Push-Wiper: Toward General-Purpose Robotic Cleaning across Varied Stains and Surfaces with Segmented Pushing Trajectories

Push-Wiper reframes viscous-stain cleaning as a state-aggregation problem rather than conventional wiping: a sponge uses segmented pushing trajectories to progressively contract scattered material into one compact region, then predefined dabbing, scraping, rinsing, squeezing, and final-wipe primitives remove the residue and clean the sponge. A Diffusion Policy receives texture-independent binary stain maps and predicts low-dimensional 3-DoF actions—planar translation and yaw—while the Arbitrary Surface Pose Interpolator, B-spline smoothing, trapezoidal velocity control, and hybrid force–position admittance controller convert them into contact-stable 6-DoF motion on arbitrary surfaces. Using 448 demonstrations augmented to 2,688 trajectories, the system runs a DDIM-based diffusion model with 16 actions per plan and a 20 N normal-force setpoint. On ketchup and peanut butter, Push-Wiper reaches an overall Cleaning Score of 89.88, versus 32.64 for repeated Full-Cover sweeping and 44.98 for one-shot global pushing, corresponding to a reported improvement of up to 130% over baselines. Without retraining, it achieves zero-shot transfer to convex and concave surfaces, where scores are 91.45 and 93.42, and to solid residues, liquid spills, and unseen viscous stains. Post-processing raises the average score from 89.44 to 99.25, including 100.00 on ketchup and 98.51 on peanut butter. The experiments use an NVIDIA RTX 4060Ti workstation, but the central advance is the policy–geometry decoupling that enables generalization beyond the training surface and stain types.