Software Architectures for Enhanced Physical Human-Robot Interaction and Mixed Reality-based Rehabilitation

Physical human-robot interaction (pHRI) and mixed reality (MR)-based robotic rehabilitation impose stringent requirements on software systems: safety-critical real-time execution, heterogeneous sensor integration, and support for collaborative multi-user workflows. Conventional research prototypes, typically monolithic and single-threaded, lack the reliability and extensibility required for large-scale studies. This dissertation develops software architectures for pHRI and MR rehabilitation, guided by two principles: (1) parallel architectures that deliver real-time performance through non-blocking execution, and (2) modular architectures that enable scalable development through clear functional separation and decoupled communication.

For pHRI, these principles are instantiated in a shoulder rehabilitation exoskeleton with lock-free multi-threaded pipelines, a wearable upper-limb exoskeleton on a layered ROS architecture, and a robotic arm manipulator driven by a distributed task scheduler. They extend to MR rehabilitation through a gait-symmetry AR visual-distortion system, a single-user AR admittance platform, and a multi-user MR platform that formalizes visualization–control separation through policy-based configuration. They further extend to multi-robot and multi-agent intelligence through an MR materials-synthesis platform, a Single Source of Truth architecture deployed across heterogeneous robot platforms at two institutions, a multi-agent AI platform with retrieval-augmented discussion and four-level human-in-the-loop approval, a personal AI infrastructure integrating four language-model providers with a real-time C++ core, and a three-rate interaction-control architecture composing language-model reasoning, diffusion-based action generation, and Lie-group impedance and admittance control.

Validation supports the architectural claims. The shoulder exoskeleton achieves 250 Hz control with high impedance reliability (R² > 0.97) in a forty-participant study; the wearable exoskeleton attains 500 Hz with >99.2% VAF in impedance replication; the robotic arm manipulator reduces interaction energy by approximately 45% through session-continuous Bayesian optimization. The multi-robot platform integrates heterogeneous robots from multiple manufacturers through YAML-only configuration with zero per-platform source-code modification. The multi-agent AI platform produced a sole-authored nuclear waste site compliance assessment that received "Paper of Note" and "Superior Paper" awards at the Waste Management Symposia 2025. The multi-user MR platform sustains real-time synchronization across head-mounted displays through a LAN-local authoritative server with a mesh VPN overlay. Together, these architectures transform prototype-level pHRI and MR rehabilitation systems into reliable platforms for long-term and multi-site studies.