Concurrent morphology-optimization and behavior-learning: Co-designing intelligent quadcopters

In this paper, a novel co-design process is proposed to accomplish the coupled objectives of maximizing the flight range and learning capacity of a quadcopter unmanned aerial vehicle (UAV), with the end goal of producing a UAV that can learn to fly bottoms up (pun intended). The approaches presented here are motivated by the observed coherence in the evolution of morphology and behavior (or rather the capacity to learn behavior) in nature. The co-design process requires concurrent execution of the following two processes: i) optimization of the UAV morphology (e.g., geometry, component choices, etc.) and ii) learning of the flight controller to achieve take-off and preliminary hovering. In order to decrease the computational burden of applying the learning process nested under each morphology design, a novel metric called “Talent” is defined, which estimates the learning capacity of the system based on the UAV morphology and control theory. A concept of “Capacity Variables” is also introduced to help in reducing the dimension of the design variable space, and enable swift inverse design. Multi-objective optimization and neuroevolution processes are respectively used to perform the multi-level design and validate the learning capacity of the UAV designs produced thereof. The results of our co-design framework demonstrate the trade-offs accomplished in terms of UAV range performance and talent. More importantly, we show that through coherent morphology-talent optimization, it is possible to produce UAVs that can learn to fly (in this case, take off and hover briefly) with no prior guidance, which is in itself a major step forward in designing autonomous aerial robotic systems.