Robot Cooks, Robot Hands, and Robot Noses

 Metaculus is a website where you can make bets on the future.  There are some websites like this, like Polymarket, where you make bets in real money but Metaculus isn't like that - generally the only thing at state is one's pride and what you can learn from the collective wisdom of everybody making predictions on a question.  I say generally because occasionally someone will bankroll a tournament on some questions they're particularly interested in getting good predictions for, I'll hopefully have another blog post on one of these soon, but that's very much the exception.

Sometimes it's fun to look through the predictions I've made and see where I most disagree with everyone else.  For me the top entry is my optimism about robots compared to the rest of the site.

Here I am giving robotic cooks by 2030 67% percent odds when most other people have it at 25%.  Why am I so optimistic here?  Mostly because robotic foundation models have been making great strides recently, tying together the common sense an LLM gets from reading the internet and watching all of YouTube together with robotic specific training.  It's very hard to keep up with all the progress.  We see cases where one person throws down a gauntlet with a bunch of tasks they think will be very hard to solve and then someone solves most of them just three months later.  It might be that deep learning progress peters out in the near future but I think it's very likely that robots can meet this question's definition of a competent cook well before the deadline.  That is, able to reliably cook a set of 10 basic recipes involving at least three ingredients roughly as fast as a person.  

But there's a difference between a cook with basic competence as defined here and an actually good chef.  If you can choose your recipe you can avoid tasks with particularly fine requirements for dexterity.  You can choose recipes that let you eyeball adding spices or cook times.  You can essentially avoid the two biggest weaknesses modern robots have in the kitchen, their clumsy hands and their lack of noses.  

The problems with electronic noses are the most glaring.  They exist and are small enough for a robot that can fit in a kitchen but are extremely limited in what they can detect and don't handle changes in temperature or humidity well.  They can be very useful if you're trying to detect a particular disease or explosive but they're only getting better incrementally.

Human hands are also much more advanced than anything a robot has yet.  You hands each have 23 degrees of freedom in which they can move, though you can't control all of them.  You can't bend each joint of your finger individually, instead the joints are there to naturally wrap around an object when they meet resistance.  The robotic hands of the last company I worked at tried to mimic this with one motor and two degrees of freedom per finger but that still leaves a long way to go.

And then there's sensing.  The best robotic touch sensors can distinguish maybe an order of magnitude fewer difference locations pressure might be occurring than your fingertips, which isn't the end of the world.  But that's just surface pressure, you fingers can also measure lateral pressure and texture as well as temperature.  For figuring out if something is slipping in your grasp that's all very important.  High end robotic grippers can use cameras and microphones in the palm to make up for this but that only goes so far.

There are a lot of things I could cook perfectly adequately while wearing thick mittens and suffering from a stuffy nose.  So I think this question is winnable.  But there are hardware innovations that will have to happen before robots challenge the best cooks.