Spotting Inventory: Autonomous Item Localization on Active Construction Sites Using a Boston Dynamics Spot® with Bayesian Filtering and Object Detection

Abstract

Accurate and complete location information is crucial on construction sites for effectively tracking resources and ensuring the successful completion of a project. However, the construction sector has numerous inventory management challenges, including dependence on outdated technology and the prevalence of manual processes. Moreover, the absence of real-time data, coupled with the dynamic nature of material movement, exacerbates the challenges of effective inventory management. Human-based tracking requires considerable manual labor and time, and when resources are managed over prolonged periods, it is also susceptible to human error.

These challenges highlight the essential requirement for automation to enhance efficiency and productivity. Previous studies have demonstrated the benefits of employing auto-identification and localization technologies to automate asset tracking on construction sites. Numerous case studies and field testing have proven the benefits of technology-enabled onsite materials monitoring compared to human onsite materials management. To this end, integrating computer vision with RFID can improve the precision of material identification and location, offering a more thorough solution for inventory management.

This dissertation has employed a sensor fusion of complementary technologies to automate inventory management on construction sites. The hybrid system consisted of a quadruped robot enhanced with RFID-based recursive probabilistic localization and custom object detection. The objective was to identify and locate inventory elements on construction sites using an autonomous quadruped robot. This integration can facilitate the advancement of more complex systems for physical asset management. Integrating RFID, computer vision, and robotics for inventory management on construction sites represents a significant advancement in the construction industry, addressing traditional challenges such as inefficiency, labor-intensive processes, and poor inventory management.

This dissertation predominantly utilized exploratory experimental quantitative research as its core methodology, supplemented by qualitative approaches to establish the research need and to validate the conceptual framework as the tangible outcome of the dissertation. Interview findings revealed that asset tracking decisions on construction sites are primarily driven by cost, visibility, and perceived risk. Large, high-value equipment is prioritized, while smaller tools are often excluded due to the high relative cost of tracking technologies. A predominantly reactive approach persists, relying on manual checks and ad hoc processes influenced by site-specific factors. This research successfully deployed a localization algorithm on an autonomous mobile platform in the context of RFID-based probabilistic localization. The performance assessment showed that the system could read all installed tags with a median localization error of 1.569 meters/ 5 feet 2 inches for the four power levels tested. The 1.569-meter localization error may suffice for coarse-grained tracking, e.g., identifying pallets of bulk materials in open laydown yards. For custom object detection-based identification and localization, the experimental findings indicate that the combination of low asset height, higher Spot® height, and a 0° incidence angle between Spot® and the asset markedly improves the system's ability to detect assets from extended ranges. Of all evaluated configurations, the combination of Medium Spot® speed, High Spot® height, and Low asset height, especially at a 0° incidence angle, yielded the greatest median distance and the most concentrated observation distribution, rendering it the most suitable configuration overall.

The dissertation's tangible outcome is a conceptual framework that optimally localizes physical inventory in an unmapped space by synchronizing two mutually exclusive localizations coming from RFID based probabilistic and custom object detection-based approaches. A weighted fusion attains the finalized localization of a typical physical asset; this approach will ensure that the more reliable sensor at any given moment contributes more heavily to the final localization estimation (i.e., optimum localization), while the less reliable sensor has a diminished influence.

For future work, the proposed conceptual framework can be enhanced to incorporate safety-critical features, including human proximity detection, real-time worker monitoring, and emergency override capabilities, to ensure practical applicability on actual construction sites. As construction sites progressively integrate mobile platforms, such as drones and terrestrial robots, future developments must facilitate localizations facilitated by multi-robotic systems, their reliable coordination by incorporating communication protocols, collision avoidance mechanisms, and collective spatial awareness. Additionally, the growing complexity of construction projects necessitates the concurrent tracking of diverse physical assets. This introduces technical challenges related to sensor data alignment, information overload management, signal interference in dense environments, and system scalability. Addressing these challenges will be critical to advancing the framework toward deployment-ready, scalable solutions for collaborative, automated construction sites.