Abstract:
In this work, a collaborative co-evolution approach is adopted to solve a joint physical design and feedback control optimization problem of a nature-inspired Unmanned Ae...Show MoreMetadata
Abstract:
In this work, a collaborative co-evolution approach is adopted to solve a joint physical design and feedback control optimization problem of a nature-inspired Unmanned Aerial Vehicle (UAV). Unlike traditional multirotors and fixed-wing aircraft, lift is achieved by spinning its entire body with attached aerofoils around a central axis and positional control is attained through regulation of 2 sets of independent aerodynamic surfaces and thrusters. The collaborative co-evolution process consists of 2 ‘species,’ the first consisting of the mechanical design variables and the second consisting of Proportional-Integral-Derivative (PID) and central pattern generator (CPG) controller variables. Each species have their own respective individual Evolutionary Algorithm (EA) solvers, Covariance Matrix Adaptation-Evolutionary Strategy (CMA-ES) and Parameter Exploring Policy Gradients (PEPG). In each optimization iteration, the parameters of one species is combined with representatives with the highest fitness from the other species and fed into a shared model for fitness evaluation, with each species taking turns to send a representative. Detailed performance comparison in trajectory tracking and power consumption between the proposed jointly optimized system against a design-only optimized, control-only optimized and unoptimized baseline were conducted. It was found that configurations with optimized designs would draw on average 18% less power than the non-optimized designs, and configurations with optimized controllers reduce error by 56% on average. The best performing configuration is the one with jointly optimized mechanical design and controller which outperforms all other configurations individually and collectively.
Published in: IEEE Robotics and Automation Letters ( Volume: 6, Issue: 2, April 2021)