Co-Designing Body and Behavior via Planning-based Hierarchical Grammatical Evolution

An agent’s intelligence tends to benefit from the co-adaption between the body morphology and the behavior policy, aka. embodied intelligence. To create an embodied agent, the joint optimization of the body and the behavior becomes the key problem. To solve that, a popular approach is to use search-based or learning-based heuristics to traverse the morphological space, and optimize a behavior policy for each morphology to evaluate their joint fitness. However, prior works often ignore the relations between body components and modular behaviors, e.g. a mechanical foot can be used to walk, run, kick, etc. To reduce the complexity for embodied intelligence, we come up with a natural idea, that is, to build a mapping from the componentized body to modular behaviors at a lifted abstraction. Based on that mapping, we propose a novel evolutionary computing framework called Hierarchical Grammatical Evolution (HGE). We use the Backus-Naur Form (BNF) to abstract both the body and the behavior space in a modular way, where Behavior Tree (BT) is used to model the behavior policy. Specifically, we leverage the planning-based BT expansion method for further improving the efficacy and efficiency of HGE. Experiments are conducted in a house cleaning scenario and computational test sets, where the results validate the feasibility and scalability of HGE compared with the state of the arts.