By: Brian Heater
January 30, 2019
Turns out training a robotic arm to play Jenga is a surprisingly complex task. There are, so to speak, a lot of moving parts. Researchers at MIT are putting a modified ABB IRB 120 to work with the familiar tabletop game, utilizing a soft gripper, force-sensing wrist joint and external camera to design a bot that can remove a block without toppling the tower.
The robot was trained with 300 attempts, rather than the thousands it would traditionally take, learning to cluster different attempts into groups as a kind of shorthand similar to how humans teach themselves. With each attempt the robot is pushing against the block, testing for tactile feedback to determine whether it’s a safe bet.