Science · robot vacuums

What the research says about robot vacuums

Robot vacuums are one of the most studied consumer robots in research labs. Here are ten papers, all free to read on arXiv, on how they map rooms, avoid socks and cables, plan a path that covers every spot, and what happens to your data. Each summary is ours, written from the paper's abstract.

Papers 10
Topics 4
Reviewed Oct 24, 2026
Source arXiv.org

Our own simplified drawing. No figures are taken from the papers.

Known so far

Every number links to the paper it comes from.

Tidying plate load: 50N10Held with its motor switched off, over a 100-cycle test.
Pushing trials: 140trials10On seven household objects with a modified robot vacuum.
Coverage papers reviewed: 125papers8In one recent survey of coverage path planning.
Sensors on one floor scrubber: 5sensors53D and 2D lidar, two depth cameras and a stereo camera.

Navigation & mapping (SLAM)

3 papers

How a robot builds a map of your home, works out where it is and gets itself unstuck.

[1]2019cs.ROPreprint

Loop Closure Detection in Closed Environments

Nils Rottmann, Ralf Bruder, Achim Schweikard & 1 other

What they foundEven a robot with no lidar or camera can tell when it is back somewhere it has been, using only its wheel odometry and wall-following, and turn that into an accurate map.

Why it mattersCheap robot vacuums and mowers often move randomly; this is a step toward planned, room-by-room cleaning on low-cost hardware.

What they studied

The authors fed the detected loop closures into standard pose-graph map building and tested it in simulation and on a real lawn mower.

arXiv:1908.04558Abstract (opens arXiv)PDF (opens arXiv)

[2]2023cs.CVPreprint

2D Floor Plan Segmentation Based on Down-sampling

Mohammadreza Sharif, Kiran Mohan & Sarath Suvarna

What they foundA simpler method splits a cluttered robot floor map into rooms by repeatedly down-sampling it while keeping its structure.

Why it mattersRoom segmentation is what makes 'clean the kitchen' possible, and this approach is much less complex to compute and implement.

What they studied

They tested it on maps made by a vacuum robot exploring unknown homes and on a standard benchmark dataset.

arXiv:2303.13798Abstract (opens arXiv)PDF (opens arXiv)

[3]2025cs.ROPreprint

Embodied Escaping: End-to-End Reinforcement Learning for Robot Navigation in Narrow Environment

Han Zheng, Jiale Zhang, Mingyang Jiang & 4 others

What they foundA reinforcement-learning policy got a robot vacuum out of narrow dead zones consistently, beating the planning and learning methods it was compared with on success rate and collisions.

Why it mattersGetting stuck under chairs and between furniture legs is one of the most common robot vacuum frustrations.

What they studied

The robot used a lidar, an inertial sensor and wheel encoders, and was tested in real-world scenes of varying difficulty.

arXiv:2503.03208Abstract (opens arXiv)PDF (opens arXiv)

Obstacle & object detection

2 papers

Telling a sock or a cable from crumbs, and spotting people and objects.

[4]2024cs.ROPreprint

VLM-Vac: Enhancing Smart Vacuums through VLM Knowledge Distillation and Language-Guided Experience Replay

Reihaneh Mirjalili, Michael Krawez, Florian Walter & 1 other

What they foundVLM-Vac sorts what its camera sees into 'avoid' or 'suck up' and beat standard vision-based methods, especially at spotting small objects, while using less energy.

Why it mattersIt points to vacuums that learn the clutter in your home over time instead of relying on a fixed list of obstacles.

What they studied

A large vision-language model labels objects at first, then gradually teaches a much smaller model; a language-guided replay method keeps it learning as the home changes.

arXiv:2409.14096Abstract (opens arXiv)PDF (opens arXiv)

[5]2020cs.ROPreprint

Robot Perception of Static and Dynamic Objects with an Autonomous Floor Scrubber

Zhi Yan, Simon Schreiberhuber, Georg Halmetschlager & 3 others

What they foundA professional floor-scrubbing robot for large public places combines 3D and 2D lidar with several depth and stereo cameras to track people and find dirt and objects.

Why it mattersThe team released a dataset recorded in several public places, including a supermarket, a warehouse and an airport, useful for anyone building commercial cleaning robots.

What they studied

The paper describes the robot's sensor layout (a 3D lidar, a 2D lidar, two RGB-D cameras and a stereo camera) and its perception software, released for ROS.

