Performance-guided Task-specific Optimization for Multirotor Design

TL;DR

This work tackles the problem of designing multirotor MAVs that excel at a specific navigation task by co-optimizing airframe geometry and the control policy. It combines reinforcement learning, Bayesian optimization, and CMA-ES in an iterative sim-to-real framework, encoding designs, simulating them, training controllers, and evaluating task performance to guide design updates. In simulation, the task-optimized hexarotors outperform planar and fully actuated baselines on two waypoint-navigation tasks; a real-world deployment demonstrates promising sim2real transfer. The approach highlights the value of task-specific co-optimization for aerial robotics, with potential impact on custom MAVs for varied missions and future extensions to motor-propeller configurations and sensor placement.

Abstract

This paper introduces a methodology for task-specific design optimization of multirotor Micro Aerial Vehicles. By leveraging reinforcement learning, Bayesian optimization, and covariance matrix adaptation evolution strategy, we optimize aerial robot designs guided exclusively by their closed-loop performance in a considered task. Our approach systematically explores the design space of motor pose configurations while ensuring manufacturability constraints and minimal aerodynamic interference. Results demonstrate that optimized designs achieve superior performance compared to conventional multirotor configurations in agile waypoint navigation tasks, including against fully actuated designs from the literature. We build and test one of the optimized designs in the real world to validate the sim2real transferability of our approach.

Figures (12)

• Figure 1: Summary diagram of the full optimization pipeline.
• Figure 2: Position and orientation of each motor (with $xz$ plane symmetry). Translation is given in polar coordinates $(r,\theta,\phi)$ and orientation is defined with two Euler angles $(\alpha,\gamma)$. The table shows the bounds of the variables.
• Figure 3: Steps of the repair procedure to ensure a feasible airframe design. (a) Initial design: interfering airflows. (b) Retrieve the pose of the motors and add the margins representing the airflow. (c) Solve Optimization Problem: repair airframe design to find the minimum translation for each motor that separates them from each other and from the electronics cage. (d) Extract the new positions of the motors.
• Figure 4: Two samples of the waypoint probability distributions of task A and task B.
• Figure 5: Contribution of each reward by component for the planar hexarotor in task A and task B. The reward function considered in this paper is a sum of different reward terms. We define the contribution of each reward as the percentage of each reward component in absolute value with respect to the sum of all of the reward terms, averaged over $4\cdot10^4$ episodes.