NVIDIA AI: Physics Simulations Now Outpaced by Artificial Intelligence
Here’s a transcription and a rewritten blog post based on the video content:
Transcription:
“You are going to see many insane things in this video because this research paper is jaw-dropping. It is teaching a robot to dream, yes, and dream so vividly that these dreams are going to start coming true. True insanity. Okay, so I bet you saw many of these videos of robots doing these really impressive acrobatic things. Parkour, flips, dance routines, you name it. These are really making the rounds in the media these days. It looks like we suddenly jumped a hundred years into the future, right? Well, not quite. But very few know that these are not the hard problems. Really? Why? Well, because these all happen in controlled environments. Every step is known in advance. Yup. This is easy. The really hard problems are the mundane, messy ones. Handling and grasping small or deformable objects. Now, that is hard. And it gets worse: doing it with new objects, on new surfaces, under unseen lighting… that is nearly impossible. These demos are super impressive, but they contain none of the real challenges, so they can be even hand-coded. They don’t need the robot to adapt to new or changing situations. So, in short, kung fu, easy-peasy, but being useful is nearly impossibly hard. I don’t need you guys to do kung fu, get out of here. I want you to fold my laundry instead, thank you very much! Okay, but how? Well, scientists typically train a robot in a simulation before deploying it into the real world. It’s like learning inside of a video game. Imagine a self-driving car, first you are in a game, and when you drive safely for a long time, you get a good score. And then you can you proceed into the real world. But only then! Sounds good, right? But whenever I visit incredible robotics labs around the world, I always come back with one conclusion: things work really well in a simulation, and when moved to reality, things break like crazy. Until this incredible new work appeared. It sets out to solve two really tough problems: one, perform predictions.”
Blog Post:
Robots are Dreaming, and it’s Changing Everything
Robots performing incredible feats of parkour and acrobatics dominate our screens, leaving us in awe. While these displays are impressive, they often mask the true challenges in robotics. The real hurdle isn’t flashy movements in controlled environments, but the mundane, messy tasks of everyday life. Imagine a robot folding laundry, reliably grasping delicate objects, or adapting to different lighting conditions. This is the “holy grail” of robotics.
The Hard Reality of Robotics
Many impressive robotic demonstrations take place in carefully orchestrated settings. Every step is pre-programmed, and the environment is perfectly calibrated. This makes complex maneuvers like flips and dances relatively simple to achieve. However, these demonstrations often fail to address real-world complexity.
The real challenge lies in enabling robots to interact with unpredictable environments. Tasks like grasping deformable objects or manipulating tools in varying lighting conditions are incredibly difficult. Robots struggle to adapt to new objects, surfaces, or unexpected changes in their surroundings.
The Simulation Solution (and Its Limitations)
Traditionally, robots are trained in simulations before being deployed in the real world. Think of it as a video game where a self-driving car learns to navigate roads. The more safely it drives in the simulation, the better its score, and the more prepared it is for the real world.
The problem? What works perfectly in a simulation often falls apart in reality. When robots transition from the virtual world to the physical world, they encounter unexpected obstacles, variations in lighting, and countless other factors that were not accounted for in their training. This “reality gap” has been a persistent challenge in the field.
Dreaming Robots: A New Approach
Exciting new research is tackling this problem head-on, exploring an innovative approach: teaching robots to “dream.” This innovative approach aims to solve two significant challenges: enabling robots to make better predictions. By allowing the robot to generate realistic simulations of potential actions and their consequences, it can better prepare for the challenges of the real world.
#NVIDIAs #Real #Physics #Slow
Thanks for reaching. Please let us know your thoughts and ideas in the comment section.
Source link
About The Author
Is it just me or are some of the NeRD trained simulations better then the ground-truth simulations?
congrats on 1000 episodes!
It's all nice and cool but can we actually use it for our own projects or is it behind closed doors. Because so what's the point. I would love to use it in like unreal project to create my own ARC robot.
We gpt street smart AI before GTA 6
Could we see finally good water simulation in games? …
Let's get 1000k likes for 1000th episode guys 🎉
This was longer than 2 minutes lol. Why do you sound like Dr. Nick from the Simpsons. very cool video and simple explanation though!
So it s faster cheaper and introduce noise automatically ?
