Learning Dexterity

We’ve trained a human-like robot hand to manipulate physical objects with unprecedented dexterity. Our system, called Dactyl, is trained entirely in simulation and transfers its knowledge to reality, adapting to real-world physics using techniques we’ve been working on for the past year. Dactyl learns from scratch using the same general-purpose reinforcement learning algorithm and code as OpenAI Five. Our results show that it’s possible to train agents in simulation and have them solve real-world tasks, without physically-accurate modeling of the world.

Directed by: Jonas Schneider
Starring: Alex Ray
Sound Supervisor: Larissa Schiavo
Production Design: Ben Barry
Production Manager: Diane Yoon
Stills Photographer: Eric Louis Haines
Music: "Bring Your Own" by Dexter Britain, "Intermediary" by Mattijs Muller

[kirarpit]

Both the robot's dexterity and the guy's hair are unprecedented

[DaveAlexKD]

2:45 when your robot hand can't stop telling you jokes

[opl]

This is starting to resemble an Aperature Science trailer...

[Molenkof]

This guy domain randomizes his hair...

[strange]

Cut to the team's "stock laughter" poses, then slow it down.
Oh, AI hand! You say the darndest things!

[PandoraMakesGames]

Cool to see other projects, next to the Dota project!

[flaccidpen]

The ultimate search for the female G-spot draws nearer

[SimulationSeries]

this is incredible work, thank you OpenAI for bringing this forth!

[mishterpreshident]

... when the most distracting & astonishing thing in the room ISN'T the functional robot hand solving logic puzzles ...

[dreamxcviii]

This is definitely the most advanced robotic hand I've see to date as far as being able to do what a normal human hand can, the problem with all the other robotic hands that I've seen other companies make are always the same 2 problems, #1: Either the base of the thumb has limited mobility such as being able to lay flat next to the side of the index finger while also being able to swivel around across the palm to touch the pinky or #2: The palm of the hand where the pinky is connected to has absolutely no movement in it whatsoever which isn't how our natural hand is for example if you grab your palm just under what would be considered your pinky finger and move it up and down you'll notice that we don't just have three joints that make up our pinky finger but actually four since that entire bone below it can swivel back and forth as well, this is also the case with your ring finger but to a more limited degree but with your middle finger and index finger those bones are more or less fused to the bones in your wrist so it gives us three points of motion for those two fingers instead of four. Another thing I noticed with this hand is they captured the ability to move the fingers side to side as well which is also a small but important detail to the mobility of our hands and all these come together to as I stated before create the most advanced robotic hand I've seen, other companies are further along tho as far as integration into the human body goes so it blows my mind that these companies just don't come together to bring all their strengths and bring us into the future at a much faster rate

[d-fan]

2:32 Hehe. "Hand-programmed robots"

[javiera.riverop.]

amazing, just amazing.


the only problem i see is the amount of energy and time one hand replacement needs to achieve a goal like this. idk how much "100 virtual years" are, but of course it is quite a lot.


i hope you guys keep up the good work and improve the learning methods. a lot of people would be in eternal gratitude with you.

[blaster]

This is very awesome. The work that OpenAI does is very impressive.

[kevin_delaney]

This is extremely interesting! I find Robotics and AI (in general) fascinating. I am in awe so often, yet I continuously want to tinker with and experiment with machine learning and robotics. Unfortunately, in high school and even in college the first few years, I was still completely oblivious of this technological advancement within computer science. It was not until after I had chosen to drop out and educate myself did I finally stumble across machine learning, I don't know...maybe two years ago. Even after many hours of trying to mark linear algebra, learning about motors, microcontrollers and computers, I still feel as though I am years away from fiddling with technology as sophisticated as this. I wish places like this were open to the public and relatively local because I genuinely find this fascinating. I wish I had been exposed to this field sooner because I am damn near obsessed, but cannot really financially afford to completely immerse myself like I'd like to. It is such an interesting field in my opinion. I wholeheartedly believe that just kind of playing with robots and machine learning would be ground breaking.

I was going to suggest a DIY kit for this until you said "cloud" as in cloud computing! 😂 Nevertheless, I wish it was easier for individuals like myself to get into fields like this. With the upfront costs of hardware and the genuine complexity of this, it is difficult to get into machine learning without being required to pay for server time or courses. With what school costs nowadays, and how profit focused schools are today, it's not always easy to find high-quality education on machine learning and advanced robotic concepts, let alone inexpensive tutorials.

Interesting for sure, it seems far out of reach for individuals such as myself though. I find that pretty ironic because high-end motors are easy to come by for me and that pretty much just looks like 30x30 or 40x40 aluminum extrusion with some cameras and some sensors connected a microcontroller with some motors that gets inputs from an outside source, in this case the cameras and computer(s) that do the computing. It looks relatively simple* (conceptually speaking), but would certainly be curious about what the code would look like.

[BLAISEDAHL]

Nice move on putting the Open AI on the cube, I like it!

[Gottii]

we need more, and we need tutorials, please openAi, you do such great work let us be part <3

[willasn]

Openai may not be as big as deepmind. But they definetely have to be taken seriously.

[chanyy]

If I want to try this myself, what are the dimensions of the cube?

[joshuacrandall]

2:27 The robot AI is still trying to figure out the correct way to tell his overlords he won't be playing with anymore blocks today. You almost got it buddy.

[SvetlinTotev]

But can it flip a butterfly knife? Or solve a Rubik's cube? I suppose at the moment it only knows how to manipulate solid rigid single object systems. Also I suppose it doesn't have pressure sensitivity which would make it way too hard for learning one of those tasks.