Automatic modular design of robot swarms using behavior trees as a control architecture.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Science

Background:

• Swarm robotics relies on efficient collective behavior design.
• Automatic design methods are crucial for scalability.
• Behavior trees offer a modular approach to robot control.

Purpose of the Study:

• To investigate the use, challenges, and limitations of behavior trees in automatic modular design for swarm robotics.
• To introduce and evaluate the Maple method for designing robot behaviors using behavior trees.
• To compare behavior trees with probabilistic finite state machines for swarm control.

Main Methods:

• Developed Maple, an automatic design method combining modules into behavior trees.
• Conducted three empirical studies on aggregation and foraging missions.
• Compared Maple (behavior trees) with Chocolate (probabilistic finite state machines).
• Assessed reality gap, design budget impact, and behavior tree structural variants.

Main Results:

• Maple successfully generated control software that addressed the reality gap.
• Performance was analyzed concerning the design budget (simulation runs).
• Variants of Maple explored different behavior tree structural constraints.
• Behavior trees, while appealing, did not consistently outperform finite state machines in swarm robotics.

Conclusions:

• Behavior trees present possibilities but also challenges and limitations in swarm robotics.
• The Maple method demonstrates an approach to utilizing behavior trees for collective behaviors.
• Finite state machines remain a competitive alternative for swarm control architectures.