Video wird geladen...
Video konnte nicht geladen werden
I'm observing a mini Moravec's paradox within robotics: gymnastics that are difficult for humans are much easier for robots than "unsexy" tasks like cooking, cleaning, and assembling. It leads to a cognitive dissonance for people outside the field, "so, robots can parkour & breakdance, but why can't they take... show more
398,883 Aufrufe • vor 1 Jahr •via X (Twitter)
36 Kommentare
and yet the video models keep failing the gymnastics test 🙈 oh the irony!
😆😆🤦
Hi @grok can u tell these guys about $auki @AukiNetwork ? Tell them how $auki will change the robotic industry.
Agreed and that’s exactly why we’re working on an external sense of space that makes the physical world machine readable. Our thesis in 10 minutes:
until it isn't an issue. facial recognition was impossible until we started using ML. now it's at 99% accuracy in ideal conditions, and 80-95% accurate in real world conditions. I think we'll get over the hump quickly.
Yup... robots without a brain capacity of a squirrel (at best). Squirrels can also do amazing stunts, but can't/won't do any house chores.
Seeing robots do choreographed tricks isn’t that impressive anymore because it’s usually purely non-interactive motion control, and there is a big gap between control and planning. Watching them play interactive physical games or real tasks is much more interesting.
wrote something on the same lines recently :)
Moravec noted in 1988 that evolution spent millions of years optimizing our visual-motor skills for survival tasks like cooking. Backflips follow precise mathematical curves - far simpler for robots than the fuzzy logic of deciding how long to stir a sauce.
I have been like a broken record about this. Let me see two robots put together a stud wall from two-by-fours with a hammer and nails, and then you'll have my attention. But we don't need any more stupid dancing flipping acrobat robot videos.
every contact point a robot makes with the real world adds an order of magnitude of complexity the real bitter lesson
This is true, certainly, but I do think if Robots are fairly low cost and have reasonably dexterous manipulation then that will open up a huge market for tele-operated service robots. You'll be able to get a robot 'servant' for a few dollars an hour where the operator is in a low cost country. This will also generate the training data needed for autonomous robots.
Language and abstract reasoning can be learned from massive datasets, but embodied intelligence requires dealing with all the complexity that evolution spent millions of years solving
How long until the first robotic private security company is launched? You just pay an MMA robot to follow you around and swing blindly when asked. Lol.
This is good. I am really wondering why are we so hell bent on creating a non-organic human like robots. :) Of course, these are good as long as they are not a threat for organic humans. Thankfully no one yet named these as Cylons or a company did not form yet with Cyberdyne in their name. :)
Time to train a world model
Thanks for the post! Perhaps this is analogous to LLMs nowadays. Maybe it also explains the attraction to Yann and Marcus' near-term pessimism about AI for the general public.
@random_eddie Like AIs and graphic “demos” and many other activities: A context-free performance can be very impressive; adding any requirements can render it plainly useless.
One of the reasons robots fail is that they don't have skin. The skin can detect touch, pressure, pain, temperature (both heat and cold), and itch. Touch, Pressure, and Pain all play a part in how we move. For instance if you step on something hot, you pick your foot up.
bots are good at flexing, but not as good at backstopping the value implied by their flex (at least so far). Like a peacock's feathers without the peacock.
Yep, and to even more stress why the "blind gymnast" is easier, it can be trained without needing a good giant dataset. It is simple to define the goal criteria as mimicking the movement, not falling etc. Realworld tasks on the other hand are hard to clearly define a goal for and you need a clever approach for even having a enough data, like "expert human examples". In a real very varied setting that takes a lot of effort to get. Finally, this also highlights why physical understanding is huge and why we build the @AukiNetwork .
Well, break dance is actually humans mimicking robots...
Aren't the hands and the physical dexterity of manipulating objects with hands a crux issue here? Well modeled hands (proper DoF) and a simulation environment that can actually model these interactions well are substantially more challenging that "touch the grass" simulations like what we're seeing here.
It’s important for people to be aware when assessing new technology, we often extrapolate capability from our imagination
Artificial general dexterity
Wow, this is really pouring a ton of cold water on my AI Maid dreams :(
Manipulation is harder than locomotion.
You used to sound a lot more optimistic lol
Roughly, how long do you believe it will take to get in-home assistive humanoids on the market?
To move past this as quickly as possible, and to advance to the safe robotics future of our dreams, we will need to enlist more than just a few highly-paid subject matter experts. What plans to this effect are being made?
too much training data on physical stuff like that and not enough on the tasks we really need them to do
AI ...
