Let the Robots Feel Like the Bees Do

The aim of this FWF project is to create a BEECLUST-driven robot swarm that operates in dynamic temperature gradient fields (like honeybees do). While the BEECLUST algorithm, which is derived from honeybees' collective thermotaxis, was implemented on my robot swarms, it was always searching in an environment that was characterized by other stimuli: light, terrain depth or magnetic field. Thus, a re-embodiment of the original source of inspiration, honeybees' thermotactic behavior, was long missing.

For this challenge, honeybees were first observed in temperature gradients in groups and alone. As they show a high diversity of behavior an individual-based model was derived from these explorations and re-embodied in several groups of agents: first in simulated agents in a computer model and then in real robots. These robots needed to mimic the sensor capabilities of measuring the local temperature that honeybees possess, thus we equipped several variants of robots with antennae with thermosensors and also thermosensors on the ground. Then the behavioral programs of honeybees were implemented on the robots and then tuned until all typical honeybee behaviors from the empirical observation were re-embodied: goal finders, wall followers, random walkers and sitters. Finally, evolutionary computation algorithms were employed in order to find optimal swarm configurations for specific search landscapes, which varied in the complexity and dynamics of the temperature gradient fields.

Prototype 1: A single HEMISSON robot, equipped with antennae and additional temperature sensors

Prototype 2: A group of e-puck robots, equipped with antennae and temperature sensors and a specifically designed PCB ontop

Typical REBODIMENT Experiment

Here a typical REBODIMENT experiment is shown with the Prototype 3 robots, which have shorter antennae and a specific sensory board mounted on the top of the e-puck robots. These robots are all from the "random walker" behavioral type, later in the course of the REBODIMENT also the other behavior types observed with bees ware analyzed in mixed swarm settings.

In this experimental condition there are two target spots: A global temperature optimum of 36 °C at the right side of the arena and a local temperature optimum of 32 °C at the left side of the arena. This setting and the whole arena setup, reflects the series of experimentations we had performed with honeybees in comparable temperature gradient fields. Only the size of the arena was scaled up proportionally from the arena-size and the agent size in the honeybee experiments.

Phase 1: Initially all robots are stared in the center the two target spots (left, right) are empty

Phase 2: Clusters start to form at both sides around the warm target spots, hence the name BEECLUST for this algorithm

Phase 3: After some time the cluster in the local optimum (left) and in the global optimum (right) are formed and all/most of the free robots are bound to either one of the clusters

Phase 4: In the later course of the dynamics the cluster at the arena side with the better environmental quality (global optimum side) can bind the agents stronger, thus it finally "sucks up" all free agents to the cost of the other side (local optimum) until it is saturated.