Technical thoughts on the state of robotics
L'Architecte by Henri Marcel Magne, 1910. First saw this in Michelle Fang's 2025 letter.
Spent a week diving into the Shenzhen robotics scene, the highest SNR place on earth for robotics.
- Every one want to build robots, everyone wants to deploy them; but no one has a bulletproof use case for them (imo, prove me wrong!)
- The big names in robot production are worried about developing a better product/reducing costs. But don't put a thought into how they are selling them.
Leaving open a market for car dealership style businesses. - People are still sleeping on hardware (as always); it has become my opinion that useful humanoids can't be created focusing on the software layer.
Future unlocks will happen 75% on hardware, 25% on software. - One of the main bottlenecks right now in mass scale deployment is the learning rate. Taking 100-200 examples is not worth it for replacing a human. You need to train a robot in 5 examples, for that, we need two things:
- A robust learning mechanism, fine tuning by back propagation won't work.
- A general model that can get out of sticky situations.
Those are the two problems to solve. The first is (on the way to be) solved, the second is the problem.