Robust Policies via Mid-Level Visual Representations:
Rationale for Some Experiments Being in Simulation Due to COVID-19

In this work, we investigate the effects of using mid-level visual representations as a prior that can both accelerate learning and generalize better to unseen scenarios. Our final goal was to demonstrate this on both navigation and manipulation tasks. In the submitted paper, we successfully demonstrated the use of several mid-level representations in real-robot navigation. However, for manipulation, the closure of our lab space due to the ongoing COVID-19 pandemic resulted in us being unable to access our robotic arms.

Nevertheless, to demonstrate the utility of mid-level representation for challenging manipulation tasks, we study simulation based environments that allow us to randomize the visual and physical properties of the world. In our manipulation experiments, we experiment with ’test-objects’ previously unseen in training, and ’test-scenes’ with randomized visual textures and lighting. Although such simulation results are not a substitute for real robotic experiments, the amount of randomization in testing closely approximates the challenges of generalization in visual reinforcement learning. We also note that prior work in sim-to-real transfer [tobin2017domain, sadeghi2016cad2rl] has shown that such randomizations suffice for direct transfer to real robots.