Push-Wiper: Toward General-Purpose Robotic Cleaning across Varied Stains and Surfaces with Segmented Pushing Trajectories
Renhao Lu, Mingxin Wang, Chenyang Cao, Yang Yang, Guoping Pan, Kangkang Dong, Yi Cheng, Houde Liu
IROS 2026)
cs.RO
2026-08-01
Wiping spreads viscous stains. Push-Wiper gathers them with a sponge via segmented pushing, scoring 89.9 vs 44.98 baseline, and zero-shot transfers to liquids, solids and curved surfaces.
Viscous stains (ketchup, peanut butter; high-viscosity, complex-rheology messes) are a hard case for robotic surface cleaning. Ordinary wiping spreads the stain; scrubbing gives enough friction but risks scratching the surface. This paper reframes the problem: stop trying to wipe the stain away, gather it and carry it off instead.
Push-Wiper uses a sponge and recasts viscous cleaning into a two-phase "gather, then finish" routine built on segmented pushing trajectories.
In the gathering phase, the observation is first abstracted into a texture-less binary stain map (stain = 0, clean = 1), forcing the policy to learn how to aggregate mess rather than fit visual textures or surface geometry. A Diffusion Policy then acts as a low-frequency macro planner that emits discrete pushing strokes, with the action space restricted to the plane (translations xb, yb plus yaw rotation Δθ), producing only planar aggregating motions.
Lifting 2D plans onto 3D surfaces is the job of ASPI (Arbitrary Surface Pose Interpolator). It projects each 2D action onto the surface, extracts the contact point and surface normal, and synthesizes a 6D trajectory with B-spline fitting, trapezoidal velocity parameterization and spherical linear interpolation, keeping the tool's z-axis aligned with the surface normal.
Contact force is handled by a hybrid force-position controller: an admittance controller regulates normal force to 20N while the other axes follow position control, staying in contact without pressing too hard.
Once the stain is gathered, the post-processing phase runs predefined motion primitives: dabbing to lift residue, scraping, rinsing and squeezing to self-clean the sponge, then a final full-coverage wipe. Sponge self-cleaning is what lets the method keep working continuously.
Cleaning score CS = (1 minus stain area after, divided by stain area before) times 100. On the two training stains (ketchup, peanut butter):
| Method | Ketchup avg | Peanut butter avg | Overall avg |
| Full-Cover | 53.99 | 11.28 | 32.64 |
| PushAll-Onetime | 58.04 | 31.99 | 44.98 |
| Push-Wiper | 92.30 | 87.46 | 89.88 |
Push-Wiper's overall 89.88 more than doubles the strongest baseline, and the paper reports scores "up to 130% higher than baseline." Peanut butter is the hardest case (full-cover wiping scores 11, essentially just smearing it), where Push-Wiper reaches 87.
Generalization is the brighter story. Training used only flat surfaces and only ketchup and peanut butter, yet zero-shot transfer reaches 100.00 on solid residue (black disks), 92.62 on liquid spills, and 94.42 on combined viscous stains; on curved surfaces unseen in training, convex scores 91.45 and concave 93.42. With the post-processing phase the overall score climbs from 89.44 to 99.25. Data collection used only 448 trajectories (75 ketchup and 75 peanut butter tasks), augmented to 2,688.
Cleaning sticky messes is one of the most common and most frustrating tasks for home and service robots, and mainstream manipulation policies are largely helpless with rheologically complex, shape-changing objects. The contribution is not a leaderboard number but a reformulation of an open problem into a solvable "aggregation" task; paired with surface interpolation and sponge self-cleaning, it generalizes across stains and surfaces from very little data. It is accepted at IROS 2026.
The hardware bar is non-trivial: it needs a six-axis force/torque sensor and a wrist camera, so a basic arm cannot just run it. Sensing is wrist-camera-only and is often occluded by the arm itself, patched with low-frequency replanning. The sponge needs self-cleaning and moisture management, but continuous-duty lifetime and maintenance are not discussed. Validated stain types are limited (ketchup, peanut butter, black pepper sauce, oyster sauce and the like), with no broader material coverage. The paper also names human-robot safety constraints as future work, meaning no explicit human-safety design exists today.