accordingly, automatically taking safer, more conservative, paths when in less predictable situations, and emphasizing efficiency when the situation is more predictable. The ability for robots to autonomously plan and execute safe, efficient paths in human environments will improve their ability to operate in collaboration with humans, especially in crowded situations, such as in factories, schools, hospitals, or malls.The navigation method developed follows a sense-plan-predict-act loop where robot plans are computed by finding medium-to-long term trajectories that avoid collision with an expected future distribution of likely obstacle states. The goals of this work split into two tasks: The first improves existing pedestrian path-modeling techniques by augmenting the predictions with estimates of uncertainty. This is accomplished through an iterative maximum-likelihood framework. The second task finds robot trajectories that minimize a path-cost function, accounting for both control and collision costs, while respecting robot dynamics. This is done using an iterative, gradient-based trajectory optimization approach. The resulting navigation approach is evaluated in both simulation and on real pedestrian data.