Neuromorphic Sensorimotor Integration for Legged Locomotion (NSILL)

在行走过程中，机器人必须使用几种感官来指导其移动。例如，它可能需要避开障碍物，抬起腿越过颠簸，或者采取更大的步伐来避免缺口。基本反射对直接接触(或缺乏接触)不平坦的地面做出反应，但行走可以通过提前感知地面变化或障碍物来帮助。从字面上看未来会发生什么，这将有所帮助。在活体动物中，行走模式通常依赖于中央模式生成器(CPG)来生成规则的步态。然而，CPG的输出必须学会控制在复杂和陌生的地形上行走-它必须使用动物的感官来调整行走模式。或者，这些模式可能出现在分布式控制中，在分布式控制中，每一条腿都会影响到它的邻居。由于复杂的地形，当“图案”变得非常不规则时，这种方法可能是有利的。这些对行走问题的生物学解决方案为行走机器人提供了工程解决方案。我们将使用模拟/数字硅片设计和建造生物灵感系统，控制一个六条腿的机器人。研究有三条主线。1)学习生成自适应行走模式的CPG芯片。生物CPG非常灵活，可以产生不同的节奏。神经调节器、中枢指令和输入信号都会影响CPG产生的模式。它们通过改变单个神经元的电学和化学性质以及不同神经元组之间的耦合来做到这一点。例如，视觉输入可以使一只动物突然奔跑。我们之前已经开发出一种CPG电路，能够产生广泛的动物步态。现在，我们将在硅中建造这种新型的CPG，并在项目期间设计和建造的一个类似昆虫的6条腿机器人上进行测试。然后，我们将评估其产生足够灵活的输出以处理复杂地形的能力，并将其与更分布式的控制系统进行比较，或者与这两种形式的步行控制的适当组合进行比较。2)一种视觉芯片，用于估计图像中对象的距离。神经形态视觉传感器已经被建造用于边缘检测、速度和深度传感。然而，这些都是孤立的案例研究，没有将视觉与其他感官结合起来，用于机器人应用。我们将设计一种硅芯片，实现一种新的视觉深度感知方法，该方法是在之前的一项拨款下开发的，基于尖峰神经元模型。这个时尚的早期视觉模型将由一个2-D光敏晶体管阵列组成，能够执行边缘检测和深度检测。3)一种受生物启发的芯片，可以整合感官做出决定。这种芯片将把视觉处理器的输出(检测到的物体及其深度)与额外的感官(触摸和关节角度传感器)结合起来，调整行走模式，以实现在崎岖地面上平稳、平稳的移动。特别是，它将学习使用视觉深度信息来预测避免碰撞和保持稳定立足点所需的动作。我们将在模拟中测试这些方法，然后使用微处理器上的软件在真实的机器人上测试结果。然后，我们将创建一种芯片，将传感器的世界观转换为对行走控制器的适当调制。该项目的最后阶段将是将所有三个芯片集成到6条腿的机器人上，以生产一种具有明确生物前兆的新步行机器人。因此，该项目将推进神经形态超大规模集成电路和传感器-电机集成的研究，直接应用于移动机器人，用于家庭应用和在危险、困难和未知环境中工作。它还将有助于识别神经系统中的一些基本计算和控制原理。

During walking, a robot must use several senses to guide its movements. For example, it may need to avoid obstacles, to lift its legs over bumps, or to take a longer stride to avoid a gap. Basic reflexes respond to direct contact (or lack of contact) with uneven ground, but walking would be aided by sensing the ground surface variation or obstacles in advance. It would help to literally see what is ahead . In living animals, walking patterns often depend on a Central Pattern Generator (CPG) to generate a regular gait. The CPG output must, however, learn to control walking on complex and unfamiliar terrain - it must use the animal's senses to adjust walking patterns. Alternatively, the patterns might emerge from distributed control in which each leg influences its neighbours. This approach might be advantageous when the 'pattern' becomes very irregular due to complex terrain.These biological solutions to the problem of walking suggest engineering solutions for walking robots. We will design and build biologically inspired systems using analogue/digital silicon chips, controlling a 6-legged robot. There are three research strands. 1) A CPG chip that learns to generate adaptive walking patterns. Biological CPGs are extremely flexible for producing different rhythms. Neuromodulators, central commands and input signals all influence the pattern produced by a CPG. They do so by altering both the electrical and chemical properties of individual neurons and the coupling between different groups of neurons. For example, visual inputs can cause an animal to break into a gallop. We have previously developed a CPG circuit capable of producing a wide range of animal gaits. We will now build this novel CPG in silicon and test it on a 6-legged insect-like robot, designed and built during the project. We will then evaluate its ability to produce sufficiently flexible output to deal with complex terrain, and compare it to a more distributed control system, or to a suitable hybrid of these two forms of walking control.2) A vision chip that estimates the distance to objects in an image. Neuromorphic vision sensors have been built for edge detection, velocity- and depth-sensing. These are, however, isolated case studies that do not integrate vision with other senses for robotic applications. We will design a silicon chip that implements a novel vision depth sensing method, developed under a previous grant and based on a spiking neuronal model. This stylised early vision model will consist of a 2-D array of light-sensitive transistors with the ability to perform both edge sensing and depth detection. 3) A biologically-inspired chip that integrates senses to make decisions. This chip will combine the output of the visual processor (detected objects and their depth) with additional senses (touch and joint angle sensors) to adjust the walking patterns to achieve smooth, stable movement across rough ground. In particular it will learn to use the visual depth information to anticipate movements required to avoid collisions and maintain stable footing. We will test the methods in simulation, then use software on microprocessors to test the results on real robots. We will then create a chip which will translate the sensors' view of the world into appropriate modulation of the walking controller. The final stage of the project will be the integration of all three chips on to the 6-legged robot to produce a new walking robot with explicit biological antecedents. The project will thus advance research in neuromorphic VLSI and sensor-motor integration, with direct applications in mobile robots for domestic applications and work in hazardous, difficult and unknown environments. It will also help to identify some of the basic computing and control principles in the nervous system.