Scalable Causal Imitation Learning

This paper addresses causal imitation learning when expert demonstrations contain unobserved confounders and the expert and imitator have mismatched observations. It introduces Causal Soft Q Imitation Learning (Causal SQIL) and Causal Inverse soft-Q Learning (Causal IQ-Learn), combining causal adjustment with off-policy inverse reinforcement learning objectives. The methods approximate the sequential π-backdoor criterion through a fixed-size sliding window over causally adjusted representations. According to the abstract, evaluations in confounded continuous-control environments show substantially better long-horizon performance than prior Causal BC and Causal GAIL methods, with some results surpassing the expert, while causally unaware baselines fail to learn meaningful behavior.