Evolving Morphological Complexity in Collective Robotics

• Behavioural Competence

  Evolve simpler team morphologies (sensory configurations) without sacrificing behavioural competence.

• Environmental Complexity

  Investigate the relationship between environmental difficulty and morphological complexity

Evolutionary methods based on NEAT and HyperNEAT used to evolve controller and morphology.

• NEAT

  Evolves both weights and topology; Phenotype (Artificial Neural Network) directly encoded by genotype (String).

• HyperNEAT

  Evolves both weights and topology; Phenotype (Artificial Neural Network) indirectly encoded by genotype (CPPN / Compositional Pattern Producing Network)

Evolving Controller and Morphology

• Evolving Controller

  Evolving an Artificial Neural Network (ANN) as the controller.

• Evolving Morphology

  Encoding a sensory configuration into the input layer of the ANN.

Experimental Simulator

Collective Gathering

Robot teams are evolved to locate and co-operatively push the yellow blocks (resources) to the rectangular gathering zone at the bottom of the environment.

Environmental Difficulty

We define co-operation as the number of robots required to push a block, and we adjust environmental difficulty by increasing the number of blocks and degree of co-operation required.

Morphological Complexity

Morphological complexity is a function of number of sensors as well as the strength (field of view and range) of each sensor.

Experiments

Experiments compared evolution of robot teams with versus without a cost of morphological complexity.

The relationship between morphological complexity and behavioural competency

In each environment, adding a cost of morphological complexity resulted in simpler team morphology without sacrificing behavioural competence. This suggests that, for real-world co-operative robot tasks, competent teams can be evolved while automatically reducing design costs which would have been spent on unnecessary sensors.

This result is illustrated visually in the videos and ANN controller depictions below, showing the best solutions evolved with versus without a cost of complexity. We see that a cost of complexity results in fewer and cheaper (smaller range and field-of-view) sensors, but without sacricficing the team's ability to locate and gather resources in the environment. Note that we only show videos and networks for the simple environment as this finding was consistent across all environments.

Evolution with NEAT

	Without Cost of Complexity	With Cost of Complexity
Controller-Morphology Coupling in Action
Topology of evolved ANN controller (input layer encodes morphology)

Evolution with HyperNEAT

	Without Cost of Complexity	With Cost of Complexity
Controller-Morphology Coupling in Action
Topology of CPPN (indirectly encodes the ANN controller)
Topology of evolved ANN controller (input layer encodes morphology)

The Relationship between Morphological Complexity and Environmental Difficulty

Robot teams did not require higher morphological complexity for competent behaviour inincreasingly difficult environments. For multi-robot system design, this implies that additional costs on sensor parts are not always necessary for optimal team behaviour in difficult environments. This could facilitate lower spending on sensors when environmental difficulty is variable.

The figures below show progression of morphological simplicity (minimal morphological complexity) produced when there is a cost of complexity (MO) and when there is no cost (SO) in the three environments over evolutionary time (generations).

NEAT	HyperNEAT

We see that with a cost of complexity (MO), evolution selected for approximately the same degree of morphological complexity over evolutionary time regardless of the environment. Without a cost of complexity (SO), evolution selected for higher morphological complexity in response to increasing environmental difficulty.