[Paper Review] Learning Vision-based Cohesive Flight in Drone Swarms

Fabian Schilling, Julien Lecoeur+2 more|arXiv (Cornell University)|Sep 3, 2018

Robotics and Sensor-Based LocalizationEngineering26 references16 citations

TL;DR

This paper proposes a fully decentralized, vision-based control system for drone swarms that learns cohesive, collision-free flight directly from raw camera images using a convolutional neural network. By training on simulated data generated from a flocking algorithm, the controller predicts 3D velocity commands from visual inputs alone, achieving robust, sample-efficient coordination without communication or position sharing—demonstrating generalization to unseen swarm sizes and speeds.

ABSTRACT

This paper presents a data-driven approach to learning vision-based collective behavior from a simple flocking algorithm. We simulate a swarm of quadrotor drones and formulate the controller as a regression problem in which we generate 3D velocity commands directly from raw camera images. The dataset is created by simultaneously acquiring omnidirectional images and computing the corresponding control command from the flocking algorithm. We show that a convolutional neural network trained on the visual inputs of the drone can learn not only robust collision avoidance but also coherence of the flock in a sample-efficient manner. The neural controller effectively learns to localize other agents in the visual input, which we show by visualizing the regions with the most influence on the motion of an agent. This weakly supervised saliency map can be computed efficiently and may be used as a prior for subsequent detection and relative localization of other agents. We remove the dependence on sharing positions among flock members by taking only local visual information into account for control. Our work can therefore be seen as the first step towards a fully decentralized, vision-based flock without the need for communication or visual markers to aid detection of other agents.

Motivation & Objective

• To develop a decentralized, vision-based control system for drone swarms that eliminates reliance on position sharing or communication.
• To enable robust, cohesive flight in dense swarms using only local visual inputs and end-to-end learning from raw images.
• To demonstrate that a neural network can implicitly learn agent localization and relative motion from visual data without explicit supervision.
• To validate the controller's generalization to unseen conditions such as varying swarm size and speed.
• To provide a foundation for real-world deployment of fully autonomous, vision-only drone swarms.

Proposed method

• The controller is formulated as a regression problem: predicting 3D velocity commands from raw omnidirectional camera images.
• A simulated swarm of quadrotors generates training data by running a simple flocking algorithm with random initializations and trajectories.
• For each image, the corresponding target velocity command is computed from the flocking algorithm's dynamics, creating a paired visual-control dataset.
• A convolutional neural network (CNN) is trained to map visual inputs to predicted velocity commands by minimizing the error between predictions and targets.
• Saliency maps are computed via gradient-based attribution on the penultimate convolutional layer to visualize which image regions most influence motion decisions.
• The trained CNN is deployed as a real-time, reactive controller in simulation and initial real-world experiments using PX4 and ROS.

Experimental results

Research questions

• RQ1 Can a vision-based neural controller learn cohesive, collision-free flight in drone swarms without relying on position sharing or communication?
• RQ2 To what extent can a CNN trained on simulated data generalize to unseen swarm sizes and speeds?
• RQ3 Can the network implicitly learn to localize other agents from raw visual inputs without explicit detection supervision?
• RQ4 How effective is the saliency map in identifying visually salient regions influencing motion decisions?
• RQ5 Can the controller be transferred from simulation to real-world flight using the same software stack?

Key findings

• The vision-based controller successfully achieves collision-free, coherent flight in simulated swarms of varying sizes (N=3,12) and speeds (v^max=2.0, 4.0 m/s), demonstrating strong generalization.
• The controller maintains stable inter-agent distances and high order parameters even when the migration point is removed, indicating self-organized cohesion.
• Saliency map visualization reveals that the network effectively localizes other agents in the visual field, with high activation on specific agents visible in the camera view.
• The saliency map is computed efficiently via backpropagation and could serve as a real-time prior for detection or relative localization systems.
• The method generalizes well to unseen conditions, with no noticeable degradation in performance under perturbations such as changes in swarm size or speed.
• Initial real-world experiments confirm that the same software stack (PX4/ROS) used in simulation can be applied to real drones, supporting a smooth transfer from simulation to reality.