Hacker News Reader

Comment by Animats | Hacker News Reader

A reasonable guess.

As far as I can tell, the number of humanoid robots doing anything productive is zero. It's all demos.

This is far harder than self-driving. As a guy from Waymo once said in a talk, "the output is only two numbers" (speed and steering angle).

Also, there are at least 18 humanoid robots good enough to have a Youtube video. Tesla is not the leader.

Remember the "cobot" boom of about five years ago? Easy to train and use industrial robots safe around humans? Anybody?

I'm not saying this is impossible, but that it's too early for volume production. This will probably take as long as it took to get to real robotaxis.

View on HN

> Also, there are at least 18 humanoid robots good enough to have a Youtube video.

Agreed, thing is the robot hardware isn't the hard part anymore, the top ten robots are all sufficient to be transformative if they had good enough AI.

My bet is on Google/Gemini being the first to market from what I've seen so far.

Boston dynamics is a leader in getting robots to do useful niche work in well bounded environments, but that's yesterday's news.

> Boston dynamics is a leader in getting robots to do useful niche work in well bounded environments, but that's yesterday's news.

BD did most of their locomotion using classical dynamics and control theory until a few years ago. So did Honda, with Asimo. I did some of that in 1994.[1]

Early thinking revolved around landing on the "zero moment point". There's a landing point which, if hit, maintains speed and balance. To speed up, you aim for slightly beyond that point; to slow down, aim for a nearer point. That was Asimo. You could push that concept to the level of BD's "Big Dog", and later, their smaller dogs. Even pre-calculated flips were possible. But that approach gets you rather clunky motion.

The next step was to use some machine learning to tweak the control system parameters. That works, but you don't get overall coordination of all the joints. That only started to appear as machine learning systems became powerful enough to take on the whole problem at once.

Hard problem. Took over three decades to get decent humanoid control. Now everybody is doing it. You can be too early.

[1] https://www.youtube.com/watch?v=kc5n0iTw-NU

Reply on news.ycombinator.com