Abstract: A quadruped robot with four actuated hip joints and four passive highly compliant knee joints is used to demonstrate the potential of underactuation from two standpoints: learning locomotion and perception. First, we show that: (i) forward loco-motion on flat ground can be learned rapidly (minutes of optimization time); (ii) a simulation study reveals that a passive knee configuration leads to faster, more stable, and more efficient locomotion than a variant of the robot with active knees; (iii) the robot is capable of learning turning gaits as well. The merits of underactuation (reduced controller complexity, weight, and energy consumption) are thus preserved without compromising the versatility of behavior. Direct optimization on the reduced space of active joints leads to effective learning of model-free controllers. Second, we find passive compliant joints with po-tentiometers to effectively complement inertial sensors in a velocity estimation task and to outperform inertial and pressure sensors in a terrain detection task. Encoders on passive compliant joints thus constitute a cheap and compact but powerful sensing device that gauges joint position and force/torque, and — if mounted more distally than the last actuated joints in a legged robot — it delivers valuable information about the interaction of the robot with the ground.

Key words: underactuated robots, legged robots, compliant joints, force and tactile sensing, haptic terrain classification, learning locomotion