AVO-QP: Task-Adaptive Real-Time Obstacle Avoidance for Redundant Manipulators on Edge Platforms

We address real-time obstacle avoidance for redundant manipulators where tracking and safety constraints can conflict and render quadratic programs (QPs) infeasible. We propose AVO-QP, a sensor-guided velocity-level planner that fuses RGB-D depth with learned detection to maintain situational awareness, and executes the full perception–planning–control loop on edge platforms without centralized servers. A task-adaptive avoidance vector <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$U_{\rm{avo}}$</tex-math></inline-formula> modifies the translational tracking constraint when obstacles enter a safety margin, reducing tracking–avoidance conflicts while preserving QP feasibility. A dual-objective cost balances joint-velocity magnitude and inter-step variation, and a projected primal–dual RNN solves the QP at control rate to produce smooth, feasible joint commands. Experiments on a 7-DOF Franka Panda (simulation) and a 6-DOF UR5 (hardware) with static and dynamic obstacles show an overall success rate of 84.09%, average planning times of 19.0–20.3 ms on edge devices, and terminal position errors down to 0.0001 m in static scenes while respecting distance constraints. These results suggest that AVO-QP enables responsive manipulation in cluttered workcells under tight latency budgets.