Cover image for I Am Mesmerized By Our New Robotic Vacuum

I Am Mesmerized By Our New Robotic Vacuum

deciduously profile image Ben Lovy ・9 min read

Disclaimer: this is more a rant than anything else. Maybe more #crazydudewontstoptalkingavoidthewatercooler but that's unwieldy as a hashtag. The #help part is in the TL;DR.

My girlfriend (somewhat impulsively) recently bought us a new robotic vacuum. We'd had the idea in the back of our heads for a while but I always assumed we were talking about at least a $500+ purchase having only heard of the big name brands like Roomba. Thanks to Lord Bezos, we got a Eufy RoboVac 30 for under $200 - selling at $270 from the manufacturer, still considerably cheaper than I had anticipated. It's a pretty well-reviewed "budget" model. I can't even bring myself to be that upset, this thing is really cool. Dammit, Amazon, it's so hard to quit you. Her only other purchase? God of War for PS4. I'm gonna keep her.

I swear I'm not a Eufy shill, I'd never even heard of this brand, I'm just geeking out over our new robot. I'm not necessarily specifically recommending this model over other similar-tier models.

We put it through its inaugural run this weekend and I am utterly fascinated by it. That little thing is pretty clever, more so than I had anticipated, and I've just pretty much wasted my morning following it around trying to get inside its software head. I know these things have been around a long time now, but I'd never really been around one long enough to see it do its thing, and I'm impressed at how thorough a job it's doing.

It seems to be a great example of emergent behavior from very simple parts, and I can't stop thinking about it.

I admit, I was initially a total elitist. Some cursory searching had led me to believe that the super fancy high end ones have this whole mapping feature that figures out our floor map and stores it to efficiently cover the area, using its knowledge of the layout to optimize the path whereas the cheaper ones bumble around with sensors to get the same job done but much less efficiently, and likely only to 100% coverage over multiple runs. Not so much a replacement for normal manual vacuuming but an extra help that's pretty good for day-to-day maintenance and touch-ups.

I hadn't really ever thought about the problem before, so when we opened it up I had initially thought the basic MO would be to just run in straight lines until its bumper hits something, then try a new angle. Rinse, repeat for X amount of time, floor will be eventually cleaned. Do it every day or multiple times a week, and over the course of a week your whole floor is clean. Our apartment is a little funky layout-wise, and with a decent amount of floor space but odd partitions and angles in places, so I figured without some sort of stored picture of the shape of the space there was no way it could reasonably cover the area in a single 100-minute shot-in-the-dark style run.

I owe the entire industry an apology for what's clearly been a gross underestimation. I know I should have known better, there's all these different companies making all these different models. It's the future now, and people have put some thought into this even (especially?) at this price point.

This thing schooled my ass on how to get the job done. I'm not even ashamed to say this is likely the cleanest these floors have been since we moved in a year ago. It's nuts that it got the job done just through an effective application of very simple movement algorithms. If it's not 100%, it's pretty damn close. I ended up following it around or watching it from my desk for the duration (which felt a little embarrassing, yes, but I couldn't stop looking) and don't believe a spot went untouched even if it got there in a pretty roundabout way. It even crawled up on an area mat without complaining and handled all raised thresholds between rooms like a champ.

It likely simply has better suction than our manual vacuum to boot, but that particular unit of measure (1500Pα) doesn't mean a lot to me. I couldn't tell you what a Pascal of suction feels like, and don't know what the other vacuum is rated. I do know it did a seriously good job considering it's so small.

While I don't have an uber-fancy model to compare to in the same space, this thing is leading me to believe the next level up probably wouldn't have even been worth it, at least for us. I suppose if we had a mansion and the battery life simply wasn't long enough for all of the floors the time saved by optimizing for our specific home might be useful. Watching it putter about, though, I actually almost believed it was mapping the space out.

Instead, the manual and some searching around suggests it simply cycles through a few modes, based on data from its sensors. It's all very vague, but I think that might mean it really is quite simple. It starts by shooting off in a random direction, as expected, and turning when it detects an obstacle. It also has an edge-tracking mode with methodically tracks an edge by continuously attempting to turn right at a low speed, and a spot cleaning mode that spirals out from a point. You can enable each mode specifically or just let it auto-cycle, and I see no reason not to do that.

I had also underestimated the number of sensors - it's not just a big front bumper like I thought. It's got an infrared sensor on the side and a sensor on the bottom for drops, as well as a magnetic sensor to detect strips you can lay out to prevent it from going in certain areas. The infrared sensor may be just for the remote, but I think it might be used for locating home base as well if it's not storing any locations.

