Baby Gym: Bridging the Gap between Reinforcement Learning and Human Infant Locomotor Development

Abstract

Learning how to move is one of the most fundamental milestones humans achieve during their development, through complex interactions between neural control, biomechanics, and the environment. However, not every human learns to locomote the same way: babies exhibit remarkable variance in the stages they undergo before crawling and walking. While there exist years of empirical research quantifying and qualifying developmental stages in infant locomotion, we lack a computational model to understand how variations during the developmental stages affect overall crawling and walking behavior, thereby allowing us to test hypotheses in simulation. In order to better understand how infants learn to move, a testable model of infant locomotion would complement experimental studies allowing for model-guided interpretations of observed phenomena. This thesis work fulfills the gap in research by introducing Baby Gym, a library for probing emerged behavior through reinforcement learning (RL) on an infant-like agent with the capacity to crawl and walk, compatible with both the OpenAI Gymnasium and DM Control APIs. Baby Gym will serve as a first step in enabling a cross-disciplinary open-source ecosystem of computational models to understand infant motor development.

The work consists of the following: an extensive literature review that justifies the foundations for a baby RL environment; a Python-based infrastructure for cross-compatibility between Gymnasium and DM Control; a reproducible RL environment with several new reward functions that yield human-like locomotor development stages; and initial methods for evaluating the "human-likeness" of the emerged locomotion.