Scalable co-optimization of morphology and control in embodied machines

Area of Science:

• Artificial Intelligence
• Robotics
• Evolutionary Computation
• Embodied Cognition

Background:

• Traditional AI and robotics design robots with fixed body plans and separately optimized controllers.
• Simultaneously evolving robot morphology and control policies is a significant challenge.
• The theory of embodied cognition highlights the interdependence of body plan and sensorimotor control.

Purpose of the Study:

• To investigate the challenges of co-optimizing morphology and control in embodied AI.
• To propose and evaluate a novel technique for 'morphological innovation protection' to overcome evolutionary stagnation.
• To explore the implications for automated robot design and embodied cognition research.

Main Methods:

• Implementing a 'morphological innovation protection' technique that temporarily reduces selection pressure on individuals with recent morphological changes.
• Allowing time for control policies to adapt to new morphologies through subsequent mutations.
• Testing the method across diverse initial conditions to assess convergence and fitness improvements.

Main Results:

• The proposed method successfully avoids local optima in the co-optimization process.
• It enables convergence to similar, highly fit morphologies regardless of initial conditions.
• The technique sustains fitness improvements throughout the optimization process, outperforming standard methods.

Conclusions:

• Morphological innovation protection is a promising step towards scalable optimization of embodied AI.
• This approach offers theoretical insights into evolutionary stagnation in robot design.
• The method provides a test bed for investigating the theory of embodied cognition and automating robot development.