Here’s a really practical way to find out how AI is evolving.
You could start with the original iRobot Roomba vacuum cleaner, which came out in 2002. It went on a search and destroy mission for dirt, often rolling around in a sporadic pattern but eventually cleaning your carpet as you sip on coffee and watch. My own writing career happens to coincide with this debut (my career started in full force in early 2002), along with the first Apple iPod.
I remember being fascinated by the Roomba at the time, so much so that I visited iRobot in person way back when the company first started. I knew something was different about the company because, even back then, the Roomba became a major hit with consumers. (The company expects to reach around $650 million in sales this year.) You could vacuum when you weren’t home! You could leave the house and come back to clean carpets!
Still, when that unnumbered model shipped, the AI was not that advanced. In fact, you might say it didn’t use AI at all, as Ken Bazydola, the director of product management at iRobot, mentioned to me recently. At least, the machine learning was in an early state. He said it was intelligent, but the decision making back then was fairly basic. It might bump into a wall, but it would try different tactics to recover and keep cleaning, maybe using a different angle or route the second time.
Fast forward to the latest Roomba, the 980 model. It uses true AI. Now, when you let the bot loose on your living room, it will scan the room size, identify obstacles, and remember how to clean the carpet and which routes work best. Interestingly, the bot resets itself after each cleaning session and starts over, in case you move the furniture around or if you leave a pile of clothes sitting out in the open. In my tests with the 980 recently, I noticed how it would “search and destroy” a lot less. When the bot encountered a table or a sofa during one session, went back to recharge, then ventured off again, it seemed to know where those objects were in the room.
The bot can also determine how much cleaning it should do based on the room size. In a small area, it will repeat a cleaning cycle three times. In a medium-sized room, it will clean twice. In my larger living room, it cleaned once. This is more like the machine learning you might encounter in a Tesla Model S (which also scans for obstructions) in that it uses a sensor for odometry (where the wheels travel), what it has encountered (a sensor for the floor), and a camera that can scan the room.
That’s a big improvement compared to the original because it shows how this decision tree works. The bot keeps a record of all obstacles. Like the guy in the Memento movie, it has short-term memory loss, but it is smart enough to vacuum more proficiently when it does record objects for each session.
In terms of the AI, Bazydola agreed with me that there is an obvious next step. The AI will eventually monitor and remember objects like a sofa and identity that it even is a sofa, maybe based on the texture or size.
Today, the Roomba knows whether it is vacuuming on a plush rug or sucking up dust on a hardwood floor, based on that floor sensor. The 980 has a boost mode that it can use when it detects heavier carpet. Someday, it will identify new objects in the room but remember that a coffee table is always in the same place (and avoid that one spot where it won’t fit in between the legs). The AI will keep a massive decision tree for this, constantly labeling obstructions as temporary or permanent, so large that they cause problems or small enough that the vacuum knows it can sneak under the obstruction and keep cleaning.
What will it take? That’s not something iRobot will reveal, but my guess is that it is partly an issue of processing power and partly determining if there is even consumer demand for that kind of long-term memory. As it stands, I know the Roomba did a pretty good job on my living room.