Abstraction, Sensory-Motor Coordination, and the Reality Gap in Evolutionary Robotics

Area of Science:

• Robotics
• Artificial Intelligence
• Evolutionary Computation

Background:

• Transferring controllers from simulation to real robots is a major challenge in evolutionary robotics.
• The 'reality gap' often causes controllers evolved in simulation to fail in real-world applications.
• Abstraction, involving preprocessed sensory inputs and higher-level motor commands, is explored as a solution.

Purpose of the Study:

• To investigate the impact of sensory and motor abstraction on the transferability of evolved robot controllers.
• To demonstrate how abstraction can mitigate the reality gap in evolutionary robotics.
• To assess if abstraction preserves the effectiveness of genetic optimization for complex behaviors.

Main Methods:

• Evolving two controllers with varying levels of abstraction for a swarm micro air vehicle formation task.
• Testing controller performance in both simulated and real-world environments.
• Analyzing the resulting behaviors for sensory-motor coordination and exploitation of environmental properties.

Main Results:

• Controllers with lower abstraction failed in the real world due to the reality gap.
• Controllers with higher abstraction performed effectively in both simulation and reality.
• Abstraction reduced computational complexity, accelerating the optimization process.
• Optimized behaviors demonstrated exploitation of environmental dynamics and input shaping for formation control.

Conclusions:

• Higher levels of abstraction are crucial for robust controller transfer in evolutionary robotics.
• Abstraction is a viable strategy to overcome the simulation-to-reality gap.
• Abstraction does not necessarily diminish the power of genetic optimization for discovering complex behaviors.