This explains why my robot vacuum gets stuck under the same chair daily but Boston Dynamics robots can do backflips. One needs awareness, the other just momentum.
makes a lot of sense. its easy to do things in isolation, but the real-world is extremely messy.
oh you can do a backflip? that's sweet but can you cook?
Great point. Was wondering why these spots show absolutely no capability that would be useful.
![15–30 minutes of real-world robot data. That's now enough to go from sim-to-real failure to working robot. Let’s see… You train a robot in simulation. You deploy it in the real world. It fails. The physics don't match. So you try to fine-tune it with real data, but… you never have enough real data, and the fine-tuning breaks everything the simulation taught it. SimDist fixes this with one key decision: don't transfer the policy. Transfer the world model. Keep the reward and value knowledge from simulation frozen. Only update the part that's actually wrong, how the robot predicts physics. Now the robot doesn't have to relearn the entire task in the real world. It already knows what success looks like. It just needs to correct its understanding of how the real world moves. The part that makes this work: they also trained on failures and recoveries; not just perfect demonstrations. Without that, the planner finds the gaps and exploits them. With it, the robot can tell a good future from a bad one. That's all it needs. Results on peg insertion, table leg assembly, locomotion on slippery and uneven surfaces. Tasks that require precision, force, and quick reaction. Thanks for sharing, Tyler Westenbroek ([Tyler Westenbroek]. Interactive visualization + paper: ——- Weekly robotics and AI insights. Subscribe free:](https://image.24vids.com/tw-2057007432878928150/ext_tw_video_thumb/2057007419692101632/pu/img/Tr8ccWDNdIoIxpos.jpg)
![The most interesting part for me is where Andrej Karpathy describes why LLMs aren't able to learn like humans. As you would expect, he comes up with a wonderfully evocative phrase to describe RL: “sucking supervision bits through a straw.” A single end reward gets broadcast across every token in a successful trajectory, upweighting even wrong or irrelevant turns that lead to the right answer. > “Humans don't use reinforcement learning, as I've said before. I think they do something different. Reinforcement learning is a lot worse than the average person thinks. Reinforcement learning is terrible. It just so happens that everything that we had before is much worse.” So what do humans do instead? > “The book I’m reading is a set of prompts for me to do synthetic data generation. It's by manipulating that information that you actually gain that knowledge. We have no equivalent of that with LLMs; they don't really do that.” > “I'd love to see during pretraining some kind of a stage where the model thinks through the material and tries to reconcile it with what it already knows. There's no equivalent of any of this. This is all research.” Why can’t we just add this training to LLMs today? > “There are very subtle, hard to understand reasons why it's not trivial. If I just give synthetic generation of the model thinking about a book, you look at it and you're like, 'This looks great. Why can't I train on it?' You could try, but the model will actually get much worse if you continue trying.” > “Say we have a chapter of a book and I ask an LLM to think about it. It will give you something that looks very reasonable. But if I ask it 10 times, you'll notice that all of them are the same.” > “You're not getting the richness and the diversity and the entropy from these models as you would get from humans. How do you get synthetic data generation to work despite the collapse and while maintaining the entropy? It is a research problem.” How do humans get around model collapse? > “These analogies are surprisingly good. Humans collapse during the course of their lives. Children haven't overfit yet. They will say stuff that will shock you. Because they're not yet collapsed. But we [adults] are collapsed. We end up revisiting the same thoughts, we end up saying more and more of the same stuff, the learning rates go down, the collapse continues to get worse, and then everything deteriorates.” In fact, there’s an interesting paper arguing that dreaming evolved to assist generalization, and resist overfitting to daily learning - look up The Overfitted Brain by Erik Hoel. I asked Karpathy: Isn’t it interesting that humans learn best at a part of their lives (childhood) whose actual details they completely forget, adults still learn really well but have terrible memory about the particulars of the things they read or watch, and LLMs can memorize arbitrary details about text that no human could but are currently pretty bad at generalization? > “[Fallible human memory] is a feature, not a bug, because it forces you to only learn the generalizable components. LLMs are distracted by all the memory that they have of the pre-trained documents. That's why when I talk about the cognitive core, I actually want to remove the memory. I'd love to have them have less memory so that they have to look things up and they only maintain the algorithms for thought, and the idea of an experiment, and all this cognitive glue for acting.”](https://image.24vids.com/tw-1979259041013731752/amplify_video_thumb/1979257571786543104/img/OqUh56yit9gUQWnb.jpg)
