Learning to walk: Reinforcement Learning for robotic locomotion

学习走路：机器人运动的强化学习

• 批准号: 2745689
• 金额: --
• 依托单位: University of Surrey
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2022
• 资助国家: 英国
• 起止时间: 2022 至 无数据
• 项目状态: 未结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2745689
• 关键词: Learning; walk; Reinforcement; robotic; locomotion

Locomotion is a fundamental skill necessary for any autonomous robot or system which wants to effect significant changes within it's environment. Only the most simplistic production line robotics can operate entirely from a fixed location. However, locomotion is also one of the most varied andchallenging skills to learn in a robust and generalisable way. In nature there are countless examples of different locomotion strategies, and they almost universally require a huge time investment to master. Attempts to go beyond wheeled locomotion in robotics tend to be very brittle. The robot often cannot adapt when the material properties of the terrain are unexpected, or the surface is irregular. This project will aim to address this gap by learning robust and adaptive walking strategies through trial and error reinforcement learning. The project will explore a range of exotic locomotion strategies,from hybrid wheel-legs to starfish-like hexapod robots. In these scenarios the robot will attempt to estimate the properties of the surface it is moving across based on both perception, and it's own locomotion behaviours. It will then use this as conditioning information to inform the following locomotion strategy. The project will specifically explore symbolic-RL as a high efficiency control loop for rapid locomotion, and may also touch on areas of automated robotic design, with regards to learning the optimal locomotion hardware.

运动是任何自动驾驶机器人或系统都希望在其环境中实现重大变化所必需的基本技能。只有最简单的生产线机器人可以完全从固定位置运行。但是，运动也是最多样化和挑战的技能之一，可以以强大而可以普遍的方式学习。在自然界中，有无数不同的运动策略的例子，它们几乎普遍需要大量的时间投资才能掌握。在机器人技术中超越车轮运动的尝试往往非常脆弱。当地形的材料特性出乎意料或表面不规则时，机器人通常无法适应。该项目的目的是通过试用和错误强化学习来学习鲁棒和自适应步行策略来解决这一差距。该项目将探索从混合轮腿到类似海星的六边形机器人的一系列异国情报策略。在这些情况下，机器人将尝试根据两种感知及其自身的运动行为来估计其表面的特性。然后，它将将此用作调理信息，以告知以下运动策略。该项目将专门探索符号-RL作为快速运动的高效率控制回路，并可能涉及自动机器人设计领域，以学习最佳的运动硬件。