PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2608.00730v1cs.RO

TL;DR

Robotic cleaning lacks a single strategy for viscous stains and varied surfaces without spreading contaminants or requiring retraining. Push-Wiper aggregates stains through segmented pushing trajectories and achieves up to 130% higher cleaning scores than baseline methods while generalizing zero-shot across stain types and curved surfaces.

  • Problem

    Robotic cleaning lacks a single strategy for multiple stain types on arbitrary curved surfaces, while viscous-stain cleaning remains specifically underdeveloped.

  • Method

    Push-Wiper reformulates viscous-stain removal as aggregation, using segmented pushing trajectories and a stepwise Diffusion Policy to gather stains before post-processing.

  • Results

    Up to 130% higher cleaning scores than baseline methods are achieved, with zero-shot generalization to solid residues, liquid spills, unseen viscous stains, and curved surfaces.

  • Takeaways & Limitations

    Push-Wiper provides a cleaning framework that handles diverse stain distributions, stain types, and surface geometries without additional training.

Abstract

from arXiv · show

Viscous stains, characterized by high viscosity and complex rheological properties, remain a major challenge for robotic surface cleaning. Conventional wiping often spreads the stain, while scrubbing provides stronger friction but risks damaging the surface. In this paper, we propose Push-Wiper, a framework that reformulates viscous stain cleaning as an aggregation problem. Push-Wiper employs a sponge to progressively gather stains through segmented pushing trajectories, followed by a post-processing phase that detaches the aggregated material and enables sponge self-cleaning. We adopt a stepwise strategy for stain gathering and leverage Diffusion Policy to generate adaptive pushing action sequences. These sequences are executed through our Arbitrary Surface Pose Interpolator (ASPI) and a hybrid force-position controller, allowing the method to generalize to stains with diverse spatial distributions. Push-Wiper achieves a cleaning score (CS), defined as the percentage of stain area removed, up to 130% higher than baseline methods. Without additional training, Push-Wiper also transfers in a zero-shot manner to solid residues, liquid spills, unseen viscous stains, and curved surfaces with varying geometries. Our experiments demonstrate the cleaning effectiveness of Push-Wiper and its strong generalization ability. The project website is available at https://push-wiper.github.io/.

I. INTRODUCTION

Viscous stains are difficult to remove because their high viscosity and complex rheology cause smearing, while effective cleaning requires controlled normal force and tangential friction. Push-Wiper addresses this challenge by reformulating cleaning as aggregation and combining a low-frequency Diffusion Policy with ASPI and hybrid force–position control.

  • Challenge: Viscous stains are difficult to remove because they often smear or spread and require controlled normal force plus sufficient tangential friction.Restricted liquid use in tabletop settings further complicates cleaning.
  • Challenge: Scrubbing can remove viscous stains but risks surface abrasion, whereas wiping can spread viscous substances and cause secondary contamination.These motions therefore involve competing effectiveness and surface-safety limitations.
  • Approach: Push-Wiper reformulates viscous stain removal as a state-aggregation problem using an “aggregate-then-finish” strategy.This paradigm is designed to overcome the limitations of conventional wiping and scrubbing.
  • Approach: Push-Wiper couples a low-frequency Diffusion Policy for 3D action generation with ASPI and hybrid force–position control for 6D trajectory.The decoupled architecture isolates the policy from 3D surface variations, reducing problem complexity.
  • Results: 130%: Push-Wiper improves cleaning score by up to 130% over baselines and generalizes zero-shot to curved surfaces, solid residues, liquid spills, and novel viscous stains.The reported generalization is achieved by abstracting the task into geometric topologies rather than complex visual textures or 3D surface geometries.

II. RELATED WORK · III. PUSH-WIPER FRAMEWORK

The paper contrasts classical and learning-based robotic cleaning methods, highlighting their limited portability and generalization to varied stains and curved surfaces. It then introduces Push-Wiper, which aggregates viscous stains through segmented pushing and adaptively controlled trajectories before detachment and sponge self-cleaning.

