Evolving Vision-Based Flying Robots

Jean-Christophe Zufferey, Dario Floreano, Matthijs van Leeuwen, Tancredi MerendaBiologically Motivated Computer Vision 2002 · DOI

TL;DR

This paper presents a novel approach to developing autonomous flying robots that can navigate complex indoor environments using only visual input. By genetically encoding and evolving spiking neural circuits directly on the physical robot, the researchers demonstrate that these robots can learn efficient, albeit unexpected, navigation strategies without human intervention.

What Problem Does This Paper Address?

Developing autonomous navigation for miniature flying robots is a significant challenge. Traditional methods relying on heavy sensors like lasers or ultrasonics are impractical due to weight and power constraints. Vision-based navigation is a promising alternative, but conventional computer vision techniques often require substantial processing power and energy, which are scarce on small robots. This paper tackles the problem of enabling miniature flying robots to navigate complex environments using lightweight, low-power vision systems, drawing inspiration from biological systems.

Methods

The core of this research lies in evolutionary robotics and neuromorphic control. The authors genetically encode the architecture of a spiking neural circuit that connects the robot's low-resolution linear camera to its motors. This genetic code is then evolved using a standard genetic algorithm on the physical robot.

Here's a breakdown of the key methods:

  • Spiking Neural Circuits: Unlike traditional artificial neural networks, these circuits mimic biological neurons by firing discrete "spikes" when a certain threshold is reached. This approach is inherently more energy-efficient and can potentially capture temporal dynamics crucial for real-time control.
  • Genetic Encoding: The structure of the spiking neural network, including connectivity patterns and neuron signs (excitatory/inhibitory), is represented by a binary string. This string is the "genome" that is evolved.
  • Evolutionary Algorithm: A population of these genetic strings (representing different neural network configurations) is tested on the physical robot. The "fitness" of each individual is determined by its ability to maximize forward motion, measured by an anemometer. The fittest individuals are selected, reproduced (with crossover and mutation), and form the next generation. This process is repeated over many generations.
  • Physical Robot Evolution: Crucially, the evolution happens directly on the physical robot. This allows the evolved controllers to directly exploit the robot's morphology and the dynamics of its interaction with the environment, leading to potentially novel and efficient solutions.
  • Vision System: A simple linear camera with a wide field of view (150°) captures images. This raw visual data is processed by a Laplace filter to detect contrast, which then serves as input to the spiking neural network.

Datasets

The paper does not explicitly mention or use pre-existing datasets for training. Instead, the "dataset" is generated through the real-time interaction of the physical robot with its environment during the evolutionary process. The environment itself is a room with randomly sized black and white stripes on the walls, serving as a textured visual input.

Key Insights and Ideas

  • Emergent Collision Detection: Perhaps the most surprising finding is that the evolved spiking circuits, despite receiving only visual input, learn to effectively detect collisions. This is inferred from the robot's behavior of turning off motors and bumping backward when it hits a wall, a behavior strongly correlated with bumper activation. This suggests that the visual system, in conjunction with the robot's dynamics, can implicitly signal impending collisions.
  • Exploiting Environmental Dynamics: The evolved robots don't just avoid walls; they actively use them to stabilize their trajectory. This highlights how artificial evolution can discover strategies that leverage the physical properties of the robot and its environment in unexpected ways.
  • Simplicity and Efficiency: The evolved controllers are remarkably simple, using a low-resolution camera and a small spiking neural network. This demonstrates the power of evolutionary approaches to find efficient solutions for complex tasks with limited resources.
  • Bio-Inspired but Novel: While inspired by biological vision systems, the evolved solutions are not necessarily direct replicas of insect behavior. The authors suggest that the evolved "tricks" might be novel, even if they are effective.

Key Results

  • Efficient Navigation: Within 20 generations (2-3 days of evolution), the best evolved individuals developed efficient strategies to navigate around the room in a forward direction.
  • Wall Stabilization: The robots learned to use walls to stabilize their course, a behavior not explicitly programmed but emerged from the evolutionary process.
  • Qualitative Performance: Figure 5 shows a typical trajectory where the robot moves forward, hits a wall, stabilizes, moves forward again, and repeats this cycle. The performance graphs show a clear correlation between collision events and changes in motor activity.

Limitations and Future Work

  • Limited Visual Information: The current setup uses a very low-resolution linear camera, which limits the complexity of the visual information processed. The authors acknowledge that distinguishing between frontal and side collisions is not possible with the current bumper setup.
  • Altitude Control: While the paper focuses on 2D navigation, the blimp is a 3D vehicle. The initial experiments were limited to 2D by a pre-designed altitude regulator.
  • Simulation vs. Real World: The authors are developing a 3D flight simulator to speed up evolutionary runs and allow for the evolution of sensor morphology. However, they recognize the challenge of bridging the gap between simulation and real-world performance, suggesting that evolving hebbian-like synaptic plasticity might be a way to address this.
  • Future Directions: Future work includes investigating the behavioral effects of different imaging devices and preprocessing filters, and eventually applying this approach to indoor slow flyers with wings.

Summary updated

BibTeX
@inproceedings{zufferey2002evolving,
  author = {Jean-Christophe Zufferey and Dario Floreano and Matthijs van Leeuwen and Tancredi Merenda},
  title = {{Evolving Vision-Based Flying Robots}},
  booktitle = {Biologically Motivated Computer Vision},
  year = {2002},
  doi = {10.1007/3-540-36181-2\_59},
}
Figure