Learning viewpoint control from human-initiated transitions for teleoperation in construction

Visual perception is critical for teleoperation in construction, where optimal visibility directly impacts task performance. Hybrid viewpoint control systems enhance the flexibility of visual perception by adaptively coupling or decoupling the viewpoint from robot movements according to situational demands. However, determining the optimal timing for transitions between these perspectives remains a major challenge, as existing autonomous methods are not directly applicable to hybrid control for construction tasks. In this work, we propose a viewpoint control mode prediction model that autonomously manages transitions during teleoperation with hybrid control. Our learning scheme with a transition-guided weighting method leverages sporadic transition commands from human interactions with the teleoperation system as demonstration data for imitation learning. User evaluation in a virtual reality (VR) environment simulating construction welding tasks shows that our model outperforms the baselines, achieving an 11% improvement over the state-of-the-art behavioral cloning (BC) algorithm and a 19% improvement over the state-of-the-art weighted BC algorithm in replicating human transition behaviors. This work contributes novel insights into the design of visual perception systems for teleoperation in construction, enabling reliable, user-aligned viewpoint transitions.