Abstract
Reliable self-localization is a fundamental requirement for autonomous mobile robots operating in indoor environments, particularly for service robots that must function safely and robustly in dynamic, human-centered spaces. This thesis presents the design, implementation, and evaluation of a particle filter (PF)–based localization system for the Toyota Human Support Robot (HSR), with a specific focus on fully passive global localization under strict stationary conditions. Unlike standard approaches that rely on exploratory motion to resolve pose ambiguity, the proposed method emphasizes particle initialization, a raycasting-based measurement model, and systematic resampling to enable convergence using perceptual information alone.
The PF is evaluated across two localization scenarios: pose tracking and passive global localization. Performance is quantitatively assessed in both a high-fidelity simulation using NVIDIA Isaac Sim and real-world experiments conducted in a laboratory environment, with operational comparison against the Toyota-provided Adaptive Monte Carlo Localization (AMCL) baseline. Experimental results demonstrate that the proposed PF achieves centimeter-level positional accuracy and sub-degree orientation error during pose tracking, while consistently recovering from complete belief resets without requiring robot motion. In contrast, the Toyota-provided AMCL baseline exhibits degraded performance or failure under the same passive recovery conditions.
These findings highlight the viability of fully passive particle filter-based localization for service robots and address a gap in existing experimental evaluations, offering practical insights for robust deployment in domestic and competition environments such as RoboCup@Home.