Learning-Based Downwash Modeling for Multirotors in Close Proximity Flight

Core Concepts

Learning-based approach for modeling downwash forces in close proximity flight improves sample efficiency and performance.

Abstract

This article introduces a novel learning-based approach for modeling downwash forces in close proximity flight for multirotors. The study focuses on the geometric properties of the downwash function and demonstrates the advantages of an equivariant model over non-equivariant models. Real-world flight experiments validate the efficacy of the proposed model in improving trajectory tracking and reducing errors. I. Introduction Multirotors flying in close proximity induce aerodynamic wake effects through downwash. Conventional methods lack 3D force-based models for robust control in dense formations. II. Problem Formulation Two multirotors operate in close proximity, with one acting as a leader and the other as a follower. Notation and dynamics of multirotor control are defined. III. Establishing Geometric Priors Geometric invariance and equivariance are defined in the context of group actions. Assumptions on the rotational equivariance of the downwash function are introduced. IV. Geometry-Aware Learning Feature mapping and equivariant model for downwash forces are presented. Proof of equivariance for the model is provided. V. Real-World Flight Experiments Sequential data collection method for training the model is explained. Study on model training efficiency and geometric learning benefits is conducted. Evaluation of model performance in dynamic flight scenarios is discussed. VI. Conclusion The proposed learning-based approach improves sample efficiency and performance in modeling downwash forces for multirotors.

Stats

Our geometry-aware model outperforms state-of-the-art baseline models trained with more than 15 minutes of data. Deploying our model online improves 3D trajectory tracking by nearly 36% on average. The equivariant model reduces vertical tracking errors by 56% and lateral tracking errors by 36%.

Quotes

"Our geometry-aware algorithm is sample-efficient." "Our equivariant model displays high sample efficiency and low bias."

Key Insights Distilled From

SO(2)-Equivariant Downwash Models for Close Proximity Flight

by H. Smith,A. ... at arxiv.org 03-27-2024

https://arxiv.org/pdf/2305.18983.pdf

Deeper Inquiries

How can the equivariant model be extended to model multi-vehicle downwash interactions?

To extend the equivariant model to model multi-vehicle downwash interactions, one approach could involve incorporating the interactions between multiple vehicles into the feature mapping and neural network architecture. By considering the relative positions, velocities, and orientations of multiple vehicles in the input data, the model can learn to predict the combined downwash effects on each vehicle. Additionally, the model can be trained on data collected from flights involving multiple vehicles to capture the complex interactions between them. By adapting the feature mapping to encode the symmetries and interactions between multiple vehicles, the equivariant model can be extended to effectively model multi-vehicle downwash interactions.

What are the implications of the rotational equivariance assumption on aggressive maneuvering scenarios?

The rotational equivariance assumption has significant implications on aggressive maneuvering scenarios where the downwash effects can be highly asymmetric and dynamic. In such scenarios, the assumption may not hold true as the downwash patterns can vary based on the rapid changes in the orientation and speed of the vehicles. This can lead to deviations from the assumed rotational symmetry, impacting the accuracy of the model predictions. In aggressive maneuvering scenarios, the model may need to adapt to the changing dynamics and incorporate additional features or mechanisms to capture the non-equivariant behaviors of the downwash forces. Failure to account for these variations can result in suboptimal performance and reduced effectiveness of the model in predicting the downwash effects accurately.

How can the geometric priors in the learning algorithm be adapted for outdoor flights with varying force magnitudes?

To adapt the geometric priors in the learning algorithm for outdoor flights with varying force magnitudes, the feature mapping and neural network architecture can be modified to account for the different force magnitudes encountered in outdoor environments. By incorporating additional features related to the force magnitudes and adjusting the model's parameters to handle a wider range of force variations, the algorithm can be tailored to outdoor flight conditions. Training the model on diverse datasets that include varying force magnitudes and environmental factors can help improve its robustness and generalization capabilities. Additionally, fine-tuning the model based on real-world outdoor flight data can enhance its performance in predicting downwash forces accurately under different conditions.