Exploiting invisible cues for robot navigation in complex natural environments

对于机器人来说，在自然环境中进行户外导航仍然是一个挑战。机器人导航的最新突破依赖于特定的传感器技术，如激光深度传感器和GPS，以及先进的图像处理。蚂蚁等动物能够在没有这种耗电和计算能力的系统的情况下有效地导航，这证明了另一种更便宜的方法是可行的。蚂蚁也有专门的感知能力，它们的外围视觉系统已经进化到对导航的关键线索非常敏感。具体来说，它们利用不可见（对人类）的光线索，以紫外线（UV）和偏振光探测的形式。紫外线检测可以很容易地区分地平线形状对天空的重要信号。偏振光探测提供了一个相对于太阳方向的外部罗盘提示，即使只有一小部分天空是可见的。我们建议建立一个感官系统，收集蚂蚁在自然生态环境下可以获得的所有光信号，并分析这些信号中包含的信息。我们还将分析特定的传感器布局（ommatidia阵列）、外周受体特征和蚂蚁的运动行为如何有助于提取显著信息。这些数据将形成一个试验台，用于比较蚂蚁导航能力背后的算法和神经模型。迄今为止，这些线索都是单独考虑的，但我们认为，该系统的导航成功取决于具体的组合。例如，偏光天空中的方向信息可能是视觉记忆的重要组成部分；和紫外线信息可能有助于消除极化模式的歧义，以及在不同云条件下该信息的稳健性。在过去的几年里，对昆虫神经通路的研究有了实质性的增加，这些神经通路涉及处理这些线索，但尚未在机器人模型中得到利用。特别是在解码偏振光以获得方向的中枢脑机制方面已经取得了突破性的进展。人们对学习回路的理解也在迅速增加，这是导航能力的关键组成部分。与许多受生物启发的项目相比，我们的方法的一个关键方面是，我们不打算从设计和制造一个尽可能与蚂蚁眼睛相似的传感器开始。相反，我们认为有必要首先了解眼睛的哪些方面实际上是支持导航的关键，以及精确的生物设计的哪些变化将在简化制造的同时保留理想的特性。本研究的最终成果将是一个小型、低功耗的户外导航视觉系统的详细设计。我们假设我们可以利用CCD相机的小型化（通常对紫外线的上范围敏感，但有过滤器来去除它）来组成一个全向复眼，其中每个小眼的过滤器特性适当地调整波长或偏振平面，模仿蚂蚁的眼睛。除了在自然环境中直接应用于机器人导航，在农业、搜救和环境清理等任务中，这种设计还具有开发其他应用的潜力，包括移动设备、检查和监视。

Outdoor navigation in natural environments remains a challenge for robotics. Recent breakthroughs in robot navigation have been dependent on specific sensor technologies, such as laser depth sensors and GPS, and advanced image processing. The ability of animals such as ants to navigate effectively without such power- and computation- hungry systems are a proof of principle that alternative cheaper approaches are viable. Ants also have specialised sensing, with a peripheral visual system that has evolved to be sensitive to crucial cues for navigation. Specifically, they make use of non-visible (to humans) light cues in the form of ultraviolet (UV) and polarised light detection. UV detection allows the important signal of the horizon shape against the sky to be easily distinguished. Polarised light detection provides an external compass cue of heading relative to the sun direction, even when only a small portion of the sky is visible. We propose to build a sensory system that gathers the full range of light cues available to the ant, in its natural ecological situation, and to analyse the information contained in this signal. We will also analyse how the specific sensor layout (ommatidia array), peripheral receptor characteristics, and the motor behaviour of the ant may contribute to extracting salient information. The data will form a test-bed for comparison of algorithmic and neural models of the processing that underlies the navigation capabilities of the ant. To date, these cues have been considered separately but we believe the navigational success of this system depends on the specific combination. For example, the directional information in the polarised sky may form an important part of visual memories; and UV information may contribute to disambiguation of the polarisation pattern and the robustness of this information under different cloud conditions. There has been a substantial increase in the last few years in research into insect neural pathways involved in processing these cues which has yet to be exploited in robot models. In particular there has been breakthrough work on the central brain mechanisms involved in decoding polarised light to obtain heading direction. There is also a rapidly increasing understanding of the circuits involved in learning, a key component of navigation capabilities. A key aspect of our method, in comparison to many bio-inspired projects, is that we do not propose to start by designing and manufacturing a sensor as similar as possible to the ant eye. Rather, we consider it essential to first understand what aspects of the eye are actually key to support navigation, and what variation from the exact biological design will conserve the desirable properties while simplifying the manufacture. The final outcome of this research will be a detailed design for a small scale and low power vision system for outdoor navigation. We hypothesis that we can leverage the miniaturisation of CCD cameras (which are normally sensitive to the upper range of UV, but have filters applied to remove it) to compose an omnidirectional compound eye where the filter characteristics of each ommatidia are appropriately tuned for wavelength or polarisation plane, in imitation of the ant eye. Beyond the direct application to robot navigation in natural environments, in tasks such as agriculture, search and rescue, and environmental clean-up, such a design has potential to be developed for other applications, including mobile devices, inspection and surveillance.