Perception and Validation for Robotic Industrial Inspection

Research Context

Robotic inspection systems are increasingly deployed in industrial and infrastructure environments where manual inspection is slow, dangerous, or inconsistent. Computer vision is central to these systems — cameras and depth sensors are the primary way robots perceive what they are inspecting. But fielding a perception system that is accurate in a laboratory is different from fielding one that is reliable in an industrial environment with variable lighting, surface conditions, occlusion, and operational constraints.

This MSc research, conducted at the University of Regina under the Master of Science in Industrial Systems Engineering program, addresses the gap between perception system development and industrial deployment readiness. The work connects directly with parallel internship experience at NSERC and Mitacs, where robotic inspection and industrial computer vision are applied in operational contexts.

Research Questions

The work is organized around three connected questions:

How should perception pipelines for robotic inspection be structured to handle real-world variation? Industrial inspection environments are not controlled. Lighting changes between shifts. Surface conditions vary with equipment age and maintenance cycles. The pipeline architecture — including preprocessing, model design, calibration, and failure handling — needs to be designed around this variability rather than despite it.

What validation methodology accurately predicts deployment performance? Standard computer vision benchmarks measure performance on data drawn from the same distribution as training data. Industrial deployment exposes systems to distribution shift, edge cases, and failure modes that held-out test sets do not capture. Developing validation protocols that predict real deployment behavior — including out-of-distribution robustness and failure mode characterization — is a core research objective.

How does sensor integration affect perception reliability on a robotic platform? Camera sensors do not operate in isolation. Vibration, thermal effects, lens contamination, and mounting geometry all affect image quality in ways that are not present in static lab setups. The research examines how these factors affect perception outputs and what mitigation strategies are effective.

Methods

Research involves developing and evaluating computer vision pipelines on robotic inspection hardware, with experimental validation designed to reflect real operating conditions rather than idealized test scenarios.

Sensor integration work is conducted on ROS2-based platforms, with perception nodes interfaced to camera and depth sensor hardware. The ROS2 framework enables structured experimentation with pipeline architecture — preprocessing stages, inference nodes, and post-processing logic can be replaced or reconfigured without changing the surrounding system.

Quantitative evaluation uses metrics appropriate to the inspection task, including precision-recall at operating thresholds, performance across deliberately varied capture conditions, and timing analysis under the real-time constraints of a moving inspection platform.

Connection to Practice

The research runs parallel to applied work at NSERC (computer vision pipelines for robotic inspection) and Mitacs at MacDon Industries (ML model development and validation for industrial perception). This connection between research methodology and applied engineering context is intentional — the validation frameworks and pipeline design principles developed in research are evaluated against the constraints and failure modes encountered in real industrial deployment.

Expected Contribution

The anticipated output is a validated framework for designing and evaluating computer vision perception pipelines for industrial robotic inspection — including pipeline architecture recommendations, validation protocol design, and experimental characterization of how sensor integration factors affect perception reliability. The target completion is Spring 2026.