  • II. RELATED WORK: Robotic surface-cleaning methods are categorized as classical or learning-based according to how they generate action trajectories.
  • II. RELATED WORK: Classical methods execute predefined trajectories with position, force, or compliant control for coverage planning and physical interaction.
  • II. RELATED WORK: Learning-based methods primarily use reinforcement learning or imitation learning to derive cleaning policies from rewards or demonstrations.
  • II. RELATED WORK: Classical methods are stable for predefined tasks but depend on precise models and hand-tuned parameters, limiting portability across platforms and environments.
  • II. RELATED WORK: Learning-based methods adapt within specific scenarios, but new conditions often require redesign or retraining; no single strategy handles multiple stains on arbitrary curved surfaces.
  • III. PUSH-WIPER FRAMEWORK: Push-Wiper reformulates viscous-stain cleaning as aggregation, progressively gathering material with segmented sponge-pushing trajectories before detachment and sponge self-cleaning.The framework uses a stepwise gathering strategy and a post-processing phase.
  • III. PUSH-WIPER FRAMEWORK: Push-Wiper achieves a cleaning score up to 130% higher than baseline methods and transfers zero-shot to solid residues, liquid spills, unseen viscous stains, and curved surfaces.Cleaning score is defined as the percentage of stain area removed.

A. Problem decomposition

Push-Wiper formulates viscous-stain cleaning as progressively aggregating stain pixels through segmented pushing, then applying post-processing to remove the compacted material. This decomposition avoids directly wiping stains and motivates a data-driven approach because hidden rheological parameters make classical planning intractable.

  • Problem formulation: The cleaning objective represents the surface D as an M × N grid whose cells Di,j ∈{0, 1} indicate dirty and clean states.The stated objective is to completely remove viscous stains from the surface.
  • Task decomposition: Push-Wiper decouples cleaning into a gathering phase and a post-processing phase, rather than directly wiping viscous stains that can worsen contamination.The gathering phase uses segmented pushing trajectories to aggregate the stain; post-processing then handles the aggregated material.
  • Task decomposition: The gathering objective minimizes the stain’s maximum spatial diameter D(St) and connectivity fragmentation K(St) across pushing trajectories τ.A weighting factor λ balances spatial spread against the number of disconnected components.
  • Task decomposition: The gathering phase progressively contracts the stain toward a compact state characterized by D < ϵ and K →1.Binary maps and robot states serve as observations during this data-driven gathering process.
  • Task decomposition: Classical planning is intractable because predicting contact-induced fluid deformation requires precise prior knowledge of hidden rheological parameters.This motivates using a data-driven approach for stain aggregation.

B. Gathering Phase

The Gathering Phase models stain aggregation from binary spatial maps and uses segmented Diffusion Policy trajectories to iteratively push stains inward. ASPI maps these planar actions onto surfaces for force-controlled execution, with scraping after each trajectory until the stain area falls below a threshold.

  • Action generation: Push-Wiper abstracts visual observations into texture-less binary stain maps, making the policy learn pushing strategies from stain spatial topology rather than visual appearance.The binary map uses stain = 0 and clean = 1.
  • Action generation: At each macro-step, Diffusion Policy receives the current stain map and capture pose and generates a complete segmented pushing trajectory rather than a short receding-horizon action.This iterative design accommodates stain redistribution that invalidates plans based on a single initial observation.
  • Action generation: The policy outputs low-dimensional planar actions a = (xb, yb, ∆θ), regularizing motions to push stain boundaries inward while decoupling topological planning from 3D surface geometry.ASPI independently reconstructs the end-effector’s 6-DoF pose from these planar actions.
  • Trajectory generation: ASPI projects each planar action onto the 3D surface, aligns the TCP z-axis with the negative surface normal, and superimposes the predicted yaw increment for stable contact.The interpolator converts the 2D trajectory into a surface-conforming 3D execution trajectory.
  • Trajectory execution: After each pushing trajectory, the manipulator performs predefined scraping to clean the sponge and repeats gathering until the stain area falls below qth.Execution maintains prescribed normal contact force through an admittance controller while other motion axes follow position control.

