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/.
Alex: Welcome to another episode of ResearchPod. Today, we're looking at a robotic system called Push-Wiper that aims to solve a surprisingly difficult problem: cleaning up sticky, thick messes.
Sam: That's right. When you spill something like peanut butter or syrup on a counter, a standard robot usually just smears it around, making the mess larger. This paper introduces a system that treats cleaning as a geometry puzzle rather than a wiping task.
Alex: So the paper argues that current robots are using the wrong strategy for sticky spills?
Sam: Exactly. Thick, sticky substances behave differently than liquids or dust. If you try to wipe them directly, you're fighting the physical properties of the material itself, which causes it to spread further.
Alex: So if wiping just makes it worse, what does this system do differently?
Sam: It uses what the researchers call an "aggregate-then-finish" approach. Instead of trying to remove the stain in one go, the robot uses a sponge to push the mess inward from the edges, gathering it into a small, compact pile. Once the stain is concentrated, the robot removes the entire pile at once.
Alex: That's like using a squeegee to gather a spill into a puddle before picking it up, rather than trying to wipe the whole surface clean. How does the robot learn to do that?
Sam: It looks at the stain as a simple, texture-less map. By ignoring the specific color or appearance of the stain and focusing only on its shape, the robot learns to push from the outside in. The researchers call this a "Diffusion Policy" — think of it as the robot's brain predicting the best sequence of pushes based purely on the shape it sees.
Alex: So the Diffusion Policy is the decision-maker — it figures out where to push. But then something else has to actually move the arm?
Sam: Right. And moving the arm precisely on different surfaces is trickier than it sounds. That's where something called an "Arbitrary Surface Pose Interpolator," or ASPI, comes in. It's essentially a translator. It takes the brain's 2D instruction — "push this way" — and converts it into the precise 3D movements the robot's arm, elbow, and wrist need to make. It ensures the sponge maintains the right angle and pressure against the surface, whether the table is flat or slightly curved.
Alex: So it's separating the "what" from the "how." The brain decides the strategy, and the ASPI handles the physical geometry.
Sam: Precisely. And because those two tasks are separate, the system doesn't have to relearn everything when it moves to a new surface or encounters a different type of stain. It just applies the same geometric logic.
Alex: How well does it actually work?
Sam: The study suggests it's a notable improvement. Using this gather-first strategy, the system achieved a cleaning score — essentially the percentage of stain removed — that was more than twice as high as standard baseline methods.
Alex: That's a significant gap. Does it need to be trained on every type of mess, or can it handle things it hasn't seen before?
Sam: That's one of the key findings. Because the system learns the basic geometric shape of a stain rather than the specific properties of any one substance, it can handle new, unseen materials without additional training. Researchers call this "zero-shot generalization" — it works on the first try with things it was never explicitly taught.
Alex: So it's not memorizing specific spills. It's learning a general rule for gathering any thick material.
Sam: Exactly. And there's another layer to how it handles the physical act of pushing. The robot needs to press the sponge down firmly enough to move the material, but not so hard that it just slides over the top of it. To manage that balance, the system uses what's called an "admittance controller." Think of it like a car's cruise control, but for pressure — if a sensor detects the sponge has lost contact with the surface, the system automatically adjusts the arm's position to restore steady contact.
Alex: So it's constantly feeling the surface and correcting itself in real time. That matters because you can't really predict in advance exactly how a particular syrup is going to behave.
Sam: Exactly. Rather than trying to calculate the fluid dynamics of every possible spill — which would be enormously complex — the system just responds to what it physically feels. It's a practical workaround for a genuinely difficult problem.
Alex: Are there any limitations the authors flag?
Sam: Yes, and they're upfront about them. The system relies on a separate post-processing step — like scraping or dabbing — to actually remove the pile once it's been gathered. The Push-Wiper handles the aggregation, but not the final pickup. It also assumes the robot has clear, unobstructed access to the surface, so it doesn't yet account for navigating around objects in a cluttered environment.
Alex: So it works well on a clear counter, but it doesn't solve the broader challenge of operating in a busy kitchen.
Sam: That's a fair summary. The authors suggest future work will focus on adding collision awareness for safer operation around people and objects. For now, the paper represents a meaningful step in moving robotic cleaning beyond simple sweeping — toward handling the kind of complex, sticky materials you'd actually encounter at home.
Alex: Rethinking the geometry of the task rather than the physics of the material — it's a practical approach to a genuinely messy problem. Thanks for listening to ResearchPod.