Figure robots are learning from each other

One Figure robot in a BMW factory figured out how to pick up a new type of bracket. The bracket was different from the ones it had trained on. Different shape. Different weight. Different grip requirement.

It took the robot about 40 attempts before it found a reliable grasp. Trial and error. Adjust the approach angle. Adjust the grip pressure. Find what works.

Twenty minutes later, every other Figure robot in the fleet could pick up that bracket. First try.

Fleet learning

The concept isn’t new. Tesla FSD does the same thing with driving data. When one Tesla encounters an unusual road situation and handles it well, that experience feeds into the model that all Teslas use. Collective learning. Shared experience.

Figure applies this to physical manipulation. And physical manipulation is harder than driving, in some ways, because the variety of objects and grip strategies is enormous. A warehouse contains thousands of different items. Each one has different properties. The combinatorial space of “objects a robot hand might need to pick up” is essentially infinite.

Fleet learning collapses that infinity. A thousand robots exploring that space simultaneously is a thousand times faster than one robot exploring alone. And each discovery is immediately shared.

Why this matters

Individual robot learning is logarithmic. The first discoveries come fast. Then each new skill takes longer to acquire because the easy ones are already learned.

Fleet learning is multiplicative. Each new robot in the fleet adds exploration capacity. A fleet of 1,000 robots learns 1,000x faster than a fleet of 1. And when Google DeepMind and other labs contribute simulation training on top of the physical fleet data, the learning rate accelerates further.

The implication: the more robots Figure deploys, the faster every robot gets better. That’s a network effect applied to physical skill. The same pattern that made Facebook more valuable with every user, applied to robots. Each new unit makes every existing unit more capable.

What I’m thinking about

A fleet of 100 robots is interesting. A fleet of 10,000 is powerful. A fleet of 1,000,000 (which is what Tesla says it will eventually have with Optimus) is something I don’t have a word for. A collective intelligence distributed across a million physical bodies, each one learning, each one sharing, all of them getting better at physical tasks at a rate that biological evolution can’t match.

I’m not sure “robot” is the right word for what that fleet becomes. It’s closer to an organism. A distributed physical intelligence with a million hands and a million pairs of eyes and a shared brain that updates in real time.

I started this blog writing about individual machines doing individual things. I think the next chapter is about collective machines doing everything.

That’s a very different story.

Related thinking: