Using Reverse Reinforcement Learning for Assembly Tasks
Keywords: Reinforcement Learning, Assembly, Planning, Lego
TL;DR: We propose to use Reverse Reinforcement Learning to speed up training of assembly tasks and demonstrate it in a simulation environment.
Abstract: Using Reinforcement Learning (RL) for assembly tasks is of high interest, both because of the high complexity of the problem and the high economic potential in its solution. Alternatively, there are approaches summarized under Assembly Planning (AP), which use reasoning, search and / or optimization methods to find assembly sequences. One strategy used in AP is assembly by disassembly: Finding a solution to the easier disassembly problem and then reversing the solution to create an assembly sequence. We take inspiration from this approach and apply it to RL: We train the agent to solve the disassembly task and use this to simultaneously teach it the assembly solution. To demonstrate the potential of this approach, we developed a simulation that features the assembly of toy bricks. Using this simulation we can show that our Reverse Reinforcement Learning (RRL) approach speeds up training of assembly processes significantly. We see this as an example of how RL approaches can benefit from inspiration from the field of planning
Submission Number: 9