Encoding mechanical intelligence using ultraprogrammable joints

摘要

Animal bodies act as physical controllers, with their finely tuned passive mechanical responses physically "encoding" complex movements and environmental interactions. This capability allows animals to perform challenging tasks with minimal muscular or neural activities, a phenomenon known as embodied intelligence. However, realizing such robots remains challenging due to the lack of mechanically intelligent bodies with abundant tunable parameters-such as tunable stiffness-which is a critical factor akin to the programmable parameters of a neural network. We introduce an elastic rolling cam (ERC) with accurately inverse-designable rotational stiffness. The ERC can closely replicate 100,000 randomly generated stiffness profiles in simulation. Prototypes ranging from millimeters to centimeters were manufactured. To illustrate the mechanical intelligence encoded by programming the ERC's stiffness response, we designed a bipedal robot with optimized ERC passive knees, achieving energy-efficient, open-loop stable walking across uneven terrain. We also demonstrated a quadcopter drone with ERC joints encoding an impact-activated, dual-state morphing.