Abstract
Bimanual manipulation is challenging due to precise spatial and temporalcoordination required between two arms. While there exist several real-worldbimanual systems, there is a lack of simulated benchmarks with a large taskdiversity for systematically studying bimanual capabilities across a wide rangeof tabletop tasks. This paper addresses the gap by extending RLBench tobimanual manipulation. We open-source our code and benchmark comprising 13 newtasks with 23 unique task variations, each requiring a high degree ofcoordination and adaptability. To kickstart the benchmark, we extended severalstate-of-the art methods to bimanual manipulation and also present alanguage-conditioned behavioral cloning agent -- PerAct2, which enables thelearning and execution of bimanual 6-DoF manipulation tasks. Our novel networkarchitecture efficiently integrates language processing with action prediction,allowing robots to understand and perform complex bimanual tasks in response touser-specified goals. Project website with code is available at:http://bimanual.github.io