Active propulsion noise shaping for multi-rotor aircraft localization

TL;DR

This work tackles the problem of MAV localization when vision-based cues are unreliable by exploiting the drone's own rotor self-noise as an acoustic beacon. It introduces a fully differentiable forward model of a rotorcraft in an acoustic environment and a neural inverse model that localizes using microphone data and rotor phase information, jointly optimizing rotor phase modulation under physical constraints. The key contributions include a neural self-noise-based localization approach in a known acoustic environment, a physically feasible rotor phase modulation framework learned alongside localization, and a differentiable forward model calibrated with real rotor pressure data, demonstrating substantial accuracy gains through phase learning and measurement aggregation. The results suggest a practical, energy-efficient, and robust localization modality for MAVs in challenging conditions, with potential for extension to 3D localization and SLAM in future work.

Abstract

Multi-rotor aerial autonomous vehicles (MAVs) primarily rely on vision for navigation purposes. However, visual localization and odometry techniques suffer from poor performance in low or direct sunlight, a limited field of view, and vulnerability to occlusions. Acoustic sensing can serve as a complementary or even alternative modality for vision in many situations, and it also has the added benefits of lower system cost and energy footprint, which is especially important for micro aircraft. This paper proposes actively controlling and shaping the aircraft propulsion noise generated by the rotors to benefit localization tasks, rather than considering it a harmful nuisance. We present a neural network architecture for selfnoise-based localization in a known environment. We show that training it simultaneously with learning time-varying rotor phase modulation achieves accurate and robust localization. The proposed methods are evaluated using a computationally affordable simulation of MAV rotor noise in 2D acoustic environments that is fitted to real recordings of rotor pressure fields.

Figures (6)

• Figure 1: Forward and inverse models. (A) Stages of the forward and the inverse models and their parameters. Learnable parameters are denoted in red. (B) The geometry of sources, microphones, and the environment. (C) The geometry of zeroth, first, and second-order image sources in a rectangular room.
• Figure 2: Simulated pressure fields generated by the aircraft in free space (A) and in a square room at different times (B-D). Positive and negative pressures are color-coded in red and blue, respectively. A circle of $0.51$m around each rotor is not modeled in the absence of data recording in blade proximity.
• Figure 3: Rotor phase modulations evaluated in the experiments. Rotors are color-coded. Counter-rotating rotor pairs are (1,4) and (2,3).
• Figure 4: Localization uncertainty in a square $5$m $\times$$5$m room with learned phase modulation. Shown are $1 \sigma$ uncertainty ellipses calculated in a $0.05$ radius over a uniform grid of $64$ azimuthal orientations. RMS errors are color-coded. Left-to-right: no angular aggregation; geometric median aggregation post-training; and training through the aggregation. Average RMS localization accuracy is reported in the captions in relative units.
• Figure 5: Robustness to various sources of modeling and sensing noise. (A-D) environment parameters mismatch between training and inference. Nominal parameters are indicated by vertical lines. (E-F) sensitivity to sensing and rotor phase noise. Shown is the performance of a model trained in noiseless settings compared to a model trained with noise injection. Shaded regions indicate $1 \sigma$ confidence intervals calculated over a uniform grid of locations in the room. A $5$m $\times$$5$m room was used at training. Phase modulation was trained in all models.