A single-demonstration guided manipulation learning with dexterous hand

Abstract

Demonstration assisted reinforcement learning has been proven to be an extremely effective method for solving complex multi-fingered dexterous hand manipulation tasks. It usually requires costly and time-consuming expert demonstrations for each task, affecting the learning efficiency of the dexterous manipulation policy. To overcome this drawback, this paper devotes to use only one human demonstration per task to obtain a generalizable dexterous manipulation policy. And a novel ‘Basics-before-Extension’ policy learning strategy (BBE) is proposed for this purpose. It consists of two learning stages. In the ‘basics learning stage’, the dexterous hand extracts hand-object contact points as key clues from demonstration, facilitating a quick learning of the expert basic skill. While in ‘extension learning stage’, the designed joint policy training scheme enables the expert knowledge to be transferred and adapted to new environments, outputting generalizable policy. We present the distinctive overall framework from the low-cost demonstration data collection to the policy learning process. Meanwhile, BBE strategy has been experimentally validated on typical grasp and manipulation tasks, including relocating objects, opening door, hammering nails and functional grasp. The results indicate that the proposed BBE strategy can empower the multi-fingered dexterous hand with the intelligence of learning typical grasp and manipulation efficiently and accurately from a single demonstration.