Soft robot hand learns how to avoid butter fingers

A soft robotic hand can pick up nearly a dozen objects and adjust the way it is holding them when they start to slip.

Many robotic arms are good at picking up objects, but in-the-moment adjustments when an object starts to slip can be challenging if many parts of the robot must be moved at once. Thomas Thuruthel at University College London and his colleagues have made a simple soft robotic hand that can prevent an object from slipping with only wrist motions.

They started out by making a human-like hand using a 3D-printed plastic skeleton and soft, moulded silicone material. To mimic the way human skin feels pressure, they added 32 flexible pressure sensors to the robot’s palm and fingers. Finally, they connected the hand to a movable arm with a motor in the wrist, which made the wrist the only part of the device that could move. All the sensors were wired to a computer to collect readings.

To train the hand, the researchers lowered it onto various objects, then tried different wrist motions, which made the springy fingers contort around the objects, to grip them and pick them up. Thuruthel says that even with immovable fingers, in more than a thousand attempts the hand could reliably grasp 11 out of 14 randomly selected household objects like a peach, a computer and a roll of bubble wrap.

To deal with slipping objects, they trained an AI software on sensor readings for both successful and failed pick-ups. The software learned to recognise when an object was beginning to slip out of grasp, and to then effect new wrist motions to prevent that from happening. When controlled by the AI, the hand successfully made this adjustment about 79 per cent of the time.

Robert Katzschmann at the Swiss Federal Institute of Technology in Zürich says that the new hand is too simple and limited in motion to mimic the different ways human hands grasp objects, but the experiment shows how it only takes a little information for an AI system to learn to discern the different ways an object can feel. He says that this AI approach will be something he thinks about when working on adding “senses” to the more complex and dexterous robotic hands his team is building.