Developing probabilistic frameworks for the detailed analysis of insect visual navigation behaviours

蚂蚁、蜜蜂和黄蜂经过漫长而曲折的觅食之旅后，能够成功地找到返回巢穴的路。尽管它们的视觉敏锐度相对较差，神经系统也很简单，但这些令人印象深刻的导航壮举却得以实现，这使它们成为研究的特别有趣的系统。目前大多数昆虫导航模型都倾向于忽略噪声问题。在这个项目中，我们打算构建一个通用的概率框架，用于模拟和分析受噪声影响的真实和合成行为。这个想法源于机器人导航的概率同步定位和映射（SLAM）方法，该方法假设：(1)所有信息和输出（至少在某种程度上）是不确定的，(2)运动命令及其感官结果是强耦合的。通过组合多个不确定信息源，实现了鲁棒的映射性能。在SLAM中，导航代理依赖于其传感器和马达的模型。这些模型包含了当智能体与被感测物体有一定距离时，传感器读数应该是多少，以及它的位置如何响应给定的电机命令而变化。通过这两个模型，智能体可以预测其动作的感官后果，并通过将这些预测与实际传感器读数相结合，获得对自身位置和映射特征的更好估计。通过用昆虫的感觉和运动能力模型代替机器人的测量和运动模型，我们可以把SLAM方法变成分析真实和模拟行为的工具。模型的唯一约束是它们指定了模型输出的概率分布，而不仅仅是单点预测。我们使用的感觉和运动模型可以应用于不同层次的问题。一旦建立，传感器模型和框架为比较不同的导航算法提供了一个标准化的测试环境。在提出的框架中，导航策略可以表示为闭环运动模型，即考虑感官反馈的运动模型。在相同的视觉敏锐度和运动精度约束下，允许实现不同的导航算法，可以进行更有意义的比较。我们还可以通过比较不同运动模式下地标估计位置的不确定性如何变化，来研究运动的精细结构如何影响定位表现。一些运动模式将提供比其他模式更有用的信息，我们可以研究观察到的行为如何与理论上最优的运动相匹配，从而最大限度地减少SLAM框架内的不确定性。在模拟中建立并测试了运动和感觉模型后，它们将被转移到一个装有全景摄像头的大型龙门机器人上，以便对不同的控制算法进行更全面的评估。架式机器人的设置也将使我们能够研究昆虫如何解决数据关联问题，这是机器人导航中的一个突出问题。数据关联指的是识别一个视觉特征之前是否见过的问题。数据关联错误是当前SLAM方法失败的主要原因。为了在真实的昆虫中解决这个问题，我们将记录它们在获取过程中的精细运动结构，然后使用龙门架模拟它们在学习过程中的视觉输入。最后，这些技术为数据采集的自动视频跟踪问题提供了一个潜在的解决方案。上述运动模型可并入视频跟踪系统。通过了解昆虫可能移动到哪里，可以更容易地从一帧到另一帧跟踪代理的位置，从而导致更健壮的跟踪性能。

Ants, bees and wasps are able to successfully find their way back to their nest sites following long and often convoluted foraging trips. The fact that these impressive navigational feats are achieved despite their relatively poor visual acuity and simple nervous systems makes them particularly interesting systems for study. Most current models of insect navigation have tended to ignore problems of noise. In this project we intend to construct a general probabilistic framework for the modelling and analysis of real and synthetic behaviours subject to noise. The idea stems from the probabilistic Simultaneous Localisation and Mapping (SLAM) approach to robot navigation, which assumes that: (1) all information and output is (at least to some degree) uncertain, and (2) motor commands and their sensory consequences are strongly coupled. Robust mapping performance is achieved by combining multiple sources of uncertain information. In SLAM, the navigating agent relies on having models of both its sensors and motors. The models incorporate what a sensor reading should be if the agent is a given distance from a sensed object, and how its position will change in response to a given motor command. With these two models, the agent can predict the sensory consequences of its actions, and, by combining these predictions with the actual sensor readings, obtain a better estimate of its own position and of the mapped features. By replacing the robot measurement and movement models with models of insect sensory and motor capabilities we can turn the SLAM approach into a tool for analysing real and simulated behaviours. The sole constraint on the models is that they specify a probability distribution over the outputs of the model rather than just a single point prediction. The sensory and motor models that we employ can be applied to problems at different levels. Once constructed, the sensor models and framework provide a standardised test environment for comparing different navigation algorithms. In the proposed framework, navigation strategies can be expressed as closed-loop motor models, that is, motor models that take into account sensory feedback. By allowing different navigation algorithms to be implemented, subject to the same constraints on visual acuity and accuracy of movement, more meaningful comparisons can be made. We can also investigate how the fine structure of movements affect localisation performance by comparing how uncertainty in the estimated positions of landmarks changes in response to different motor patterns. Some patterns of movement will provide more useful information than others and we can investigate how closely observed behaviours match theoretically optimal movements that maximally reduce uncertainty within the SLAM framework. Having built and tested the motor and sensory models in simulation, they will be transferred to a large gantry robot fitted with a panoramic camera to allow a more comprehensive evaluation of different control algorithms. The gantry robot set up will also enable us to investigate how insects solve the data-association problem, which is an outstanding issue in robot navigation. Data-association refers to the problem of recognising whether a visual feature has been seen before or not. Errors in data-association are the main reason for failure in current SLAM approaches. To approach this question in real insects, we will record the fine structure of their movements during acquisition and then use the gantry to emulate their visual input during the course of learning. Finally, these techniques provide a potential solution to the problem of automated video tracking for data acquisition. The movement models described above can be incorporated into a video tracking system. By having some idea of where an insect is likely to move to, it is easier to track the agent's position from frame to frame, thereby leading to more robust tracking performance.