C. Post-processing Phase

The post-processing phase removes the residue gathered during cleaning through predefined motion primitives and completes the process with a final full-coverage wipe. It also self-cleans the sponge and regulates its moisture.

  • Residue removal: After stain gathering, the robot executes predefined motion primitives using the same hybrid force–position controller to remove the compacted residue.The primitives include dabbing, scraping, rinsing, and squeezing.
  • Sponge maintenance: Dabbing lifts the residue, while scraping, rinsing, and squeezing self-clean the sponge and regulate its moisture.
  • Cleaning completion: A final full-coverage wipe follows the post-processing primitives to complete the cleaning process.

IV. EXPERIMENTS

The experiments compare Push-Wiper with baselines on representative viscous stains, test generalization to unseen objects and curved surfaces, and evaluate post-processing effectiveness.

  • Comparative evaluation: The study compares Push-Wiper without post-processing against two baselines on two representative viscous stains.
  • Generalization: The experiments validate Push-Wiper’s generalization on unseen objects and arbitrary curved surfaces.
  • Post-processing: The study separately evaluates the effectiveness of the post-processing stage.

A. Experimental Setup

The experimental setup uses a force- and vision-equipped UR7e robot with a sponge end effector to evaluate cleaning on simple and complex viscous-stain tasks. Performance is measured by Cleaning Score (CS), the percentage reduction in stain pixels after cleaning, using ketchup and peanut butter as representative stains.

  • Platform: The platform combines a UR7e robot, KWR75B six-axis force/torque sensor, RealSense D435 wrist camera, and an 11 cm × 7 cm sponge block.The system uses hybrid force–position control and requires no specialized mechanisms for sponge mounting.
  • Tasks: Tasks are categorized as simple when stains form one connected region below 20% of the image, and complex when regions or coverage exceed those conditions.Complex examples include multiple scattered stain regions.
  • Metrics: The evaluation uses ketchup and peanut butter, whose reported viscosities are 1000 mPa · s to 1500 mPa · s and 29 452 mPa · s, respectively, at 10 s−1.These stains represent distinct physical properties under the same shear-rate condition.
  • Metrics: Cleaning Score (CS) is defined as the percentage reduction in stain pixels after cleaning, based on stain masks extracted before and after each trial.Nbefore and Nafter denote the total stain pixels before and after cleaning, respectively.

B. Baselines and Implementation

The study compares Push-Wiper with full-coverage wiping and one-shot pushing under a shared perception and execution stack. Demonstrations use segmented planar pushing, while implementation specifies diffusion-policy planning, hybrid force–position control, and consistent baseline settings.

  • Baselines: The baselines isolate aggregation-first cleaning from implementation differences by comparing full-coverage wiping and one-shot pushing within the same perception and execution stack.Full-Cover executes repeated fixed sweeping trajectories, while PushAll-Onetime predicts one global continuous pushing trajectory per replanning step.
  • Baselines: Full-Cover repeatedly executes a fixed sweeping trajectory spanning the entire cleaning area.The trajectory may be arched or boustrophedon-like.
  • Baselines: PushAll-Onetime replaces segmented aggregation with a single global continuous pushing trajectory intended to traverse the current stain region in one stroke.Its global sweep plans often produce long paths and large orientation changes.
  • Implementation: 150 planar cleaning tasks produce 448 segmented pushing trajectories, with each trajectory represented by three DoFs, a = (x_b, y_b, ∆θ).The tasks include 75 ketchup and 75 peanut-butter trials, each with 25 simple and 50 complex stain patterns.
  • Implementation: Push-Wiper uses a convolution-based diffusion policy with DDIM sampling, predicts n=16 actions per plan, and executes actions with a 100 Hz hybrid force–position controller.Training uses 100 diffusion steps and inference uses 10 steps; baselines share the force controller, PO uses identical diffusion settings, and all methods scrape the sponge once after each trajectory.

