Don't do things you can't undo: Reversibility models for generating safe behaviours

We argue that an ability to determine the reversibility of actions allows a robot to identify safe behaviors autonomously. We introduce a notion of reversibility model and give a definition of model refinement. We implement this on a real robot and observe that, when a reversibility model is refined by the addition of proximity sensors, obstacle avoidance emerges as a side-effect of avoiding irreversible actions. We interpret this as evidence of a deep connection between reversibility and safe behaviour. We also observe that, on the real robot, reversiblities are learned as efficiently as a dedicated reward function. We conclude that reversibility identification may provide an abstract and yet practical method of generating a variety of safe behaviours.