Energy-Efficient Autonomous Aerial Navigation with Dynamic Vision Sensors: A Physics-Guided Neuromorphic Approach

TL;DR

This work tackles energy-efficient autonomous aerial navigation under dynamic obstacles by combining neuromorphic event-based perception with physics-guided planning and neurosymbolic reasoning. A shallow SNN processes Dynamic Vision Sensor data to detect moving gates, while a PgNN learns near-minimum-energy flight times using depth inputs and a physics-constrained energy model, optimized via a composite loss that includes data, physics, and energy terms. A symbolic planner fuses neural predictions to anticipate obstacle motion and compute collision-free trajectories, demonstrated in Gazebo/ROS with measurable gains in speed and energy efficiency. The results show robust, low-latency perception and energy-conscious navigation, illustrating the potential of neuromorphic perception plus physics-guided planning for real-time UAV autonomy.

Abstract

Vision-based object tracking is a critical component for achieving autonomous aerial navigation, particularly for obstacle avoidance. Neuromorphic Dynamic Vision Sensors (DVS) or event cameras, inspired by biological vision, offer a promising alternative to conventional frame-based cameras. These cameras can detect changes in intensity asynchronously, even in challenging lighting conditions, with a high dynamic range and resistance to motion blur. Spiking neural networks (SNNs) are increasingly used to process these event-based signals efficiently and asynchronously. Meanwhile, physics-based artificial intelligence (AI) provides a means to incorporate system-level knowledge into neural networks via physical modeling. This enhances robustness, energy efficiency, and provides symbolic explainability. In this work, we present a neuromorphic navigation framework for autonomous drone navigation. The focus is on detecting and navigating through moving gates while avoiding collisions. We use event cameras for detecting moving objects through a shallow SNN architecture in an unsupervised manner. This is combined with a lightweight energy-aware physics-guided neural network (PgNN) trained with depth inputs to predict optimal flight times, generating near-minimum energy paths. The system is implemented in the Gazebo simulator and integrates a sensor-fused vision-to-planning neuro-symbolic framework built with the Robot Operating System (ROS) middleware. This work highlights the future potential of integrating event-based vision with physics-guided planning for energy-efficient autonomous navigation, particularly for low-latency decision-making.

Figures (9)

• Figure 1: Event-Based Vision and Spiking Neuron Model: Illustration of the conceptual flow from event-based sensing to biological inspiration and spiking neural networks (SNNs). Left: An event camera outputs discrete intensity changes over time (red/blue dots) rather than continuous image frames. Center: Biological neurons communicate via discrete spikes, with dendrites receiving inputs that the soma integrates before generating an output spike. Right: In SNNs, inputs are represented as spikes over time, and the neuron body (soma) converts these event-driven signals into an output spike train, resembling biological neuronal firing.
• Figure 2: System architecture: Integrating neuromorphic vision, physics-guided neural networks, and symbolic rule-based reasoning. Event and depth streams feed the SNN and PgNN, informing real-time navigation decisions. Adapted from evplanner. The motion planner is based on the work in mellinger2011minimum.
• Figure 3: Event-based Object Detection at Various Depths. Each sub-panel shows neuromorphic event output and a bounding box around a moving gate. Although event density decreases with increasing depth, the SNN continues to isolate and track the gate in real time.
• Figure 4: Optimal velocity and energy-consumption patterns. Each depth features a characteristic velocity $v_{\mathrm{opt}}$ that balances time and power usage.
• Figure 5: Drone navigating through a moving gate in Gazebo simulation. The gate moves from right to left, while the drone starts at position $(1, 1)$. The timestamps ('t = 0s', 't = 1s', 't = 2s', 't = 3s') indicate the time elapsed at each key stage: (a) Start, (b) Approach, (c) Entry, and (d) Pass-Through.