Vision-Guided Nonlinear Model Predictive Control for Coordinated Dual-Manipulator Pre-Contact Capture of Tumbling Space Targets

Abstract

This paper presents a vision-guided nonlinear model predictive control (MPC) framework for coordinated dual-manipulator pre-contact acquisition of tumbling targets. Temporal monocular estimation supplies pose, motion, and uncertainty; rigid-body prediction generates fixture trajectories; and model-based or diffusion candidates initialize exact Denavit–Hartenberg (DH) refinement before clearance-checked execution. Synthetic tests give mean position and attitude errors of 0.0170–0.0182 m and 1.05–1.19°. With initial-state uncertainty and ±6% inertia mismatch, 3 s fixture-prediction root-mean-square error is 0.0822 m with 94.5% covariance coverage. In 30 paired stress scenarios, original-criterion safe capture is 19/30 for exact-DH MPC versus 14/30 for linearized MPC (exact McNemar p=0.180). A separate reviewer-motivated post hoc sensitivity imposes three-step angular-rate matching uniformly across five methods: safe capture is 19/30 versus 8/30 (Holm-adjusted p=0.0171). Diffusion and uncertainty inflation add no safe captures. Actual-cycle analysis finds a small 1.15% iteration reduction with diffusion, without an end-to-end timing benefit. Sequential mean update times are 1.234 s without diffusion and 1.392 s with diffusion, exceeding the 0.10 s simulated update interval. The evidence concerns delay-free, fixed-base kinematic supervisory planning with target-specific synthetic vision, not real-time hardware control or flight readiness.