Safety-oriented human-robot collaboration in construction through alignment of human preference

The construction industry faces significant challenges, including high incident rates and a shortage of skilled labor. With the sector’s increasing willingness to integrate robotics to improve productivity, this thesis addresses the emerging safety challenges in human-robot collaboration (HRC). A novel approach is presented in this study, which leverages human feedback regarding construction robot behaviors to define safety-oriented control policies (i.e., a series of control actions by robots). The proposed approach begins with designing an online labeling tool tailored for collecting human preference data regarding robot behaviors in collaborative tasks. Utilizing this data, this study develops and trains a score model using RankNet, a network architecture for learning pairwise comparisons, enabling prioritization of safer robot behaviors. Afterwards, safety-oriented policies can be inferred to guide robot actions. The proposed methodology extends to the practical application where three specific construction tasks (hammer handover, travel handover, and rebar lifting) involving robots are used to test the effectiveness of the generated policies. The results reveal that the preference-based score model not only effectively predicts safer robot behaviors but also closely aligns with the collective safety preferences of multiple annotators who might provide different opinions towards the same robot behavior. Given the subjective nature of safety evaluations in HRC, the proposed approach is capable of mitigating the biases inherent in individual preferences, providing a more objective and reliable means of identifying and implementing safer robot behaviors in construction environments. Real robot experiments demonstrate the ability of the proposed approach to adjust robot behaviors dynamically in response to the movements and actions of human collaborators. The robustness and adaptability of the proposed approach indicate that it can be deployed across a wide spectrum of construction activities without the need for extensive retraining or customization for each specific task. In summary, this research underscores the importance of aligning construction robot behaviors with human preferences, offering a scalable solution to enhance occupational safety for construction.