State-Conditional Adversarial Learning: An Off-Policy Visual Domain Transfer Method for End-to-End Imitation Learning
For imitation learning practitioners, this work provides a principled off-policy adversarial framework to handle domain shifts with limited target data.
The paper addresses visual domain transfer for end-to-end imitation learning when target-domain data is off-policy, expert-free, and scarce. The proposed SCAL method achieves robust transfer and strong sample efficiency in visually diverse autonomous driving environments.
We study visual domain transfer for end-to-end imitation learning in a realistic and challenging setting where target-domain data are strictly off-policy, expert-free, and scarce. We first provide a theoretical analysis showing that the target-domain imitation loss can be upper bounded by the source-domain loss plus a state-conditional latent KL divergence between source and target observation models. Guided by this result, we propose State- Conditional Adversarial Learning, an off-policy adversarial framework that aligns latent distributions conditioned on system state using a discriminator-based estimator of the conditional KL term. Experiments on visually diverse autonomous driving environments built on the BARC-CARLA simulator demonstrate that SCAL achieves robust transfer and strong sample efficiency.