Geosteering Through the Lens of Decision Transformers: Toward Embodied Sequence Decision-Making

Intelligent agents in embodied environments face the challenge of acting under uncertainty, where every decision both responds to incomplete information and reshapes the observations that follow. Building reliable world models is therefore central to long-horizon decision-making. While Reinforcement Learning (RL) has been a dominant approach, it often suffers from instability, limited interpretability, and reliance on costly online interaction. In this work, we explore Decision Transformers (DT) as a sequence modeling framework for uncertainty-aware control. As a testbed, we consider geosteering, where drilling trajectories must be continuously adjusted in real time based on indirect and noisy subsurface measurements. Our training data is generated by a dual-network deep RL (DRL) agent coupled with a particle filter (PF), embedding geological variability through probabilistic boundary estimates and noisy logs. Experiments demonstrate that longer temporal contexts allow the DT to capture delayed structural signals, leading to more consistent long-horizon trajectories. These findings position sequence modeling as a promising foundation for embodied world models in complex, uncertainty-laden decision-making domains.