The Saros Z70 mostly works like any other robot vacuum-mop combo, sucking up debris and wiping up light grime. But the main attraction is the OmniGrip arm, which promises to grab and sort the type of clutter that robot cleaners, until this point, haven’t been able to tidy up on their own.
The arm spends most of its time tucked inside the robot, behind a hatch. It’s programmed to recognize socks, slippers, small towels, and tissues (that’s all, for now), extend the claw, pick up the items, and then drop them off in designated areas. Socks, towels, and tissues go into a cardboard bucket that comes with the robot, and slippers go to another floor-level spot of your choice.
Setup took me about 10 minutes. You need to start by sending the robot out on a training run to draw a map of your home’s floor plan, a process that’s quick and mostly automatic. Then you coach it, through the smartphone app, so that it knows where you’ve put the sock bin and where you want the slippers to go.
Depending on which settings you have enabled, either it’ll sort your things as it finds them or you can set it to finish cleaning first and then go on a separate sorting-only run. If it comes across other obstacles, it’ll just steer around them. It also has a remote-control mode that lets you peer through the bot’s camera, drive up to obstacles, and then tell the claw to do its thing.
The claw isn’t exactly a piece of heavy-duty precision machinery, but it doesn’t strike me as flimsy, either. When it pops out of its hatch, tiny motors whirring, sometimes I stop to marvel at how far robot vacuums have come. Other times it strikes me as cartoonish and over-the-top, and occasionally I shudder, as though I’m gazing upon a techno-dystopian torture device from an H.R. Giger painting. My dog hates it.
Down to business: The claw doesn’t collect stuff as reliably as advertised. I gave it a few chances to pick up a normal mess that my dog and my 7-year-old leave out on any given day, but it grabbed only about a quarter of the items.
If a sock was anywhere near a piece of furniture — as it often is when you live with a first-grader who reflexively tears off their footwear as soon as they plop down on the couch — the Z70 wouldn’t even attempt to pick it up. It’s probably a good thing that the robot is smart enough not to snap off its own appendage while driving under a coffee table.
But even when I carefully staged a bunch of socks and slippers that didn’t seem like they would cause the Z70 any obvious problems, it still worked only about half as often as I thought it should.
It didn’t accurately identify every piece of detritus, even from its approved-item list. One of my black socks blended in with my navy blue rug, and the Z70 couldn’t spot the sock. A Roborock representative told us that the vision system struggles to spot objects on dark carpets in general. The vision system later mistook that same sock for a dog turd, even though it was on a bare wood floor that time. It didn’t register my L.L.Bean slippers at all.
Then again, it did pick up a Birkenstock sandal, which I guess it had pegged as a slipper (it must be hard to train AI on those subtle differences). It’s a positive sign that the arm is strong enough to handle that much weight, and I was impressed when the bot set the sandal down exactly where I had coached it to. The other half of the Birkenstock pair wasn’t so lucky, though; it sat sole-side up on the floor, and the Z70 stared at it for a second and then left without attempting to lift it.
Sometimes, when the Z70 did identify a sock correctly, the arm just couldn’t quite grab it. Considering the slew of cameras and sensors built into the arm and the main chassis, you’d think it would find the right angle, but it’s not there yet.
The Z70 was also way too conservative around furniture and fixtures for my tastes. It wouldn’t even try to pick up socks that were within a foot of my TV stand or toilet, and it balked at grabbing a sock under my dining room table, where there’s ample height for the arm to do its thing.
Then there are other obstacles that it should have steered clear of but didn’t. It promptly tried to pick up a sock next to my dog’s wire playpen — and ended up hooking itself to one of the crossbars and not letting go until the front of the bot had risen a few inches off the floor.
Occasionally, it just looked lost, twitching in place with its claw hanging out, calling out “sorting items, sorting items” over and over for about a minute at inconsistent intervals, before retracting its arm and moving on.
I tested it for about three weeks. It remains to be seen whether the claw will hold up over time, what with all its moving parts and all the crevices where dust and grit can gather. My sense is that these robots are not easy to repair. When I’ve spoken to Roborock representatives about the durability and repairability of the company’s products in the past, they’ve told me that it doesn’t plan to offer major replacement parts because it expects customers to upgrade every three years or so anyway.
The Z70’s arm is very much a work in progress, and Roborock has a big challenge on its hands. Getting the arm tuned correctly, across all the image recognition, mechanical engineering, and choices about risk management and bot behavior, will be quite a feat of engineering. And the solution probably won’t come to this first-generation model: In an email, a representative from Roborock said that the company probably won’t be able to solve some of the pickup and image-recognition problems with software updates alone.
For now, it falls well short of being Roborock’s promised panacea for all your floor-level messes. You shouldn’t count on Roborock (or any company) to catch up with its promises through software updates over time. It might — but I’ve seen too many examples of gadget makers not delivering on their promised improvements to navigation systems and other features.
