ANYmal Parkour: Learning Agile Navigation for Quadrupedal Robots
HTML-код
- Опубликовано: 28 сен 2024
- In this video, we demonstrate how our fully-learned method enables robots to conquer challenging scenarios reminiscent of parkour challenges.
The paper introduces a hierarchical formulation that trains advanced locomotion skills for various obstacles, including walking, jumping, climbing, and crouching. A high-level policy is used to select and control these skills, allowing the robot to adapt its behavior based on the environment at hand. Furthermore, a perception module is trained to reconstruct obstacles from occluded and noisy sensory data, enhancing the robot's scene-understanding capabilities.
Unlike previous attempts, our method does not require expert demonstration, offline computation, prior knowledge of the environment, or explicit consideration of contacts. It achieves impressive results solely through training on simulated data.
Our real-world experiments showcase the successful transfer of these learned skills onto hardware.
For more information:
- Visit our project website at sites.google.c...
- Read the paper: www.science.or....
The software and this robot is really turning into something amazing. It's interesting how it decides to walk diagonally when there's large gaps, as it gives it longer reach.
This is absolutely splendid! I don't have a science robotics subscription but I think from the abstract I recognize the genre of approach (learned/weighted accumulation of a set of pre-learned primitive **trajectories** at runtime). Other than some really in-depth stuff with bio-inspired circuits, this is the only approach that strikes as "animal like" instead of "robot like". Truly well done!
I used to think that the large knee (elbow ?) joint actuator protrusion on the mid span of the legs was a mistake.
However; I have reversed my position and now believe that this feature is a HUGE advantage.
Watch how the robot uses the " knee -elbow joint actuator" when climbing; it is truly remarkable.
It's likely that among their next set of goals; this team will teach the robot to crawl directly on its knees and elbows to reduce its overall height while moving in crouching mode.
Adding a heavily textured rubberized covering to the outer surfaces of the knees and elbows would enhance the grip of these points when using them for leverage.
I didn't think it through as far a you have, but I did think protective gripping "kneepads" would be a useful addition. Does the "shoulder" joint have the necessary articulation for crawling? Vertebrates use spine flexing, and hip / shoulder rotation to maximise crawling capability.
Each development step has made it better and far more capable. Great!
You should add a "dancing" skill to the locomotion module that activates randomly.
Glory to Machines, glory to the Universe.
Anymal will always live in the Spot's shadow, but we know he is the goodest boy out there.
This is so great, guys. This is the flexibility that robots need. I just wonder how some sort of language model can be integrated into this marvelous thing. For that, I imagine you have to simplify actions into commands the transformer can use, but how much can you simplify before you start losing flexibility
Yea it's starting to resemble an actual animal in how it move
I listened with the sound off and I genuinely can't tell whether this is real or cgi. It moves kinda weird and unnatural, I think that's why. Not necessarily a bad thing if it does the job well - certainly the design and especially the software is very cool :)
If this thing is chasing me, is my only defense an emp or jumping in the water? Please tell me this thing isn't waterproof.
This is amazing, only downside is that quadrupeds didn't look creepy to me before this. if this shit came after me with a self aiming turret on the back, i'm dead. the military will definitely find this interesting, which sucks
From learning to sharpen flint, through melting metal, to microchips... every time someone comes up with a fantastic advancement, someone else finds a way to use it destructively. That is the way humans are.
very cool! and funny to watch
It's crawling 😳
such a good boy ;)
I thought this was boston dynamics for a second
fake
Andy Serkis is so talented ❤️
😂
Your robot's legs joints and articulation are my fave of all the quadbots (I know it is Anymal but anyway). Everyone else is copying BD's Spot, but yours is uniquely yours and it seems way more capable than the other configuration. I also like that training a model by letting it have a go over and over end up with a very organic final technique which given the similarities to how an animal might learn should not be a surprise. Love it. I want one.
I agree, this leg configuration is the most capable of all, in my opinion. It largely removes the requirement for reversing down steps etc.
Boston dynamics started with these clumsy but intelligent robots. The more a robot looks like it is searching, failing, and trying another approach, I consider it more intelligent.
So, this one seems to be more advanced.
Standards for a good boi are rising.
impressive!
Looks like a puppy learning the world lol
Wow! great work! the different modules integrate so well with each other, will you consider making the code open and sharing your work?
Such incredible agility - wonderful work!
I love this channel, amazing work!
Maybe it could be an option to enhance the locomotion module's capacity by enabling it to generate distinct skills based on the environment
You've got intelligent animal 👏
Nice idea to give it knees and elbows it can push with.
Aww it's left handed
Nightmare fuel
Humanoid when?
You need to make shorter videos with the real-life demos only. These are the ones that will go viral.
99.99% of people don't care about technical stuff.
Literally Jimmy from South Park
You are on fire. Your research is looking amazing.
That robot is scary mobile! 🤯
The perception is insane.
:O :0 :O :O :0
Wow so impressive!! Are you guys using the Transformer architecture? or even generative motion by giving it input via text for what it should do then it generates the motion required based on its perception to accomplish the task? I think that would be the natural evolution of general purpose robots by giving it text on what to do like "walk to that tree then climb it" something like that.
To be clear I don't mean a language model per say but it would need to understand language. The perception via Transformer is the key part I think.
The dad and little kid at 3:45 have no idea that they are some of the first people to witness this.
It amazes me how many people walked on by without seeming to notice. Is this a normal activity at parks in Zurich?
Amazing and frightening!
Robots: Cool. Parkour: Super lame.
You come from the home of Josef Müller-Brockmann. It shouldn’t be difficult to find a good designer.