Neuro-Embodied Adaptive Robotics (NEAR) Lab

Neuro-Embodied Adaptive Robotics (NEAR) Lab

Near the future of embodied intelligence.

Spiking Neural Network Control of a Flapping-Wing Robot on Resource-Constrained Hardware

Published in 10th Annual Conference on Robot Learning, 2026

Preprint Spotlight

This work presents a hierarchical neuromorphic control framework for a flapping-wing micro aerial vehicle under severe onboard computing constraints. The system deploys lightweight spiking neural networks on an ESP32-class microcontroller for state estimation and closed-loop control.

arXiv PDF Code

Neuromorphic control framework, robot hardware, and flight demonstration
Original paper teaser figure showing the data-to-deployment workflow, the butterfly-inspired FWMAV hardware, and untethered flight demonstrations.

Why This Matters

Small flapping-wing robots face strict limits in payload, power, and onboard computation. Conventional neural controllers can be difficult to deploy directly on such platforms. Neuromorphic approaches offer a biologically inspired alternative that may reduce latency and power while preserving responsive behavior.

Main Contributions

  • Proposes a hierarchical neuromorphic control framework for a flapping-wing micro aerial vehicle.
  • Deploys two lightweight spiking neural networks onboard: one for state estimation and one for control through central-pattern-generator modulation.
  • Trains the controller by imitation learning and evaluates closed-loop pitch and heading tracking in untethered flight.
  • Demonstrates spike-based computation on a low-cost, widely available ESP32 microcontroller rather than specialized neuromorphic hardware.

Key Findings

  • The SNN-based controller reduces reported inference latency from 1059 us to 680 us relative to an ANN baseline.
  • The SNN-based controller reduces reported inference power from 0.033 W to 0.027 W relative to the same baseline.
  • The experiments demonstrate fully onboard neuromorphic control for autonomous flapping-wing flight under stringent SWaP constraints.
Latency and power comparison for SNN, ANN, and event-driven SNN inference
Original paper figure comparing inference latency and power on resource-constrained hardware.

Citation

El Filali, Rim, Chenrui Feng, Chao Gao, and Weibin Gu. “Neuromorphic Control of a Flapping-Wing Robot on Resource-Constrained Hardware.” arXiv preprint arXiv:2605.19430 (2026).