From the long silence here, you might think I'd stopped think and writing about robot matters. Not so!
In fact, one reason I've been quiet here is that I've been working on a large-ish robot-related journalism project, which I hope will see the light of day very soon. (Its timing is up to its editors, not to me.)
And I've been working on some other projects, because much as I love this newsletter, it does not yet pay anything.
These other projects take time. A problem for me because I'm not temperamentally a quick-hits kind of guy. I prefer to offer longish posts, but I can't produce those every week.
I am not happy with these long gaps in publication.
So from now on, I will be posting once a week.
Every few weeks, those posts will be essays of the kind I've described. In the weeks between, I'll be posting links to robot news, research papers, and any other sources that I think might be worth a look. My goal with these links is to share what I'm learning as I learn it. I suspect that at least one or two links each week will be news to you, and news worth knowing.
So here is this week's Five Bites of Robot News. See you next week.
1. A plan to create a shared dictionary of robot actions.
Once you get past the carefully curated demo videos on YouTube, most robots strike regular people as pathetically limited in the things they can do. The slightest variation in its environment, the most meager surprise, and klompf!, the robot isn't achieving human-like behavior any more.
Now that Large Language Models like ChatGPT sound convincingly human at least some of the time, the things a robot can do are much, much less than the things it can say.
The problem is that there is no standard, shared database of robot actions. To teach your robot how to "grasp firmly," you invent your own solution. It may work on a robot very like yours, but it won't work on every robot with a grasper.
The contrast with LLMs is stark: When you ask the bodiless AI to talk to you about carrying a plate to the table, it runs its algorithms on a huge trove of words whose meanings are shared. "Carry" is roughly the same for ChatGPT as it is for Bard or Bing. But "carry" for Robot A is a set of instructions that's different from the ones in Robot B.
That's why this announcement could be huge: Google DeepMind is partnering with 33 other research teams to build a common database of skills and tasks for 22 different robot types. Such a database could bring robots much closer to humanlike flexibility, because it will make it possible for many different kinds of robot to "know" skills and tasks without having to learn them in advance, by themselves.
2. How much autonomy does a combat robot need?
As the U.S. Air Force gets serious about a "robot wingman" for fighter pilots, the question is becoming both urgent and practical.
3. Has the sacred cow of "shareholder value" hobbled American robotics and AI?
4. What are AI predictions good for?
Machine learning can make great predictions about outcomes -- will this student pass the test or will this patient get the disease you're worried about? But predicting a future outcome is not the same thing as improving it. With a prediction that a kid will flunk a test, do you get her a tutor? Or make sure she receives financial assistance? Or put food in her fridge? This paper argues that we shouldn't get carried away about AI's capacity to make predictions -- because, after all, prediction for prediction's sake is beside the point. Can AI actually help improve the future it predicts?
5. Will the Robot That Delivers Your Pad Thai Spy On You?
"Burrito bots" like Starship and Serve (which does Uber Eats deliveries in Los Angeles) are impressive feats of real-world robotics. Some people find them charming; others like to fuck with them, because maschinenschadenfreude is not rare. But not enough has been said about the surveillance possibilities of all those sensor-laden units as they criss-cross the streets. Recently the LAPD subpoenaed Serve for video footage from one of its robots -- because some mooks had tried to steal it, and the police were investigating. It's more than reasonable to wonder what they might subpoena next.