The frequent introduction of new objects to manipulate in factory lines is a big challenge for industrial robotics, where the classical approach requires the intervention of a human expert to program the robot for each new reference. Reinforcement Learning methods allow the robot to adapt its grasping parameters autonomously, however it is known as time consuming, reducing its usefulness in real-world settings. We propose to reduce the training time by providing previously optimized sets of parameters for objects similar to the new ones as starting points of the RL algorithm, using a multi-dimensional based similarity algorithm (e.g. shape, weights, material).
Keywords
Reinforcement Learning, Transfer Learning, Computer Vision, Industrial Robotics, Adaptive Robotics