What got me to keep watching is how this cycling of simple algorithms managed to still cover all previously uncovered territory. As I was watching it I kept thinking "oh no, it doesn't know that it just missed a whole corner over here", watching it scoot off at some random diagonal. I actually had assumed it would be a little more methodical, tracing a zigzag across a space, but no, this thing was all random unrelated directions, which kept it interesting enough (apparently). Inevitably ten or fifteen minutes later some other seemingly random tack across the house would bring it exactly where it needed to go. It triggered edge mode at least once on every wall of every room and every weird nook and corner, got under all chairs and the kitchen table, found the bathroom a few times and got wherever it could fit, just really did a smashing job of it all. We never had to intervene.

These new vectors its choosing can't be random. It works too well. They sure seem like they must be informed by some sort of previous understanding of found edges, but one guess I had is that it actually just tries a range of vectors over time to ensure it hits a wide range. My gut feeling is that it's actually tuned even better than that but I couldn't tell from watching the first run, and I'd be curious to learn more about that process. I also don't know if that behavior is specific to this model, and others do take a more zig-zag approach. Ideally I'd try it out, but I'm not ready to bring a second vacuuming robot into our lives. I'm not that excessive, I swear.

As far as I can tell it's not storing where specifically it finds edges, and I did notice it repeat some work - for example, it methodically scooted along the same wall of the living room three times over the course of it's 100-minute run. If it were storing specific edges, it wouldn't have done so, so it must have just sensed an edge and kicked in to that mode.

It must just be calibrated properly to avoid the problem. For instance, it must know how often to kick in to edge mode to maximize for exploration and minimize repeated work without sacrificing thoroughness. 100 minutes is a long time to vacuum, but not that long if you waste 20 of it on one wall over and over again. Most of the time it just turns away, though, and sometimes it even anticipates a wall and turns before it hits it well before (I think) any of its sensors would have noticed it. I can't figure that out either if it's not keeping track of anything. It still manages to activate for all the different walls, though!

Due to a connecting door we keep closed for airflow purposes, in the robot's universe our unit is kind of U shaped. We have furniture in odd places and some large open areas and some small, more fractured non-rectangular areas. All six of the rooms we want it to cover are different shapes and sizes, and all have more than four walls due to nooks and crannies and closets of varying sizes. Most of the doorways are your standard door size, but the living room and dining room are connected by a big opening almost as wide as the wall but not quite. This creates even more weird edges and corners to confuse it. Optimizing a filing path in a given time constraint by hand even with prior knowledge would probably not be simple.

Somehow, amazingly, it just kept passing through rooms at different enough angles to cover the whole floor. Sometimes it'd be stuck in a room for a while, sometimes it would get on a tack that has it passing between two rooms regularly, back and forth in a big line, but by the time it hit low power mode and returned itself to the charger it had actually hit all the spots I had made mental notes to look out for.

It even surprised me once again when I got worried it ran out of battery too far from home. I heard it kick into low power mode, and it was in a completely different room facing the wrong way. It did know how to orient itself towards home and try to get there, but on the way there it hit the ottoman (so, it didn't anticipate that was coming) and turned around the wrong way again. It even still with no line of sight managed to get itself oriented towards the station, bump its way around the ottoman to a straight shot, and amble its way back onto the contact pins. Even better, this straight shot included a bit of edge mode - which I was pleased to see still enabled even in low power mode - along the front edge of the couch. I'd been taking note of unexplored edges and that was the only one remaining it hadn't done!

Okay. Enough gushing. Here's the point.


I'm fascinated by how effective this budget-end robot vacuum's seemingly simple set of algorithms is at cleaning our whole weird space. I want to try to model the problem and implement the movement algorithms myself to see if I can replicate that emergent behavior with totally random inputs or if it requires more tuning, and what sort of tuning if so. It's not a type of programming I have much experience with. What would you use to explore this sort of thing?

I have some general space-filling/path-finding algorithms in my toolkit to start from, but I also know that some environments are easier than others for this sort of modeling. Some options I've heard of but don't know anything about:

  • Processing - Java (I don't really know Java, but am studying C++? Similar-ish?)
  • The python turtle - Python. I don't know anything about turtle and very little about Python but it sounds like it can be used to explore this problem space. Python seems like a good choice for this but I don't know the ecosystem at all.
  • I guess the HTML canvas element but that sounds unwieldy and complicated. Is there a framework you recommend?
  • I know Rust has some geospatial crates and image/graph crates for visualization, but nothing that I know of integrated for ease of use, experimentation and one-offs.
  • Game engines? I've never used any, is that the right genre of tool for this? Love2d for LUA looks like it might be a good choice but it also might be overkill.

...that's it, really. Is there anything else I should be aware of?

Also, if you know more about these robot vacuums than I do and can enlighten me/us, please do!

Photo by Rock'n Roll Monkey on Unsplash

Posted on Jul 21 '19 by:


markdown guide

Considering you mentioned Processing and HTML Canvas, definitely do check out p5.js. It’s what I would use to model this problem.
Very interesting read too — I’d love to get one as a pet. If it’s a pet, you’re supposed to play with it and feed it and all that, right?


