Overview
This research presents a hierarchical neuromorphic control framework designed for fully onboard, closed-loop flight of Flapping-Wing Micro Aerial Vehicles (FWMAVs). The framework utilizes lightweight Spiking Neural Networks (SNNs) deployed on a widely available, resource-constrained ESP32 microcontroller, which has a unit cost of approximately $5. The system was applied to a butterfly-inspired robot weighing less than 30 grams. The objective was to address challenges associated with onboard control for FWMAVs, including nonlinear dynamics and stringent Size, Weight, and Power (SWaP) constraints.
Research Context
Flapping-Wing Micro Aerial Vehicles (FWMAVs) are recognized for their exceptional maneuverability and aerodynamic efficiency. However, the development of onboard control systems for these vehicles is complicated by their inherent nonlinear dynamics and the strict limitations on size, weight, and power consumption. Traditional control approaches often struggle within these stringent SWaP constraints, particularly when considering fully autonomous operation on low-cost, resource-limited hardware.
Approach
The hierarchical neuromorphic control framework incorporates two distinct Spiking Neural Networks (SNNs) for onboard operation. One SNN is responsible for state estimation, processing raw sensory feedback from the FWMAV. The second SNN manages control by modulating a Central Pattern Generator (CPG) that drives wing actuation. The entire system was trained using an imitation learning methodology. This setup enabled stable pitch and heading angle tracking during untethered real-world flight of the butterfly-inspired robot.
Findings
Experimental results demonstrated the viability of spike-based computation without requiring specialized hardware. The SNN-based controller achieved stable pitch and heading angle tracking during untethered real-world flight. A comparative analysis revealed that the SNN-based controller reduced inference latency by 36%, specifically from 1059 microseconds to 680 microseconds. Furthermore, the power consumption for inference was reduced by 18%, decreasing from 0.033 Watts to 0.027 Watts, when compared to a conventional Artificial Neural Network (ANN) baseline. This work represents the first demonstration of fully onboard neuromorphic control for autonomous flight of a FWMAV.
Why This Matters
This work highlights the potential of Spiking Neural Networks to enable energy-efficient autonomy under stringent Size, Weight, and Power (SWaP) constraints, particularly for platforms like Flapping-Wing Micro Aerial Vehicles. The successful deployment on a low-cost, resource-constrained microcontroller suggests a pathway for developing highly maneuverable and aerodynamically efficient robots with integrated control systems that are both power-efficient and cost-effective.