CubeDAgger: Interactive Imitation Learning for Dynamic Systems with Efficient yet Low-risk Interaction

This research presents CubeDAgger, a novel method for interactive imitation learning designed to enhance robustness in dynamic systems while minimizing stability violations. Building on the EnsembleDAgger baseline, it introduces three key improvements: regularization to explicitly activate supervision timing thresholds, transformation of the expert-agent switching system into an optimal consensus framework for multiple action candidates, and injection of autoregressive colored noise for time-consistent exploration. The method is validated through simulations demonstrating robust policies with maintained dynamic stability, and real-robot experiments show successful learning from scratch using only 30 minutes of interaction with a human expert.