Oh cool, thanks for the tip! P5 looks neat, much easier than what I had in mind doing it manually. I've been focusing more on JS than Python lately, too, so it makes sense to not hop to something brand new.

I hope so. I'll start leaving dust bunnies around for it to find :p

EDIT: I think I'm gonna go with p5.js, it's exactly what I was looking for in a language I already sorta know! Thanks again.


We got one a few months ago. Bissell had it on sale for 200. I will say our experience has been different, and actually even better. We have 4 dogs. Two of which shed buckets of hair. Yet our vacuum has done a fantastic job at keeping the whole floor clean. With the exception of our half bathroom, but to be honest, the bathroom is super duper small relative to the robot. I will say however, there's more maintenance involved then I expected (naively). Because of the dog hair we end up cleaning it just about every day, and we've had to actually pull the sensors out and clean it a couple times (it's worth mentioning our dogs were not particularly friendly the first two days and did in fact step on it a couple of times...). But it's definitely less work for us then sweeping every other day (when I said "shed buckets" I meant it).

I did the same thing and sat there trying to figure out why it wasn't using a simple path. We have a circular swivel chair, and when I saw how it went around the chair my mind was officially blown. It has brushes that are larger than the robot. But what it did was rotate back and forth around the edge to pull dirt and hair out from under the chair, and then went back and sucked it all up.

And I realized after we bought this one, that iRobot (aka Roomba) actually sells a 200 version that you can program and supposedly do cool stuff with!


Great report! Good to know it's handling the dog hair - there's more likely than not a dog in our future, and it'd be great t be able to continue to use this device.

I was also pleasantly surprised by the brushes. I wasn't expecting them to protrude past the body of the machine but you're right, it's very effective at kicking out dust from under cabinets and along hard-to-reach areas, then passing back over to pick it all up. I'm unclear on whether it's doing so intentionally or just happens to then find itself passing back over the dust it kicked up, though - either way, it works well!

Good note about cleaning the sensors, thanks. We'll probably keep dumping the dust trap after each run for now even if it's overkill, it only takes a few seconds.


Awesome, thank you! I think Python is the language I'd prefer, I'm less convinced about about Turtle specifically, definitely going to check these out.


Nice algorithmic puzzle you've got going! I'm definitely curious now.

Shameless self-promotion, but my "Dead Simple Python" series should help you transition from whatever language you're using (C++?) into Python pretty seamlessly. It's written for that purpose, basically.

Awesome, definitely will check it out! I have used python a little bit here and there but never for anything larger than a hundred lines or so. I basically just use it like C that doesn't bite, I think. I've got a lot to learn about its idioms.


I'm SO interested in one of these... but have no clue how it would behave in a home with children, where there are regularly things like toys, tiny Legos, & soft edges like dropped blankets on the floor.


It might be tricky, we did have to do a pass an pick up cords and things, and it does treat blankets as walls which can un-optimize it pretty quickly. The magnetic boundary tape it comes with can be useful for cordoning off a part of a room, though, of the legos are relatively localized! This one you can also tell to specifically just do a room, which may be easier one at a time than needing to keep the whole house ready to be crawled over.


I too am interested, but I have approximately enough open floor space to fit a pair of shoes so it's not entirely practical.

I shall educate myself vicariously through you lot.


I was also mesmerised by my Roomba a few years ago. But recently got a neato, and i have to say using a map to specify no go zones is much more effective and less hassle than other methods. Also, if a room needs attention I can do just that one. After getting mapping, I wouldn't want to go back to not having it.


That makes sense - my rose colored glasses may be wearing off, but the second run I noticed a lot more repeated work. It's definitely not well optimized, I think our apartment is just small enough to get covered well anyway. We'll eventually move and likely revisit then!


I got one, too.
Three key take aways after a few days of robotic cleaning:

  • It will not replace manual vacuuming. It's great for general cleanup, but I'd I wouldn't trust it to have gotten every spot before the in-laws come visit.
  • I have a year-old son, who is always exploring the flat and testing out when his parents tell him to leave something be because it's dangerous or fragile. The robovac told me a lot about cables and wires we had still lying around.
  • Miss-scheduled cleaning cycles at night are scary as hell.

That's interesting because most Roomba vacs have similar algorithms I think but are more expensive. Only higher end roomba will actually remember a map and as I was told, the top end model coming over 1k take a couple of hours mapping.

Check out xiaomi's roborock. We got one recently and did the same thing of following it around. A bit more expensive but uses laser distance sensor to map out a room and continuously map out a room during the cleaning session and can store one map currently. For about an 800sqft space, it cleaned in about 60min time using a third of the battery. Difference for us is I'm working on some major renovations and it's a pretty good test of what it can do. We're pretty impressed so far. One major flaw is when it ends up inside drapes and decides for a while that it has nowhere to go based on laser distance.


