Policy Transfer and Limitations of Spatial Action Maps

Abstract

Spatial action maps developed by Wu et al. have shown promising results in applications such as foraging, search and rescue, and exploration. These policies are trained in simulation on simple agents with no onboard sensing. Given the success that these agents demonstrate, the question of whether these policies can be transferred to new agents with more capabilities naturally arises. Here, I implement spatial action maps for foraging in simple environments on a TurtleBot3 Waffle Pi and deploy it in a Gazebo simulation. Results show that the policies only transfer effectively if measures are taken to mimic the movement patterns of the agent the policies were originally trained on. Furthermore, results indicate that path planning from sensor data in adverse environments limits the transferability of the policies.