C. Cleaning results of three methods

Push-Wiper achieves the strongest and most stable cleaning performance on ketchup and peanut butter by progressively aggregating stains through segmented pushes. Baselines instead smear or spread viscous stains, while Push-Wiper operates under a matched execution budget with an average runtime of 130 s per trial.

  • Cleaning effectiveness: 89.88 overall average CS makes Push-Wiper outperform FC (32.64) and PO (44.98) across ketchup and peanut-butter experiments.Each method was evaluated in 20 ketchup and 20 peanut-butter trials covering distinct stain distributions, with 10 simple and 10 complex cases per stain type.
  • Baseline failure modes: FC smears viscous stains through sponge deformation, causing secondary contamination, substantial residues, and potentially negative CS.These failures are especially severe on peanut butter because force-controlled wiping redistributes material along the wiping path and has limited absorption.
  • Baseline failure modes: PO adapts trajectories to stain distributions but lacks explicit aggregation, so global wipe/push paths mainly traverse and spread stain mass.Push-Wiper instead uses step-wise pushes that progressively concentrate stains into compact regions for reliable cleanup under complex, scattered distributions.
  • Cleaning mechanism: Push-Wiper uses segmented, step-wise pushing to progressively concentrate stains into compact regions before cleanup.This segmented pushing process is illustrated in Fig. 5.
  • Runtime and evaluation budget: 130 s is Push-Wiper’s average wall-clock runtime per trial, including fixed sponge scraping after each pushing trajectory.Baseline evaluations use Push-Wiper’s completion time as the matched execution budget T, fitting as many full baseline trajectories as possible within T.

D. Generalization on unseen curved surfaces

Push-Wiper generalizes to unseen curved surfaces despite training only on planar data, achieving high cleaning scores on both convex and concave geometries.

  • Curved-surface generalization: 91.45 CS was achieved on the convex surface, while 93.42 CS was achieved on the concave surface.Each geometry was evaluated in 10 complex-stain trials: five ketchup and five peanut butter.
  • Curved-surface generalization: The curved-surface results were comparable to those obtained on planar surfaces.The evaluation used one convex and one concave geometry, although curved surfaces were absent from the training data.
  • Curved-surface generalization: Push-Wiper was evaluated on curved surfaces that were not included in its training data.The tested surfaces comprised convex and concave geometries with complex ketchup and peanut butter stains.

E. Generalization on unseen objects and stains · F. Evaluation of cleaning effectiveness of post-processing · V. CONCLUSIONS

Push-Wiper is evaluated for zero-shot generalization to unseen solids, liquids, viscous stains, and post-processing-enhanced cleaning. The framework aggregates viscous material through segmented pushing, adaptive Diffusion Policy actions, ASPI, and hybrid force–position control.

  • E. Generalization on unseen objects and stains: Push-Wiper evaluates zero-shot generalization on unseen solids, liquids, and viscous stains including black pepper sauce and oyster sauce.For black disks, five objects are randomly placed, and success is assessed by whether they are aggregated into one connected region.
  • E. Generalization on unseen objects and stains: Black-disk evaluation tests aggregation rather than removal because a sponge cannot remove solids from the surface.The stated success criterion is a final count of one connected region.
  • F. Evaluation of cleaning effectiveness of post-processing: Post-processing is added after gathering, which already achieves near-complete cleaning, and is evaluated in five ketchup and five peanut-butter trials.The experiments compare cleaning with and without post-processing.
  • F. Evaluation of cleaning effectiveness of post-processing: 100 for ketchup and 98.51 for peanut butter are the CS values achieved with post-processing.These results show that combining gathering with post-processing enables virtually complete cleaning.
  • V. CONCLUSIONS: Push-Wiper cleans viscous stains from arbitrary surfaces by aggregating material with segmented pushing trajectories to mitigate secondary contamination.Its stepwise policy is governed by Diffusion Policy and adapts to diverse stain spatial distributions.
  • V. CONCLUSIONS: Generated actions are executed through the Arbitrary Surface Pose Interpolator and a hybrid force–position controller.The conclusion presents these components as part of the Push-Wiper framework for arbitrary-surface cleaning.
Loading 2608.00730v1…