Swarm exploration: multi-robot visual navigation through insect-inspired strategies

这个跨学科项目将成为EPSRC资助的Brains on Board（BoB; brainsonboard.co.uk）计划资助的一部分，这是一个多大学项目，我们的目标是创造具有蜜蜂学习能力的机器人。机器人穿越路线所经历的视图用于训练人工神经网络（ANN），以学习路线视图熟悉度的紧凑编码。一旦经过训练，该网络就会估计新的视图-以及新的姿势-以前是否经历过。对于路径引导，机器人移动，采样不同的视图，并在熟悉的方向，如指定的人工神经网络的头部。由于路径知识编码在一个人工神经网络的权重，它可以很容易地在机器人之间传输。然而，为了有用，信息需要以这样一种方式编码，即它独立于特定机器人的视角，并且还需要仔细组合来自多个机器人的知识。我们的项目包括三个不同的项目阶段。阶段1：自然环境下的视角不变路线编码[12个月]。全景图像保持冗余信息，用于在低频分量中恢复方位角方向，低频分量可能是透视不变的（例如，蚂蚁使用来自天际线轮廓的视觉特征）。我们将探索视角不变的视觉处理（通过视觉神经科学的灵感），并将IMU信息纳入路线学习和导航算法。第二阶段：无人机到轮式机器人的转移[12-24个月]。第二年将对第一阶段开发的算法进行现场测试，以便在越来越困难的现场测试中实现无人机和轮式机器人之间的导航信息传输。第三阶段：路线组合[24-36个月]。最后，我们将研究从多个机器人从一个集线器探索合并路线知识的方法。为了做到这一点，我们将把嘈杂的位置估计（从里程计或GPS），以建立'辐条'与估计的位置到地形图的形式。这将包括一些元学习和排练，以测试路由辐条

This interdisciplinary project will form part of the EPSRC funded Brains on Board (BoB; brainsonboard.co.uk) program grant, a multi-university project in which we aim to create robots with the learning abilities of bees. The views experienced as a robot traverses a route are used to train an artificial neural network (ANN) to learn a compact encoding of the familiarity of route views. Once trained, the network estimates whether new views - and thus new poses - have been experienced before. For route guidance, robots move, sampling different views, and head in familiar directions, as specified by the ANN. As the route knowledge is encoded in the weights of an ANN, it can be easily transmitted between robots. However, to be useful, the information needs to be encoded in such a way that it is independent of that particular robot's perspective and knowledge also needs to be carefully combined from multiple robots. We project entails three distinct project stages.Stage 1: Perspective-invariant route encodings for natural environments [12 months]. Panoramic images hold redundant information for recovering azimuthal directions in low-frequency components that can potentially be perspective-invariant (e.g. ants use visual features from the skyline profile). We will explore perspective invariant visual processing (through inspiration from visual neuroscience) and incorporate IMU information to route learning and navigation algorithms.Stage 2: UAV to wheeled robot transfer [12-24 months]. The 2nd year will field-test the algorithms developed in Stage 1 to allow transfer of navigation information between a UAV and a wheeled robot in field-tests of increasing difficulty.Stage 3: Route combination [24-36 months]. We will finally investigate methods for amalgamating route knowledge from multiple robots exploring from a hub. To do this, we will incorporate noisy positional estimates (from odometry or GPS) to build up 'spokes' with estimated positions into a form of topographic map. This will incorporate some meta-learning and rehearsal to test route-spokes