Source for: Sensors on one floor scrubber

arXiv:2002.10158Abstract (opens arXiv)PDF (opens arXiv)

Coverage path planning

3 papers

Planning a route that cleans every reachable spot without wasting time or battery.

[6]2022cs.ROPreprint

Simulating Coverage Path Planning with Roomba

Robert Chuchro

What they foundA deep reinforcement-learning approach can model the job of visiting every free spot in an unseen room, and was compared with a Roomba's built-in algorithm.

Why it mattersIt notes how little earlier work had applied learning to coverage, the core job of every robot vacuum.

What they studied

The study simulated a vacuum robot covering rooms it had never seen, using deep reinforcement learning.

arXiv:2210.04988Abstract (opens arXiv)PDF (opens arXiv)

[7]2022cs.ROPreprint

Path Planning of Cleaning Robot with Reinforcement Learning

Woohyeon Moon, Bumgeun Park, Sarvar Hussain Nengroo & 2 others

What they foundOne learned path planner worked across different room layouts without retraining, trained faster than plain PPO and outperformed random and zigzag patterns.

Why it mattersRandom and zigzag are the classic robot vacuum strategies, and efficient paths save battery and household electricity.

What they studied

The authors combined the PPO reinforcement-learning method with transfer learning, reward shaping and other training tricks.

arXiv:2208.08211Abstract (opens arXiv)PDF (opens arXiv)

[8]2026cs.ROPreprint

Coverage Path Planning: Classical Foundations, Recent Advances, and Future Directions

Zongyuan Shen, Shalabh Gupta, Shancheng Zhao & 6 others

What they foundA survey reviews 125 representative coverage-planning papers from roughly 2015 to 2026 and sorts them into single-robot, multi-robot, 3D, constrained, learning-based and visual coverage.

Why it mattersIt is a good map of the field if you want to know how cleaning robots decide where to go.

What they studied

The authors compared the strengths and limits of each family of methods, building on the classical methods that came before.

Source for: Coverage papers reviewed

arXiv:2607.10649Abstract (opens arXiv)PDF (opens arXiv)

Privacy & new skills

2 papers

What a connected robot vacuum can reveal about your home, and new jobs it could take on.

[9]2024cs.LGPreprint

Investigating the Privacy Risk of Using Robot Vacuum Cleaners in Smart Environments

Benjamin Ulsmaag, Jia-Chun Lin & Ming-Chang Lee

What they foundEven with encrypted app traffic, some of a robot vacuum's cleaning events can be identified from the timing and size of its network packets alone.

Why it mattersConnected cleaning robots can leak hints about what happens at home, even when the content is protected.

What they studied

The researchers watched the network traffic of a popular robot vacuum running in a real smart home.

arXiv:2407.18433Abstract (opens arXiv)PDF (opens arXiv)

[10]2026cs.ROPreprint

Enabling Tidying Tasks in Robotic Vacuum Cleaners via a Geometry-Induced Self-Locking Deployable Linkage

Eunseok Song, Jinwoong Kim, Sohee Lim & 5 others

What they foundA fold-out plate that locks through its own geometry let a commercial robot vacuum push clutter aside; it held a 50 N load with the motor off and survived a 100-cycle test.

Why it mattersIt shows a cheap path to vacuums that tidy as well as clean, without adding a robot arm.

What they studied

The mechanism was built into a commercial robot vacuum and tested in 140 pushing trials on seven household objects, against a commercial robot with an arm.

Source for: Tidying plate load · Pushing trials

arXiv:2610.20150Abstract (opens arXiv)PDF (opens arXiv)

New research

New papers on robot vacuums, checked weekly, summarized after a person reads it.

RSS feed

Week of Oct 12, 2026

Oct 14Enabling Tidying Tasks in Robotic Vacuum Cleaners via a Geometry-Induced Self-Locking Deployable LinkagePrivacy & nextarXiv:2610.20150Summarized

How we choose papers

We search arXiv for robot vacuums papers and keep the ones that add something new, are clearly written up and are relevant to a general reader. A person reads every abstract.

Our own words, our own drawings

Summaries are written by us from each paper’s abstract, not copied. Diagrams are our own simplified drawings, not figures from the papers.

Preprint vs. peer-reviewed

“Peer-reviewed” means arXiv lists a journal reference for the paper. Everything else is marked “Preprint”: shared openly, possibly before or without journal review.

Paper metadata courtesy of arXiv. Summaries written by us; read the original papers via the links. Paper metadata from arXiv.org, which does not endorse this site.