Here's the funny thing about reality vs models:
Reality is always right. The trick we're doing with AI solves problems in any domain reliably: Test, fail, iterate. Test, fail, iterate. It's why Thomas Edison found 1000 ways lightbulbs didn't work and also one that did. The models only point you in the right direction. Reality tells you what's true.
it just gets better and better. And 1000 episodes!! WHAT A TIME TO BE ALIVE!!!
I'll take, "things no one wants", for a thousand Alex.
Step 1: motion-capture yourself folding laundry, applied to a virtual Unitree G1 robot
Step 2: this special NeRD sauce
Step 3: Profit!
so iRobot it will be two papers down the line
Ah yes, lets use AI to train humanoid robots to get very precise motion. That will surely turn out great for humanity. This honestly scares me a lot. AI 2027 is looking more an more probable.
Idk man, when I see "AI" in the title and I am no longer excited..especially in combination with Nvidia.
I feel like we still need the regular old "deterministic" solutions to do actually useful stuff.
AI deals with mainly hype.
these changing graphics are making me dizzy and sick…. like hypnosis
AI is actually more similar to how us humans move and do tasks because we are not constantly doing math calculations. We just look at things and act.
Just like the simulations.
1:29 "… fold my laundry instead!…" // https://www.youtube.com/watch?v=KKUaVzf3Oqw&t=1s (to be fair… that was 15 years ago… 😉
cure cancer when? reverse aging when? nuclear fusion power plants when? power bills super high, quality of life dropping, food prices and power bills skyrocketing, but i get pron, cat videos, and ai (actual indians) cratering the job market. honestly, i dont care about nvidia propaganda anymore.
I don't find all the explanations clear.
But…
I love Kung Fuuuuuuu!
Great stuff but 1000 episodes of great stuff is that much greater, congrats!
I must say, I'm very hesitant about a probabilistic physics approximation as opposed to an actual simulation. it looks good on these small scale tests, but they're not identical, and as we've learned time and time again, the devil is in the details and we best not get them wrong. I'm worried about these small inaccuracies compounding in practical application.
that being said, the tech itself is very impressive. but I'll only consider this an appropriate substitution for actual simulated physics when the results are identical.
So the first 6:39 of the 9.5 minute video was apparently just the intro?
That's cool how do I put it in my game engine?
Love the indian accent
I'm so sick of seeing a new "me too" Chinese robot video of 20 year old tech doing remote control, pre-scripted playback. And guaranteed there is an army of bot-like comments saying "China # 1 leader in robotics"
No, I'm sorry Chang, I know your Great Internet Firewall keeps you ignorant, but the rest of the world is not impressed.
Is this voice AI? 🤔
I don’t know what to say.
I can't wait for all these new physics techniques and neural learning to be combined and efficient enough to use, the day this trickles down to be used in engines and frameworks will be amazing.
Congratulations on the 1000!! What a time to be alive!
What a great paper for 1000 too! NeRD, how about that 😂
Super exciting stuff! Lovely to see some optimism with all the doomer attitudes these days.
one day I will open my YouTube and two minute papers will release a video on how AI as actually taken over the world.
OMG, it went beyond graphics. AFAIK – it was the greatest challenge in robotics – training the robots to move. It was slow and expensive. This tech will make it a lot faster and cheaper. For me it looks like a singularity in robotics.
When you realize our reality is just one of those simulations. And that our god is training us to be his slave to do laundry amazing. This is the cosmic horror truth XD
It's called 2 minute papers but I'm 2 minutes in and we haven't seen the paper yet.
love your content ❤
Incredible 1000th episode and super fascinating paper on display! it'd be awesome to see it when they add humanoids into NeRD
If I was working in robotics I'd think twice about how my robot might be used – to put people out of jobs, or to kill people I don't know.
Congrats for 1,000 endangered papers saved!
Wow its better than real life
Incredible! Can't believe it's been 1000 videos! Congrats and thank you for your dedication!!
They gonna take your job and not gonna use the laundry 🙂
Congrats on the 1000th episode! Just imagine where we'll be just 1000 more episodes down the line!
Make it happen!
Yeah, good luck catching up with Tesla's billions of miles of real-world video footage with simulations. Pretty easy to make the car self-drive but insanely hard to make it work in edge-case. This is why Optimus 3 will also be 2-3 years ahead of everybody staging cool mechanical demos when it comes to deploying in the wild.
1000 episodes! Wow!