Applying Natural Design Principles For Biomimetic Robotic Control

深度学习受到灵长类动物大脑中分层信息处理架构的启发，已被证明是自主机器人控制背景下的强大工具。然而，现代深度学习已经大大偏离了它的生物学起源，然而动物大脑可以解决复杂的任务，具有深度学习系统无法实现的灵活性，鲁棒性和能源效率。这样看来，自主控制架构将从进一步利用支配大脑信息处理架构的自然设计原则中受益。该项目的目的是应用这些原理来开发自主系统的新型控制架构，并将其应用于动物类机器人的控制。自然设计的一个关键原理是约束闭合，即生物系统中在不同时间尺度上运行的动态过程相互促进和维持[1]。限制闭合的一个例子是遗传同化，这是一个通过环境学习的行为表型通过选择压力被遗传编码的过程。类似的动力学在大脑分层控制架构的背景下被描述为脚手架[2]，其中在不同层和时间尺度上操作的过程可以相互约束，以加速对神经电路的有用适应的获得。递归神经网络是大脑动态分层架构的强大抽象，初步的工作表明，脚手架可以用来加速对递归网络拓扑的进化搜索，该拓扑可以解决给定的输入-输出映射[3]。后来人们发现，脚手架可以用来演化复杂的网络状态链，每个状态链都是网络动力学的点吸引子。这一原理可以用来存储机器人动作序列，这样动作就可以存储为吸引子，直到网络受到干扰，例如通过传感器设备检测到的外部信号。博士项目将在这项工作的基础上建立新的认知系统设计和仿生机器人控制的脚手架的实际应用。重要的第一步将是研究最近开发的用于在非对称非二进制神经网络中存储点吸引子的算法，以期在此方向推广脚手架概念，并进一步开发这种方法以利用脚手架对仿生机器人平台进行仿生控制[4]。该项目将理想地利用潜在行业赞助商Opteran开发的自主机器人系统。

Deep learning, which is loosely inspired by the layered information-processing architectures in primate brains, has proven to be a powerful tool in the context of autonomous robotic control. However, modern deep learning has deviated significantly from its biological origins, and yet animal brains can solve complex tasks with a combination of flexibility, robustness, and energy-efficiency that deep-learning systems cannot. It seems then that autonomous control architectures would benefit from further exploiting the natural design principles that govern the information-processing architectures of brains. The aim of this project is to apply such principles to develop novel control architectures for autonomous systems and apply them for the control of animal-like robots.One key principle of natural design is constraint closure, whereby dynamic processes in biological systems operating on different timescales contribute to and maintain each other [1]. One example of constrain closure is genetic assimilation, a process by which environmentally learnt behavioural phenotypes become genetically encoded through selection pressures. Similar dynamics have been described in the context of layered control architectures in brains as scaffolding [2], where processes operating on different layers and timescales can constrain each other to accelerate the acquisition of useful adaptations to neural circuitry.Recurrent neural networks are a powerful abstraction of the dynamic, layered architecture of brains, and preliminary work has shown that scaffolding can be exploited to accelerate the evolutionary search for a recurrent network topology that solves a given input-output mapping [3]. It has since been discovered that scaffolding can be harnessed to evolve complex chains of network states, each a point attractor for the network dynamics. This principle could be harnessed to store sequences of robot actions, such that actions can be stored as attractors that persist until the network is perturbed e.g. by an external signal detected by the sensor apparatus.The PhD project will build on this work to establish new practical applications of scaffolding for the design of cognitive systems and for the control of biomimetic robots. An important first step will be to study a recently developed algorithm for storing point attractors in asymmetrical non-binary neural networks, with a view to generalising the scaffolding concept in this direction, and developing this approach further to harness scaffolding for biomimetic control of biomimetic robot platforms [4].The project will ideally take advantage of the autonomous robotic systems developed by potential industry sponsor, Opteran.