U-GPMP: A probabilistic inference-based motion planner for underwater mobile manipulators

• Introduces U-GPMP for underwater vehicle-manipulator systems using continuous-time probabilistic inference on factor graphs. • Achieves 8-25% shorter paths with improved smoothness and sub-second planning times compared to sampling-based planners. • Integrates hydrodynamic disturbances directly within Gaussian process priors for environmental uncertainty management. • Employs signed distance fields for smooth trajectory optimization with nonlinear constraint encoding in complex marine environments. Underwater mobile manipulation systems face significant challenges from hydrodynamic disturbances, vehicle-manipulator coupling effects, and limited sensing capabilities that compromise mission success in complex marine environments. This paper introduces U-GPMP, the first Gaussian process motion planning framework specifically designed for underwater vehicle manipulator systems (UVMSs), which addresses these challenges through continuous-time probabilistic inference on factor graphs. The method integrates a commercial underwater vehicle with a multi-degree-of-freedom manipulator, embedding ocean current disturbances within Gaussian process priors while encoding obstacle avoidance and system constraints as nonlinear factors. Experimental validation demonstrates superior performance over sampling-based planners, achieving 8-25% shorter paths with improved smoothness and sub-second planning times, positioning inference-based planning as a compelling solution for autonomous underwater intervention tasks.