How robots learn to adapt by mixing learned patterns?

Robots struggle to handle new situations because the world's structure shifts constantly. This work proposes that generalization comes from adaptively combining learned regularities—reliable patterns in how the robot and environment interact—into situation-specific behaviors. Using AICON, a differentiable framework that represents regularities as interacting processes, the authors show the system generates appropriate behavior across novel conditions by automatically weighting which patterns matter most for each situation.