Stepping Up | Reinforcement Learning with Spot | Boston Dynamics
Vložit
- čas přidán 20. 05. 2024
- In release 4.0, we advanced Spot’s locomotion abilities thanks to the power of reinforcement learning. Paul Domanico, Robotics Engineer at Boston Dynamics talks through how Spot’s hybrid approach of combining reinforcement learning with model predictive control creates an even more stable robot in the most antagonistic environments.
Learn more about Spot's locomotion control system: bostondynamics.com/blog/start... - Věda a technologie
Spots fall at 02:35 was one of the funniest things I saw today.
**swoop flomp**
I imagine him saying: "no no no no no no no" plop
Their robots are very cute even when they fail.
Great goat impression, Spot!
If you have ever lived in a country with ice on the streets, you have witnessed something similar or it has happened to you. That fast movement as you try to regain control xD
4:50 - amazing. I like how the folks in the background light up with joy. We only see highlight reels of the pretty stuff and I know there will be blood sweat and tears put into this, but even still, looks like a dream job.
This is not the first time BD is using machine learning, but integrated RL with their already successful models and methods are a big step forward.
Cheers to the Boston Dynamics team!
4:50 YOOO this jump was so impressive!
4:51 Staff in the background real happy, he made the jump 😂
0:54 seems pretty accurate to how dogs would act in that scenario of rubble 😂
i am assume they will try to jump, pass around and if nothing can help will pass through this things any way
I love that they show the "fails" to 👍🏻
Without failure, you can't succeed
Too
failings its the best way to learn and make progress. i wish humanity was less ashamed in their failings
I love boston dynamics!
Awesome, combining your previous control architecture with reinforcement learning!
Can't wait to see that applied to atlas!
アトラスは比較的壊れやすいので、私はボストンダイナミクスがどのようにして強化学習をアトラスに適用するのか興味があります。
Reinforcement learning feels like the next wave of innovation in legged robotics. It’s awesome to see boston dynamics getting onboard as it becomes feasible.
4:51 That moment when the engineers sit back.... GOLDEN.
You guys are incredible 👏.
Showing the difference between each version is very interesting.
Great job Boston Dynamics!
I saw Atlas in a recent video with NVIDIA CEO in a line up with a bunch of humanoid robots.
We live in amazing times, and I hope that in our very near future that we can have versatile personal robots, especially if robots help in the raw materials, transport, and mining to self build, eventually even help us colonize space, by building cities & infrastructure well before we arrive.
Ah, the optimism of the 00s, a rare sight nowadays lmao
I was lucky to see Spot 2 years ago just randomly on the side walk on a Sunday afternoon, and upon seeing it for the most few minutes it became obvious just what you guys would focus in future- and now it’s played out just as so
This guy is so excited with his job.
1:33 Mujoco simulator, its open source and deserves more recognition
Commenting for later
How does it compares to Issac lab? Always want to try doing RL
You guys are doing such great work
Спасибо Вам за то, что Вы делаете 💣💥
2:33 takes me back to big dog. Its still so funny to watch them react and flip on a slippery surface.
Excited with all you're doing!!!
Makes you appreciate the human brain all the more. We take for granted things that come naturally to us.
The hybrid approach seems to be a great way to go!
These crashes of the older version made my day 😂
Boston Dynamics makes such good promo videos, it is amazing!
Great work
Awesome, you're simply top-notch guys 👍
Happy to hear Boston Dynamics took a step toward involving machine learning to train robotics.
To be fair, they've had some level of ML for a long time, but now they are having reinforced leaning integrated, it's gonna take it to the next level for sure 😎
Incredible combining your previous control architecture with reinforcement learning!
Can't wait to see that applied to Tesla car!
this sounds exciting i cant wait to see what you do with it
Amazing ❤
Thany you for posting
I like that they show a "crappy situation" to 👍🏻
very cool results, nice to see you guys embracing this approach
Fascinating
1:42 Mr. Robot style forced rule of thirds framing. In fact he is the Mr. Robot of Boston Dynamics probably. How cool is that.
B2studios has some wonderfully visualized learning algos. Nice box climb.
This is about to be a massive step for Boston Dynamics.
It's fascinating how they can execute numerous forward passes in such a brief period.
glad to see it
I love robots boston dynamics!
Spots clashed at 02:35 was one of the cutiest things I saw now. =)
How long should it swing its leg, how long should it swing that leg, and where should it swing its legs, these are all good notes on constructing robots. Thank you! 1:10
Nadie nace aprendido. Todo a su debido tiempo para poder seguir adelante por medio del aprendizaje y experiencia !!
I've been waiting for reinforcement learning applied to Boston dynamics for a long time, I'm glad it's becoming a reality. Now it's only a matter of time before we get mech fights xD
This is pretty big. Congrats.
As someone who works with RL, this was super interesting! Thank you for slight the peek behind the curtain :)
a video thats actually useful. wow
Very cool!
Great, thanks.
"And it seems to be able to balance these solutions in ways that we were better than writing ourselves"
love it... Spot is my favourite by far...
Spot on comment
With Boston Dynamics making some serious headway. And NVIDIA's GR00T/ISAAC Robotics and systems, I can't wait for the future! I need my robo-buddy!
Boston dynamics, the leader in the robotic Armageddon
the dancing and extremely cute robotic armageddon.
well that sums it up!
Since Hyundai took over these videos are so corporate.
A business survives by how it profits, Boston dynamics was previously just a research company but now they're trying to sell products and I hope they succeed
@@notebe528and when the business aspect becomes the focus we get another Boeing
@@notebe528not what he was getting at in the original comment but ok.
always were. just worse disguised now
Making goofy dance videos is cool and all but these videos are targeted towards potential customers. You know, other companies that PLAN on buying these. Not people who only cares about "woah cool robot does a backflip!".
