LG AI Research: Presents an Offline Imitation Learning Methodology That Learns from State Information Alone

Traditional Imitation Learning has required expert action information, but humans can also learn from observation alone, without action information. Research on Learning from Observation, which mimics this, has been conducted, but it has the drawback of requiring direct interaction with the environment, which entails high cost and risk.

This paper aims to solve the offline Learning from Observation problem, which utilizes pre-collected, incomplete demonstrations instead of risky real-time interaction. Existing Off-policy approaches are unstable due to the use of out-of-distribution action values, while approaches utilizing an Inverse Dynamics Model have the limitation of difficulty in accurately predicting actions in stochastic environments.

To address this, this research presents a new algorithm based on mathematical formulations. The core idea is to constrain the model from deviating from the data support via KL Regularization, achieving stable performance through an optimization problem that balances state-transition distribution matching with the incomplete demonstration distribution.