Task-priority-aware MPPI control for human–multi-robot collaborative assembly

We propose a task-priority-aware motion-control method that links task progress to online modification of model predictive path integral (MPPI) control parameters. Human–robot collaborative assembly with multiple robotic arms requires both task-level coordination and reactive motion generation in a shared workspace. A task planner can determine which robot should proceed first, whereas a motion controller must continuously respond to the positions of the human and the other robot. However, if these functions are designed independently, fixed avoidance settings can cause symmetric yielding, deadlock, or unnecessary suspension of robot motion in spatially conflicting tasks. Therefore, we propose a method in which a task-management layer determines the current priority relation between two robotic arms and updates selected goal-tracking and collision-related cost weights, together with the scales of the robot collision models used by the local MPPI controllers. This mechanism creates task-dependent asymmetric yielding while retaining reactive avoidance of the operator and the other robot. The method does not jointly optimize the trajectories of all controllers; instead, each robot computes its motion locally under task-level priority coordination. The proposed method was implemented in a collaborative assembly-support system involving one operator and two robotic arms. The experiments demonstrate inter-robot conflict resolution, avoidance of an unexpected human approach, and concurrent task execution during human-guided positioning of a heavy object. In the demonstrated task, the mean completion time was 203.7 s under the proposed control condition and 255.4 s under manual sequential control, corresponding to a 20.2% reduction for this specific experimental workflow. Finally, we discuss the practical implementation requirements, current limitations, and scope of applicability of the proposed framework. The current method relies on predefined task sequences, event-based sequence detection, and task-specific parameter design.