Dancing Between The Lines: Enabling Hierarchical Preference Learning for Constraint-Safe Human–Robot Co-Planning

Abstract

As robots transition from controlled settings into shared workspaces, they must continuously personalize their behavior to individual human collaborators while respecting safety constraints. This thesis introduces Dancing Between the Lines (DBTL), an extension of the Coloring Between the Lines (CBTL) framework for constraint-safe human–robot co-planning. DBTL replaces CBTL’s supervised classification approach with a hierarchical Bayesian preference model that enables cross-context and cross-human transfer through variational inference, a transient state model that separates persistent preferences from session-level effects and a variance-weighted exploration mechanism that leverages posterior uncertainty for active learning. We evaluate DBTL in two simulated collaborative environments and a 16 participant real-human case study. In multi-context settings with varying preferences, DBTL outperforms both flat and context-agnostic baselines, demonstrating that the hierarchical structure provides meaningful value. On cross-recipe transfer, the learned hierarchical prior yields a consistent warm-start advantage on held-out contexts. A case study with real participants confirms that the system produces interpretable, group-coherent posteriors from limited interaction data. Together, these contributions provide a framework for constraint-safe, personalized human–robot collaboration in long-horizon settings. The implementation and all experiment code are publicly available at https://github.com/arav-raval/co_plan_multitask-personalization.