Identifying Reusable Early-Life Options

Abstract

We introduce a method for identifying short-duration reusable motor behaviors, which we call early-life options, that allow robots to perform well even in the very early stages of their lives. This is important when agents need to operate in environments where the use of poor-performing policies (such as the random policies with which they are typically initialized) may be catastrophic. Our method augments the original action set of the agent with specially-constructed behaviors that maximize performance over a possibly infinite family of related motor tasks. These are akin to primitive reflexes in infant mammals - agents born with our early-life options, even if acting randomly, are capable of producing rudimentary behaviors comparable to those acquired by agents that actively optimize a policy for hundreds of thousands of steps. We also introduce three metrics for identifying useful early-life options and show that they result in behaviors that maximize both the option's expected return while minimizing the risk that executing the option will result in extremely poor performance. We evaluate our technique on three simulated robots tasked with learning to walk under different battery consumption constraints and show that even random policies over early-life options are already sufficient to allow for the agent to perform similarly to agents trained for hundreds of thousands of steps.