Behavioural cloning for robots in unstructured environments

Behavioural cloning is an approach where an agent observes the actions of a human operator and then tries to generalise their behaviour. Successful applications include pilotting aircraft in simulation (Sammut et al., 1992) and operating container cranes (Urbancic and Bratko, 1994). In this work, we examine the application of behavioural cloning to autonomous navigation of a robot in an unstructured environment. In particular, we examine the traversal of the random stepfields introduced in the Robocup Rescue Robot League competition using tracked vehicles. One critical issue in behavioural cloning is representing the state of the environment in a manner amenable to machine learning. Our sensing equipment allows a faithful reproduction of the terrain. The representation we use is based on discussions with operators as to how they view the task; and involves breaking up the problem into the general ”lay of the land” around the robot, and the particular obstacles (both protrusions and holes) that must also be taken into account.