Industrialized construction imposes stringent precision requirements on the robotic assembly of modular components such as prefabricated window units. In such tolerance-critical operations, the central bottleneck is not only mechanical clearance but also the difficulty of converting tacit installer expertise into data-efficient autonomy: sparse acceptance feedback, pronounced contact variability, and millimeter-scale geometric constraints jointly constitute a formidable sample-efficiency challenge. This study presents an installer-in-the-loop interactive reinforcement learning framework that acquires such expertise through three complementary channels: offline teleoperated demonstrations, sparse event-driven binary takeovers at contact-failure boundaries, and acceptance-aligned terminal rewards, all logged under a unified schema for traceable offline-to-online adaptation. A temporally abstract action-sequence policy built on Q-chunking with Flow Q-Learning captures multimodal recovery maneuvers under sparse terminal rewards, while a non-updating warm-start phase stabilizes the offline-to-online transition. The framework is evaluated in MuJoCo across the full operational workflow, from suction acquisition through clearance-limited seating, under both structured staging and end-to-end randomized placement. Within a defined stress-test regime characterized by 2 mm per-side clearance, bounded pose perturbations, and friction randomization, the proposed pipeline attains 100% autonomous seating with 12–15 min of cumulative installer supervision over 3.0 h of online training (12 min in Experiment A; 15 min in Experiment B), and reaches the 95% success milestone in approximately 0.5 h and 1.5 h, respectively. Beyond success rate, wall-clock adaptation time, cumulative takeover minutes, intervention-rate decay, and stage-wise failure attribution are reported to inform supervision budgeting. Ablations isolate the complementary contributions of temporal abstraction, installer intervention, and warm-start value calibration.
Jin et al. (Wed,) studied this question.