neuromorphic-fw-mav-snn-control

name	neuromorphic-fw mav-snn-control
description	Neuromorphic control of flapping-wing micro aerial vehicles using SNNs on resource-constrained ESP32 microcontroller. Hierarchical SNN framework: state estimation + CPG modulation for wing actuation. 36% latency reduction, 18% power reduction vs ANN. First onboard neuromorphic autonomous flight.
tags	["neuromorphic-control","spiking-neural-network","micro-aerial-vehicle","embedded-ai","central-pattern-generator","edge-computing"]
category	ai_collection

Neuromorphic Control of Flapping-Wing MAV on Resource-Constrained Hardware

arXiv: 2605.19430 (May 19, 2026) Authors: Rim El Filali, Chenrui Feng, Chao Gao, Weibin Gu Category: Robotics (cs.RO)

Overview

First demonstration of fully onboard neuromorphic control for autonomous flight of a Flapping-Wing Micro Aerial Vehicle (FWMAV). Deploys two lightweight Spiking Neural Networks (SNNs) on a $5 ESP32 microcontroller for closed-loop flight control of a butterfly-inspired robot (<30g).

Core Innovation

Hierarchical Neuromorphic Control Framework

Raw Sensors -> SNN #1: State Estimation -> Estimated State
                                           |
                              SNN #2: CPG Modulation -> Wing Actuation

1. State Estimation SNN: Processes raw sensory feedback -> estimates pitch/heading angles
2. CPG Modulation SNN: Takes estimated state -> modulates Central Pattern Generator for wing actuation

Key Results

Metric	ANN Baseline	SNN Controller	Improvement
Inference Latency	1059 us	680 us	36% reduction
Inference Power	0.033 W	0.027 W	18% reduction
Training	Imitation learning	Imitation learning	Same
Flight Performance	Stable tracking	Stable tracking	Equivalent

Platform Details

• Robot: Butterfly-inspired FWMAV, <30g
• Controller: ESP32 microcontroller (~$5 unit cost)
• Control Tasks: Pitch and heading angle tracking
• Flight Type: Untethered real-world flight

Methodology

SNN Architecture

State Estimation Network

• Input: Raw sensor data (IMU, etc.)
• Output: Estimated pitch and heading angles
• Lightweight design for ESP32 constraints

CPG Modulation Network

• Input: Estimated state from state estimation SNN
• Output: Modulation signals for wing actuation CPG
• Maps desired trajectory to wing kinematics

Training: Imitation Learning

• Teacher: Conventional controller (ANN or model-based)
• Student: SNN learns to imitate teacher's control policy
• Deployment: SNN replaces teacher for inference

Central Pattern Generator (CPG)

• Bio-inspired oscillatory pattern generator for wing actuation
• SNN modulates CPG parameters (frequency, amplitude, phase)
• Enables stable flapping patterns with SNN-level control

Implementation Considerations

ESP32 Deployment

• Resource constraints: Limited RAM, CPU, power budget
• SNN advantages: Event-driven computation, sparse activation
• No specialized hardware: Runs on widely available ESP32

SWaP Constraints

• Size: Sub-30g total system weight
• Weight: Minimal payload capacity
• Power: Battery-limited operation
• SNN's 18% power reduction directly extends flight time

Latency-Critical Control

• Flapping-wing dynamics require sub-millisecond control loops
• SNN's 36% latency reduction enables tighter control
• Critical for stability of inherently unstable FWMAV

Applications

• Micro aerial vehicles: Insect-scale robots, surveillance
• Embedded AI: Ultra-low-power autonomous systems
• Bio-inspired robotics: Biomimetic flight control
• Edge computing: AI on constrained microcontrollers
• Swarm robotics: Multiple low-cost autonomous agents

Comparison with Existing Approaches

Approach	Hardware	Latency	Power	Onboard Control
Conventional ANN	ESP32	1059 us	0.033 W	Yes
SNN (this work)	ESP32	680 us	0.027 W	Yes
Neuromorphic chip	Loihi/SpiNNaker	Lower	Lower	Specialized HW required
Offboard control	GPU server	Lowest	Highest	No (tethered)

Limitations

• Single platform: Demonstrated on one FWMAV design
• 2D control: Pitch and heading only (no full 3D position control)
• Indoor flight: Real-world but constrained environment
• Imitation learning bound: SNN performance limited by teacher quality

Activation Keywords

• neuromorphic fw mav
• spiking neural network control
• flapping wing robot control
• ESP32 neuromorphic
• CPG modulation SNN
• imitation learning SNN
• micro aerial vehicle neuromorphic
• embedded spiking control
• resource-constrained SNN
• butterfly robot control
• neuromorphic autonomous flight

Citation

@article{elfilali2026neuromorphic,
  title={Neuromorphic Control of a Flapping-Wing Robot on Resource-Constrained Hardware},
  author={El Filali, Rim and Feng, Chenrui and Gao, Chao and Gu, Weibin},
  journal={arXiv preprint arXiv:2605.19430},
  year={2026}
}