Comment by Nevermark | Hacker News Reader

Comment by Nevermark | Hacker News Reader

> Artificial Intelligence researchers have been trying to get [X] to [Y] for over 65 years

For 10,000 different problems. A great many of which have been solved in recent years.

Robotics is improving at a very fast clip, relative to most tech. I am unaware of any barrier, or any reason to infer there is one, for dextrous robots.

I think the primary difference between AI software models and services, and robotic AI, is economics.

The cost per task for AI software is .... very small. And the cost per task for a robot with AI is ... many orders of magnitude over that.

The marginal costs of serving one more customer are completely incomparable.

It's just a push of a button to replace the "fleet" of chatbots a million customers are using. Something unthinkable in the hardware world.

The seemingly lower level of effort and progress is because hardware that could operate in our real world with the same dexterity that ChatGPT/Claude can converse online, will be extremely expensive at first.

Robotics companies are not just focused on dexterity. They are focused on improvements to dexterity that stay within a very tight economic envelope. Inexpensive dexterity is going to take a while.

One very important task to solve is the ability to select a box from a shelf and set it neatly on a pallet, as well as the reverse. People have been working very hard on this problem for a long time, there are impressive demos out there, yet still nobody is ready to set their best box manipulating robots loose in a real warehouse environment.

How hard can it be to consistently pick up boxes and set them down again in a different location? Pretty hard, apparently.

I mean with rigid plastic containers robots are 'pretty consistent' at it now.

The problem with things like cardboard boxes, especially at any size is internal weight distribution and deformation of the box. If you take someone that is pretty new to stacking boxes at a wearhouse and give them sloppy boxes (ones that bend or otherwise shift) they are going to be pretty slow at it for the first hour or so, then we'll internalize the play in the materials and start speeding up considerably while getting a nice result.

It's pretty amazing how evolution has optimized us for feedback sensing like this.

> I am unaware of any barrier, or any reason to infer there is one, for dextrous robots.

I don't think there's a fundamental barrier to building a humanoid robot but the cost will be an extremely high barrier to adoption.

A human is nature's ultimate robot: hundreds of servos, millions of sensors, self-assembling from a bag of rice, self-repairing for minor damage. You just can't beat that, not for a very long time.

Economy of scale. Musk in his vaporware style said a humanoid robot will cost 10k usd.

> I am unaware of any barrier, or any reason to infer there is one, for dextrous robots.

Pretraining data?

Synthetic data works better for robots since you can generate endless scenarios based on real physical laws.

Do you know of success stories here? Success of transferring models learned in physics simulation to the real world.

When we (ZenRobotics) tried this 15 years ago a big problem was the creation of sufficiently high-fidelity simulated worlds. Gathering statistics and modelling the geometry, brittleness, flexibility, surface texture, friction, variable density etc of a sufficiently large variety of objects was harder than gathering data from the real world.

Google has done training in simulation: https://x.company/projects/everyday-robots/#:~:text=other%20...

I believe this is the most popular tool now: https://github.com/google-deepmind/mujoco

Thanks for the links.

AFAICT these have not resulted in any shipping products.

We have massively better physics simulations today than 15 years ago, so the limitations you found back then don't apply today. It might still not be enough, but 15 years is such a long time with Moore's law and we already know all the physics so we just needed more computation to do what is needed.

Example of modern physics simulation: https://www.youtube.com/watch?v=7NF3CdXkm68

So much easier in many ways.

You can train in stages.

First stage, either digitally generate (synthetic) basic movements, or record basic human recorded movements of a model. The former is probably better and an generate endless variation.

But the model is only trying to control joint angles, position etc. no worries about controlling power. The simulate system has no complications like friction.

The you train with friction, joint viscosity, power deviance from demand based on up, down times, fade, etc.

Then train in a complex simulated environment.

Then train for control.

Etc.

The point being, robotic control is easy to be broken down into small steps of capability.

That massively improves training speed and efficiency, even potentially smaller models.

It is also a fear simpler task by many orders of magnitude to learning the corpus of the written internet.

Comparable to that, would be training an AI to operate with any land, sea or air device. Which, nobody today is trying, (AFAIK)

It's so easy! I hope all the leading robotics researchers come to find this comment and finally deliver us the dexterous humanoid robots we've all been waiting for

Well, in fairness, the kind of deep neural architectures needed to do this stuff have only been available for a relatively short period. The robotics researchers in my institution are basically racing to put all this new capability to work.

Eg: https://hub.jhu.edu/2025/07/09/robot-performs-first-realisti...

Reply on news.ycombinator.com