A Plan Executive Architecture for Transferable Robot Behavior: Generalized Action Plans and Their Context-Adaptive Execution in Real-World Settings

To enable the deployment of robots beyond constrained laboratory settings into open-world environments, such as homes, retail stores or agile factories, robot software must be capable of transferring to novel execution contexts with minimal reprogramming effort. While state-of-the-art systems demonstrate impressive capabilities in controlled settings, transferring them to new environments, applications and hardware platforms remains a labor-intensive process. This thesis addresses the challenge of transferability in autonomous mobile manipulation systems by proposing a novel plan executive architecture that enables the reuse of robot behavior specifications across diverse execution contexts. The central innovation lies in combining robot control programs written in an expressive robot programming language with a novel mixed symbolic–subsymbolic action representation, termed action designators. This pairing yields generalized action plans that effectively combine action control flow and contextual reasoning. A hierarchical generalized action plan for the mobile pick and place action category is developed as a case study. It demonstrates context-adaptive execution by dynamically grounding its action designators through context-specific action parameter inference. This inference is carried out via a modular infrastructure that allows to integrate alternative and / or complimenting parametrization engines: a geometric world-state based engine, a heuristics-based engine, an experience-based engine trained on execution logs and an observation-based engine that learns from human demonstrations in virtual reality. A rapid simulation step is used to validate inferred action parameters prior to real-world execution. The architecture further supports self-specialization by refining generalized plans via learning and template-based plan transformations, improving performance in specific contexts. The approach is implemented in a fully integrated robot system that includes motion control, perception and learning components. It is empirically validated across 40 simulated and six real-world execution contexts, involving five robot platforms and a variety of environments and applications. Experimental results support the thesis hypothesis that a single generalized plan can effectively transfer across a broad range of execution contexts with limited reprogramming. The plan executive satisfies key requirements for transferability, scalability, extensibility, reactivity, failure tolerance, self-improvement, usability and explainability, advancing the capability of autonomous robots to competently act in diverse, real-world contexts.