4:40 Agreed. A model remains just an approximation, or a "map". So it is basically just a guideline. And guidelines are there to be bent and broken.
4:51 emploers in the background real happy, he made the jump 😂
ボストンダイナミクスが強化学習を使用していることを初めて知りました。私はロボットの動作を洗練させるためには強化学習が最良だと思います。
Cool cool cool. Finally, Boston Dynamics use RL.
"Just remember you break it and it's coming out of your salary" - some factory manager somewhere who keeps finding broken Spots next to a roll of shiny wrap...
boston dynamics the best
0:54 seems nice accurate to how pigs would act in that motion of rubble 😂
I believe in Boston dynamics other than any robotics company...😊
This is very cool. Reinforcement learning is the way to go. Keep up the good work, guys 👍
Yay… RL in real world use. Really fascinating technology with a lot of potential.
Опять нам это гавно подсовывают😅 пятилетней давности. раньше за 5 лет Прогресс был, а сейчас они стоят на месте. Хотят чтобы мы в это поверили😅 они уже давным давно должны прыгать и бегать и скакать галопом, а они всё так же ходят как и раньше, как 10 лет назад 😅 хотят чтобы мы в это поверили.😅😅😅
@@floks700you gonna come up with something original or just parrot the same moot point?
@@thecookiemeister5374я должен доказывать тебе чтото, или иы сам поймёшь свою глупость?😅
Silly question: Do these robots have a camera capable of seeing where their feet are going at all times? Does this information feed back into some kind of model that enables them to correct on the fly?
The leg actuators know the positions and the loads on all joints, there is no need for a camera in order to know where the legs are. There are cameras that look all around the robot though.
@cogoid thanks for the reply. My nervous system knows where my legs are in relation to my body too but sometimes I like to triangulate that information with my eyes. Was wondering whether having multiple sources of info might help with balance.
That's so cool. I wanna do this job so bad. I have so many ideas.
Great work, interesting
Really need tactile feedback on the toes so it knows when it's slipping or when there is nothing there so it can react instead of stepping and then making the calculations and adjustments. Then the hard part is matching the cameras with the tactile feedback to get more accurate steps.
4:50 wow
Are you using the same model as the bare robot when the agent is loaded with accessories?
This has more chances of producing a true inteligence. Motion is almost everything.
what do you guys think about figure one being the next leap in technology with ai and robotics
Incredible they started using RL just now. I was amazed the first time I knew about boston dynamics and they didnt use machine learning, only precise physics calculations.
They've used machine learning for a while now, they are now using integrated RL, which is indeed a huge step forward
@@MelloCello7 what do you mean integrated?
@@BrianSalamoneThey are able to, somehow, blend some of the traditional static dynamics programing that had with RL to get some of the staggering results you see in the video today!:)
when is the release that learns CQB?
Do you simulate in nvidia omniverse? Are there relations to Huang's yesterday presentation?
Is there a schedule for when we can watch robots dance on ice?
Figure skating robots would be the most extremely impressive demo ever.
Has any paper about this ? Paper about Spot working with RL?
We published a technical blog on the topic which includes more detail on our approach and how we're collaborating with others in the robotics research community. bostondynamics.com/blog/starting-on-the-right-foot-with-reinforcement-learning/
Это невероятные вещи
How does the advancement in spots locomotion benefit atlas? Can the same methods be used there?
They do indeed need to step it up, getting behind with all the other robotic developments! Still cool tho!
Man that's so cool. can it see glass.
В гражданской версии демо ролика врядли они это покажут на публику. Идёт война. А они как будто остановились в развитии на 10 лет. Кто в это поверит?😅
I though reinforcement learning were already the main training method for the spot since day 1... nice to learn that...
Ok, now I want to see how atlas handles a slippery surface 😂 I can imagine it doing the arm circling technique to stay balanced.
That's great! How do I apply?
Very cool. Seems like a genuinely intelligent and nice guy. Seems like he is severely underslept.
I wanna see someone teach Spot to do cartwheel locomotion. That would be truly terrifying 😂
03:18 on the left: me on beer. just kidding. no beer needed. it would accelerate result tho. i hope someday i can ride home on one of these dudes - when got stuck in a bar. cheerz from germany
❤
Interesting
3:20 And this is why it won't be long before tons of labor is automated, same hardware but you just quickly replace the software it uses with a newly trained version and it performs much better. Image the recent viral figure 01 robot having a multimodal GPT5 instead of just GPT4 vision for example.
Это всё дорого и уязвимо, есть дешовые люди, эти бандиты не остановятся ни перед чем.
the reflections in the simulation seem like an odd detail. why did you feel the need to program reflections?
"There's no perfect model only useful one."
-B.D. Engineer.
I don't understand a bit, but I'm impressed ...
Traction Control and ABS for robots.
That's what he's talking about.
@@Damien.D ...Oh, I see, so it is comparable to how a car works?! 🤔
@@a.mie.533that the point but on legs and on multiple terrain and obstacle types
yes
I'm begging for robots with flexible joints trained in a soft body world
What is reinforcement learning?
A program that randomly selects actions and learns by getting rewards (higher scores) based on results they achieve. Over time the probability of selecting the right actions grows larger based on what actions achieved high scores. The system still can randomly choose a slightly different way of doing things, but in general it will over time succeed in choosing the right actions.
what chip is in that bad boy?
Are you using something that would be kind of analogous to a human's reflexes? Like, an additional "short circuit" system, to handle e.g. things like slipping ;)
..looking at those slippery tests, it feels like a freeze reflex would come in handy, same way humans immediately stiffen their joints the moment the reflex system recognizes a sudden slip motion.