BRAC-ACT: a task-structured robotic assembly framework for multi-stage Song-style dougong assembly

Traditional timber assembly, especially dougong assembly, involves strict sequential dependencies, complex spatial relationships, and long-horizon execution. This study aims to develop and evaluate BRAC-ACT, a task-structured robotic assembly framework that translates Song-style dougong tectonic logic into learnable, executable, and evaluable robotic tasks. A preliminary whole-task pilot was first conducted to examine complete-sequence ACT execution and informed the subsequent task-structured design. The formal experiment used seven task-specific ACT policies trained from 50 demonstrations per task on a bimanual robotic platform with 1:10 scale 3D-printed, magnet-assisted specimens. Across 300,000 training steps, all seven policies showed progressive stabilization of the recorded training-loss and gradient-norm metrics. Across ten evaluation rounds, six completed the full Task 1–7 sequence, yielding a 60.0% round-level completion rate under a retry-enabled procedure in which individual tasks could be attempted repeatedly before progression or round termination. Task-wise analysis identified Tasks 6 and 7 as the main bottlenecks, with failures concentrated in placement and spatial alignment rather than grasping. The study demonstrates the feasibility of structuring traditional architectural assembly knowledge into task-specific imitation-learning policies, chained execution, and process-level diagnostic evaluation, providing a methodological basis for translating complex tectonic knowledge into robotic assembly workflows. The results are limited to scaled magnet-assisted specimens in a controlled workspace and should not be interpreted as intervention-free autonomous or full-scale timber assembly performance.