Rolling robots offer efficient locomotion and terrain adaptability, making them ideal candidates for environmental exploration and search-and-rescue missions. However, enabling these robots to autonomously navigate complex environments typically requires sophisticated sensory networks and centralized control architectures, which increase structural complexity and energy consumption. Here, we present a rolling robot based on mechanical intelligence that achieves autonomous obstacle avoidance through structural responsiveness alone, eliminating the need for electronic sensors or processors. Our design leverages a spatial parallel mechanism coupled with a continuously driven pendulum. Upon encountering an obstacle, the pendulum spontaneously shifts from a forward to a backward bias, reversing the driving torque and inducing a passive diameter contraction on one side of the robot. This geometric asymmetry triggers a deterministic “backward-turn-forward” maneuver, allowing the robot to steer away from the obstruction. We derive the kinematic constraints to optimize the radius contraction ratio and establish a dynamic model to predict the steering trajectory. Experimental validation demonstrates that the lightweight prototype can autonomously navigate a maze with 90° turns using only a single actuator. This work illustrates how mechanical intelligence derived from intrinsic structural dynamics can replace complex control loops, providing a robust, energy-efficient approach for designing autonomous machines in unstructured environments.