In “Efficient Bayesian Exploration for Soft Morphology-Action Co-optimization”, written with Perla Maiolino and Fumiya Iida, we propose a framework for actively exploring robot control and morphology to aid discrimination tasks.
In soft robotics, a sensor’s morphology, meaning its shape and the materials it is made of, affects what it can feel. The right morphology can make a hard discrimination task easier, or even possible. But a robot that can change both its sensor morphology and the way it moves faces a large search: trying every combination of the two is slow, and unsuited to real-world use.
We developed a framework based on Bayesian exploration that lets a robot co-optimize both at once. The robot’s flat end-effector carries a capacitive tactile sensor, and we made three soft silicone filters for it, each with different morphological properties that change how the sensor responds to touch. Mounted on a robotic arm, the end-effector performed repeated, parameterized touches on eight objects that differ in geometry, surface texture, and stiffness. The tasks were to tell round objects from edged ones, rough surfaces from smooth ones, and stiff objects from soft ones.
The experiments show that morphing is necessary: with this sensor, some object properties can only be told apart by changing its morphology. The framework consistently found the best morphology-action configurations in about half the time of a systematic search over all parameters. It is a step toward robots that use both their bodies and their movements to help them perceive.
The work shares a question with our study of robotic palpation: how the way a robot touches shapes what it feels.