Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- Amazing job guys. You've just created a base for robotic auto-evolution.
Now we need a model with a sole purpose of improving itself (procreation) and it's mobile platform. For the later this framework can be used. We need another model which will then do full CAD spec for each iteration and small automated factory which will implement this all in hardware ...
- I think this kind of automated design process can un-hype humanoids. The point of humanoid robots is one-size-fits-all -- a generalist can readily fit into various scenarios. However, if custom designs are cheap and optimized, there's no reason to stick to inefficient generalist designs.by esjeon
- Let's be honest: people want humanoids because half the world was used to "domestic servants" (ie. slaves) (and even is, middle east, Pakistan, India, large parts of Africa) and they can't have that because colonialism "took that from them".
What people keep forgetting is that colonialism, and with it the large-scale disappearance of slavery, and so the disappearance of domestic servants/slaves, lasted easily into the 1990s. You don't have to be very old to remember having that (or you cheat, ie. rich middle eastern countries, and still have that in 2026, just called slightly different)
by spwa4 - Watching the video, 11:37 [1] made me feel very uneasy about the validity of those results. CMA-ES optimization times in seconds? That only works if you do no physics, no collision checks, no trajectory roll-outs. (Otherwise you're looking at 300+ iterations with 2000+ lanes each with 1000+ physics steps each, so millions to billions of simulation steps.)
And that really makes me wonder, if 30 tokens for a robot can be so expressive that they replace 600 mio simulation steps?
Or the alternative would be that they just assume that the given trajectory is equally optimal for each robot, which to me seems like a HUUUUGE assumption. But if that's the case, then the results of this technique would be highly misleading, as they would tell you the robot is excellent for 1 example movement, while you wouldn't know that a tiny change to your example movement might make the situation much better overall with a different robot.
by fxtentacle - I'm also interested in how they figured out how to get the CMA-ES baseline to be this fast (less than 35 samples in Figure 5).
Maybe they are warm-starting CMA-ES with a pre-trained controller (although at a glance it doesn't seem they do). Haven't read the paper yet though.
by dadoomer