Real Robot Challenge Phase 2 - ThriftySnipe

Anonymous

Keywords: Reinforcement Learning

TL;DR: We attempt to use reinforcement learning and image processing to solve Phase 2 of the Real Robot Challenge, but ultimately come up short.

Abstract: In our attempt to solve Phase 2 of the Real Robot Challenge, we reinforcement learn a control policy, in simulation, conditioned on delivering a single chosen dice to a single chosen goal. To further simplify the problem faced by the policy, we provide it with privileged state information from the simulator. Upon transfer to the real robot, an image processing technique is employed to estimate this state information. Ultimately, Phase 2 proves significantly more difficult than Phase 1 and our method struggles to solve the task, both in simulation or indeed reality.