Data-Automated Policy Learning for Nonlinear Welfare

Two papers on policy learning and the ExpertGen framework were published on arXiv.org, detailing advancements in automating policy learning for nonlinear welfare criteria and sim-to-real transfer.

A paper submitted on June 1, 2026, to arXiv.org^[1] explores policy learning from observational data, focusing on nonlinear welfare criteria in binary treatment settings. The authors model a nonlinear welfare criterion using a utility function that encompasses potential outcomes and intermediate parameters, capturing higher moments of outcome distributions. To address bias in machine learning estimates, the paper introduces a novel reweighting-based debiasing approach. The policy-learning process is automated using sieve approximations and K-fold cross-validation for model selection. The proposed method satisfies an oracle inequality, providing theoretical guarantees on its performance. A separate paper on arXiv.org^[2] introduces ExpertGen, a framework that automates expert policy learning in simulation for scalable sim-to-real transfer. ExpertGen initializes a behavior prior using a diffusion policy trained on imperfect demonstrations and uses reinforcement learning to optimize the diffusion model's initial noise. ExpertGen achieved a 90.5% overall success rate on industrial assembly tasks and 85% on long-horizon manipulation tasks^[2].