This year's research activities focused on developing bridging techniques across the perception and planning layers of intelligent robotic systems, addressing core challenges in how autonomous agents understand their environment and act within it.

On the planning side, a recurring difficulty in deploying robotic systems across varied environments is that symbolic representations — the predicates and interfaces through which agents interpret and act on the world — cannot be assumed stable or semantically transparent. To address this, we developed a closed-loop planning pipeline capable of operating without lexical priors, using a modular architecture in which specialized components handle constraint extraction, plan generation, symbol-level translation, validation, and iterative repair. The coordination of these components reflects principles from multi-agent design, where distinct functional units collaborate across boundaries to achieve outcomes no single module could reach alone. Evaluation on established planning benchmarks confirmed that the architectural integration itself, rather than any individual component, drives performance gains — a finding with direct implications for how robotic planning systems should be structured.

On the perception side, meaningful human-robot interaction demands that robotic agents accurately interpret human affective states, of which facial expression is among the most immediate signals. Existing approaches treat detection and recognition as sequential stages, sacrificing joint optimization and overlooking the geometric structure of the face. We developed an end-to-end dual-stream framework that unifies face localization, expression classification, and keypoint regression within a single network, with cross-stream fusion modules bridging visual semantics and structural geometry. Experiments confirmed improvements in both detection accuracy and robustness, particularly for expression categories that are geometrically distinctive and affectively salient in human-robot interaction scenarios.

These two works together address complementary requirements for robotic agents operating in human environments — the capacity to plan reliably under symbolic uncertainty, and the capacity to perceive human intent through facial expression — contributing toward the broader goal outlined in the original research proposal.