Accepted to IROS 2026

Push-Wiper

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

Renhao Lu^1,2,† Mingxin Wang^1,† Chenyang Cao³ Yang Yang¹ Guoping Pan^1,2 Kangkang Dong¹ Yi Cheng² Houde Liu^1,2,*

¹Tsinghua University ²Z-Lab, Zerith Robotics ³University of Toronto ^†Equal contribution ^*Corresponding author

arXiv Videos BibTeX

Overview of the Push-Wiper framework — From a single observation, Push-Wiper plans a segmented push, adapts it to the surface, and iterates until the stain is compact enough for post-processing.

89.88overall cleaning score

+130%up to baseline improvement

99.25with post-processing

93.42concave-surface score

Overview

Aggregate first. Finish clean.

Viscous stains resist conventional robotic cleaning: wiping spreads them, while aggressive scrubbing can damage the surface. Push-Wiper instead treats cleaning as a topological aggregation problem.

A low-frequency Diffusion Policy predicts complete segmented pushing strokes from a binary stain map. An Arbitrary Surface Pose Interpolator and hybrid force-position controller then convert each planar action into a smooth, contact-aware 6D trajectory.

Iterative perception-planning-execution progressively contracts the stain into one compact region. Five post-processing primitives detach the residue and self-clean the sponge, completing the task without retraining for new stains or surface geometries.

Video

Cleaning is a sequence of deliberate pushes.

IROS 2026 video overview: stain gathering, surface-aware execution, and post-processing.

Baseline comparisonFull-Cover, Push-Wiper, and PushAll-Onetime on peanut butter

Varied stain distributionsKetchup trials across diverse spatial layouts

Post-processingFinal stain removal after segmented gathering

Unseen curved surfacesCleaning on concave and convex geometries

Unseen objects and stainsLiquids, solid objects, and viscous stain mixtures

Method

Decoupling 2D aggregation from 3D surface geometry.

Perceive

Abstract the current stain into a texture-free binary map at a fixed capture pose.

Plan

Generate a complete 16-action pushing stroke in the compact space of x, y, and yaw.

Execute

Use ASPI and 100 Hz hybrid force-position control to follow arbitrary surface geometry.

Finish

Detach the gathered residue, clean the sponge, and perform the final wipe.

Examples of segmented pushing trajectories — Segmented pushing contracts multiple disconnected stain regions into one compact target.

Five Push-Wiper post-processing motion primitives — Rinsing, squeezing, dabbing, scraping, and final wiping complete the cleaning cycle.

Results

High cleaning scores without secondary contamination.

Cleaning score across ketchup and peanut butter trials
Method	Ketchup	Peanut butter	Overall
Full-Cover	53.99	11.28	32.64
PushAll-Onetime	58.04	31.99	44.98
Push-Wiper	92.30	87.46	89.88

Mean cleaning score over 20 trials per method and stain type.

Radar comparison of cleaning methods — Push-Wiper remains stable across simple and complex distributions of both evaluated viscous stains.

Comparison of cleaning results across three methods — Full-Cover and PushAll-Onetime spread viscous material; Push-Wiper consolidates it.

Zero-Shot Generalization

New stains, new physical states, and unseen curved surfaces.

Training uses planar viscous-stain tasks. At test time, the same policy transfers to convex and concave geometries, solids, liquids, and novel viscous materials without additional data or retraining.

Push-Wiper operating on unseen convex and concave surfaces — ASPI keeps the end-effector normal aligned with unseen convex and concave surfaces, achieving scores of 91.45 and 93.42.

Cleaning novel viscous stains — **94.42** on combined unseen viscous stains.

Aggregation of unseen liquid and solid objects — **100.00** on solids and **92.62** on liquids.

Final Step

Post-processing lifts the average score from 89.44 to 99.25.

Ketchup: 100.00 +6.66%
Peanut butter: 98.51 +15.73%
Average: 99.25 +10.97%

Citation

BibTeX

@article{lu2026pushwiper,
  title     = {Push-Wiper: Toward General-Purpose Robotic Cleaning across Varied Stains and Surfaces with Segmented Pushing Trajectories},
  author    = {Lu, Renhao and Wang, Mingxin and Cao, Chenyang and Yang, Yang and Pan, Guoping and Dong, Kangkang and Cheng, Yi and Liu, Houde},
  journal   = {arXiv preprint arXiv:2608.00730},
  year      = {2026},
  url       = {https://arxiv.org/abs/2608.00730}
}