CCIL: Continuity-Based Data Augmentation for Corrective Imitation Learning

We present a new technique to enhance the robustness of imitation learning methods by generating corrective data to account for compounding errors and disturbances. While existing methods request interactive experts, additional offline datasets, or domain-specific invariances, our approach requires minimal additional assumptions beyond access to expert data. Our key insight is to leverage local continuity in the environment dynamics to generate corrective labels. Our method constructs a dynamics model from expert demonstrations, emphasizing local Lipschitz continuity in the learned model. In regions exhibiting local continuity, our algorithm generates corrective labels within the neighborhood of the demonstrations, extending beyond the actual set of states and actions in the dataset. Training on the augmented data improves the agent's resilience against perturbations and its capability to address compounding errors. To validate the efficacy of our generated labels, we conduct experiments across diverse robotics domains in simulation, encompassing classic control problems, drone flying, navigation with high-dimensional sensor observations, legged locomotion, and tabletop manipulation. See all our experiments at https://personalrobotics.github.io/CCIL/ .