Experiments in Evolutionary Synthesis of Robotic Neurocontrollers

Karthik Balakrishnan, Vasant G. HonavarAAAI 1996

Introduction

Designing effective neural controllers for autonomous robots is a significant challenge in robotics. Engineers must balance speed, obstacle avoidance, and computational complexity, often navigating a vast search space of potential network architectures. Traditional manual tuning is frequently inefficient and prone to getting stuck in local optima. This research explores an alternative: using Evolutionary Algorithms (EAs) to automatically synthesize neurocontrollers tailored for specific environmental tasks.

The Problem: Navigation Without Feedback

The study focuses on a box-pushing task where a robot must clear an arena by pushing objects to the surrounding walls. A primary difficulty in this simulation is the lack of physical feedback; the robot does not naturally "bounce" off obstacles. Consequently, a simple feedforward controller—which lacks internal state or memory—often leads the robot to drive fruitlessly into a wall, resulting in low fitness scores. The goal was to determine if evolution could discover architectures capable of overcoming this limitation.

Methodology

The researchers utilized Genetic Algorithms (GAs) to evolve high-performing neurocontrollers. Their experimental framework included:

  • Architectural Comparison: Comparing standard feedforward networks against recurrent neural networks (RNNs), which allow for internal feedback loops.
  • Evolutionary Growth: Instead of starting with complex, fully connected RNNs, the researchers allowed mutation to incrementally add recurrent links to simpler feedforward structures.
  • Co-evolution of Sensors: The study also experimented with evolving the robot's sensory configurations (its "eyes") alongside its neural architecture, allowing the robot's physical perception to adapt to its environmental constraints.
  • Fitness Metric: Performance was measured by the number of boxes successfully pushed to the arena walls within a fixed time limit.

The "Toggle" Mechanism: A Simple, Elegant Solution

The most significant finding of the study was the emergence of a specific neural structure within the evolved recurrent networks. Analysis revealed that evolution consistently "invented" a large, negative self-loop on the output unit controlling movement.

This self-loop functions as a biological "toggle switch." By creating a negative feedback loop, the network forces the robot to frequently alternate between moving forward and turning. This rhythmic "jitter" prevents the robot from becoming trapped in a loop of driving into a wall, providing a robust solution to the lack of environmental feedback. This emergent behavior demonstrates how evolution can find non-obvious, minimalist solutions to complex navigation problems.

Key Results and Efficiency

  • Performance Superiority: Recurrent networks achieved significantly higher fitness scores than feedforward networks, primarily due to the discovery of the toggle mechanism.
  • Minimalist Architecture: Evolution produced highly efficient controllers. Networks that evolved recurrent links through mutation used 30% to 70% fewer links than networks that were pre-designed to be fully recurrent. This suggests that evolutionary synthesis can find the "minimal necessary complexity" for a task.
  • Robustness: Preliminary findings suggest that these evolved controllers are inherently fault-tolerant, maintaining performance even in the presence of noisy sensors or component failures.

Limitations and Future Work

While successful, the researchers noted that the evolved behaviors are highly specific to the current simulation's constraints, such as the lack of wall-bounce feedback. Future research will focus on:

  1. Environmental Complexity: Testing how evolved controllers adapt to changing arena shapes or the introduction of sensory noise.
  2. Scaling: Determining if these evolved "toggle" behaviors and minimalist architectures can scale to more complex, multi-task robotic applications.
  3. Constraint Interplay: Further investigating how different environmental pressures shape the resulting neural architecture.

Summary updated

BibTeX
@inproceedings{balakrishnan1996experiments,
  author = {Karthik Balakrishnan and Vasant G. Honavar},
  title = {{Experiments in Evolutionary Synthesis of Robotic Neurocontrollers}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1996},
}
Figure