Learning coupling for pendulum swing-up using physical reservoir control

The coupling of dynamical systems within living organisms frequently gives rise to emergent synergies, as seen in animal movement and homeostatic control. These natural mechanisms have inspired robotic control strategies, but existing approaches require precise models and significant manual or computational effort to engineer coupling. We present physical reservoir control, in which a physical reservoir learns and continuously adapts its coupling with a plant to carry out nonlinear control tasks without external intervention. Our results demonstrate that a physical reservoir composed of a pair of double pendulums can swing up and balance a torque-limited single pendulum. We show that the coupled plant-reservoir system exhibits a clear attraction in the presence of stochastic noise and, importantly, can re-establish its coupling following disturbances like external impulses or signal loss. Further experiments show that breaking plant-reservoir coupling results in failure of swing-up, while successful control can also be achieved using an alternative oscillator reservoir. Beyond enabling model-free control, our findings reveal how natural dynamical systems leverage physical coupling to generate nonlinear behavior, enhance robustness, and preserve stability, leading to the next generation of robot design.