CRII: RI: Lyapunov-Certified Cognitive Control for Safe Autonomous Navigation in Unknown Environments

在运输、配送、农业、环境监测和建筑等领域，需要在未知、混乱和动态变化的环境中自主导航的无人驾驶飞机和地面车辆的应用越来越多。为了实现安全、弹性和自我改进的自主导航，该项目专注于自适应在线环境理解的设计，以确保在具有挑战性的条件下稳定、无碰撞地运行。这项提议的研究很重要，因为目前的实践依赖于先前的地图或手工制作的在线地图，这些地图试图捕捉整个环境，即使某些部分与特定的导航任务无关。这增加了内存和计算需求，扩大了噪声的影响，并使当前的方法变得脆弱，特别是在涉及动态障碍物、不可靠的定位或光照变化的情况下。该提案提供了两项技术创新，以实现安全的自主导航。首先，它开发了一个可学习的神经地图，基于三维卷积的层次（八叉树）空间划分来提取导航特定的特征，并基于可微分记忆来推断特征之间的长期依赖关系。从导航经验中训练神经地图参数不是为了生成精确的地图，而是为了准确地量化预期运动轨迹的碰撞概率。第二个创新是李雅普诺夫理论控制方法，该方法利用自治系统相对于虚拟运动系统（可以立即停止）的总能量来推导保证神经网络提出的轨迹跟踪稳定和无碰撞的条件。提出的可学习神经地图显著提高了碰撞预测的鲁棒性，而李雅普诺夫理论控制保证了在新的、不可预测的和混乱的环境中稳定和安全的导航。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。

Applications for unmanned aerial and ground vehicles requiring autonomous navigation in unknown, cluttered, and dynamically changing environments are increasing in fields such as transportation, delivery, agriculture, environmental monitoring, and construction. To achieve safe, resilient, and self-improving autonomous navigation, this project focuses on the design of adaptive online environment understanding that guarantees stable and collision-free operation in challenging conditions. The proposed research is important because current practices rely on prior maps or hand-crafted online mapping that attempt to capture the whole environment, even if parts are irrelevant for specific navigation tasks. This increases memory and computation requirements, spreads the effects of noise, and makes current approaches brittle, particularly in conditions involving dynamic obstacles, unreliable localization, or illumination variation.The proposal offers two technical innovations to achieve safe autonomous navigation. First, it develops a learnable neural map based on 3-D convolution over hierarchical (octree) partitioning of space to extract navigation-specific features and on differentiable memory to infer long-term dependence among the features. The neural map parameters are trained from navigation experience not to produce accurate maps but to quantify the collision probabilities of intended motion trajectories accurately. The second innovation is a Lyapunov-theoretic control approach that uses the total energy of an autonomous system with respect to a virtual kinematic system (that can stop immediately) to derive conditions that guarantee stable and collision-free tracking of the trajectories proposed by the neural network. The proposed learnable neural map significantly increases the robustness of collision prediction, while the Lyapunov-theoretic control guarantees stable and safe navigation in new, unpredictable, and cluttered environments.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.