That's interesting, but I guess it makes sense that most would end up on similar solutions. I'm just still surprised it isn't doing a space-filling zig-zag, but maybe it just can't even bother with absolutely no state stored. It's crazy how steep those get.

That's the second shout-out to that brand, and that sounds pretty clever. Definitely more efficient than whatever this is or isn't doing. I don't think there's another in our future any time soon, but this brand will be my first go-to if/when it does happen.


The power of random with simple heuristics indeed.
You could check out our PyBullet, it is a Python module suitable for this task. It has optional OpenAiGym environments, and install is one line: pip3 install pybullet --upgrade --user
Check out pybullet.org for a quickstart.


This is awesome! Thanks a lot.


I have some general space-filling/path-finding algorithms in my toolkit to start from

What you want is to study search algorithms. Norvig et al's "Artificial Intelligence" is really good here (though it is a dense, weedy subject). A lot of AI work is about cleverly pruning search space. You can do surprisingly well at a lot of things with very basic ideas (just generally speaking) and it's not until the scale gets large that it makes sense to focus on anything more efficient


A guy I worked with previously worked on guidance systems, he informed me that they use SLAM algorithms - Simultaneous Localization and Mapping. en.wikipedia.org/wiki/Simultaneous...


Aha! This is exactly what I was looking for, thanks very much for the link.


We've got one of these vacs as well - super cool, isn't it! I'd love it to be a bit more intelligent in terms of mapping the space, but it definitely gets the job done.

Now, if only if it could hover up the stairs... ;-)


I have a feeling as I watch it over time the "dumb"ness of it will become more and more apparent...I have a feeling its only a matter of time until they crack that problem too!


Unity3D all the way. For simulation it just does so many things right out of the box and in one integrated package its just a no brainer for me. C# is also not terrible and is easy to get started with if you have any exposure to other C family languages.


What a fun article to read, thanks!

I once owned one of these, some Chinese model rebranded and sold in Brazil as 'Ropo'. It was dumb and annoying as. Sometimes it would go about 1/3 of the apartment and come back to base thinking 'job done' (battery still full). Often got stuck in the kitchen and started beeping for help. It was also one of cheapest options on the market, double the price of yours but 5 years ago, so it's good to know they evolved.


Oof, that sounds frustrating. Looking around it seems like there's definitely still a lot of variance between these knockoff brands, I think we just got pretty lucky.


Not sure how I missed this when it was fresh...

Robot vacuums and their kin were languishing in a corner of my mind reserved for things that seem cool but aren't ready yet. Mostly as a result of a co-worker that was an early adopter and unimpressed.... 10+ years ago.

Having just moved to a new apartment and with my pets are en-route, it's time to re-evaluate.


That's how it went for me too - my opinion was formed somewhere around generation one and never revisited.

Now that the novelty has worn off it's definitely not a life changing device, but it's quite nice to have.


I suspect, but don't know, that one thing that makes it work much better than expected is sensing if it's picking up any dirt/dust. You can avoid following the same edge if you notice you're not picking up any dirt. You can also go off into a random direction when you're not picking up any dirt. If you're methodical when detecting dirt, and random when you're not, I bet you do ok.


Do you have Telegram messenger?
There's a group I should probably add you to, it's people jailbreaking their Xiaomi vacuums.
They are OBSESSED with maps.


I do not, I'll check it out after work today! Sounds like a fun group, hadn't thought to try anything like that on this device but its the next inevitable step!


Wow, I thought these robot vacuum cleaners can't clean carpets. Now I feel like a caveman (and like I have to get one ASAP).


I know! All my assumptions about these things were created in like 2005.


God of war PS4! Yup defo a keeper!


I should be going to bed but just spent 5 more mins skimming this and 5 more mins reading comments.


Please do a collab with computerfile on youtube if you find anything interesting.


This sounds awesome! Such a fun project!

I don't have much domain experience in anything like this either, but Python definitely feels like a good choice for this kind of thing!


I agree, seems like the right style for this task! A good opportunity to learn more about that ecosystem too.


😂😂😂😂 I could see myself doing this! We are such nerds!!🤓🤓🤓


Dude, I got a Roomba knockoff and watched it for days. I'm with you. It's so fascinating.


Does your also seem to take totally random directions each turn, or is there a clearer pattern?


Robotic vacuums changed my take on cleaning the house. Frankly speaking, I thought they are useless, low-power, and such. But I was recommended to read an article here cleanup.expert/robot-vacuums/#rating and buy one of these models. I didn't expect this to be that useful because I almost forgot about dust (well, its quantity reduced significantly) and pet fur on the floor.