The Distracting Control Suite -- A Challenging Benchmark for Reinforcement Learning from Pixels

Robots have to face challenging perceptual settings, including changes in\nviewpoint, lighting, and background. Current simulated reinforcement learning\n(RL) benchmarks such as DM Control provide visual input without such\ncomplexity, which limits the transfer of well-performing methods to the real\nworld. In this paper, we extend DM Control with three kinds of visual\ndistractions (variations in background, color, and camera pose) to produce a\nnew challenging benchmark for vision-based control, and we analyze state of the\nart RL algorithms in these settings. Our experiments show that current RL\nmethods for vision-based control perform poorly under distractions, and that\ntheir performance decreases with increasing distraction complexity, showing\nthat new methods are needed to cope with the visual complexities of the real\nworld. We also find that combinations of multiple distraction types are more\ndifficult than a mere combination of their individual effects.\n