Autonomous Robot Evolution (ARE): Cradle to Grave

点击翻译按钮获取中文摘要

Robotics is changing the landscape of innovation. But traditional design approaches are not suited to novel or unknown habitats and contexts, for instance: robot colonies for ore mining, exploring or developing other planets or asteroids, or robot swarms for monitoring extreme environments on Earth. New design methodologies are needed that support optimising robot behaviour under different conditions for different purposes. It is accepted that behaviour is determined by a combination of the body (morphology, hardware) and the mind (controller, software). Embodied AI and morphological computing have made major progress in engineering artificial agents (i.e., robots) by focusing on the links between morphology and intelligence of natural agents (i.e., animals). While such a holistic body-mind approach has been hailed for its merits, we still lack an actual pathway to achieve this.While this goal is ambitious, it is achievable by introducing a unique methodology: a hybridisation of the physical evolutionary system with a virtual one. On the one hand, it is appreciated that an effective design methodology requires the use and testing of physical robots. This is because simulations are prone to hidden biases, errors and simplifications in the underlying models. Simulating populations of robots (rather than just simulating specific parts) leads to accumulated errors and a lack of physical plausibility: the evolved designs will not work in the real system. This is the notorious reality gap of evolutionary robotics. On the other hand, evolving everything in hardware is time and resource consuming. One of our major innovations is to run simulated evolution concurrently with the physical and hybridise them by cross-breeding, where a physical and a virtual robot can parent a child that may be born in the real world, in the virtual world or in both. The advantages of such a hybrid system are significant. Physical evolution is accelerated by the virtual component that can run faster to find good robot features with less time and resources; simulated evolution benefits from the influx of genes that are tested favourably in the real world. Furthermore, monitoring of and feedback from the physical system can improve the simulator, reducing the reality gap.