00:00Hi everyone. So we started Genesis about a year ago with a simple question of like,
00:05within the next 10 years, how can we get a billion robots on the planet?
00:09And if you think of what robots do today, it's typically industrial arms, and they do the same thing all
00:13day.
00:14And what would it take to get robots to do everything else?
00:17And the key is AI. So the idea is to scale AI like we've been able to scale LLMs,
00:22and that means scaling the data that goes into training general purpose manipulation.
00:27And so this was a vision about a year ago, and a year into it, we've built every piece of
00:33vision that I'm going to walk you through.
00:37The first piece is the model. That's probably like the robot brain.
00:43So here you can see results from the model we released about a month ago.
00:48And the idea is to have a single robot brain that can control the robot arms and hands to do
00:53any types of tasks.
00:54Like for example, very dexterous tasks like making a 20-step omelette, as well as lab pipetting in a pharmaceutical
01:03lab,
01:04or cable harnessing in industrial use cases.
01:08All of these were really complex, and a lot of people thought they were not possible with hands today.
01:18So for example, this one is solving a Rubiky Cube with dual arms.
01:21Also, a lot of people thought this was not possible.
01:24And the idea behind this is to scale data, which I'm going to walk you through as well.
01:37So you can see the hands using a lot of tools.
01:41This is the value of having hands.
01:42You can use tools like humans do, and do any kind of manipulation that humans do.
01:47This is making a smoothie.
01:51Or even playing piano.
01:54And all of this is completely autonomous at 1x speed of what the robot actually does in the real world.
02:06And so the key to get a general-purpose AI brain is to be able to scale data like we
02:11have for internet.
02:12And the big issue has been that the way data is collected for robots today is really hacky.
02:18At core, you have robots that typically have grippers, while we humans have very complex hands.
02:24And that forces the data collection to have these very complex setups of either the human moves plastic exoskeletons.
02:30You try to move this piece of plastic to move the robot, but you can't feel what the robot feels.
02:34So it's really hard to scale data.
02:37Or another paradigm today is have humans adapt to the robot's embodiment and wear grippers.
02:44But this is also hacky, because if you're an expert mechanician or an expert in a lab or a cook,
02:49you just can't do your job wearing these grippers.
02:51So it makes data collection in real use cases really difficult.
02:55So if we take a step back, what we thought is, how about instead of having the human adapt to
03:00the robot,
03:01have the robot adapt to the human and build a dexterous hand that has the same form and function as
03:06a human hand.
03:07Full 20 degrees of freedom.
03:09As well as a glove that lets you collect data and directly transfer human manipulation data,
03:15which is the best source of data we have for any types of use cases, to the robot.
03:20So this is the way we scale data much beyond, like to tens of millions of hours,
03:24much beyond what has been done so far.
03:30So the second piece of this is, even if you're able to scale data to orders of magnitude more than
03:35what we can do today,
03:36there is still a lot of workflows in robotics that are not scalable.
03:40Like the first one is, how do you test your models?
03:42Typically, you have your model weights that you put on the robot and then you have a human trying to
03:47reset the scene many times in a row
03:48and let the robot try some stuff 200 times in a row for the full day and try to compute
03:54a metric of how well do you do at a specific task.
03:57And you need to do this, like you have a general purpose model.
03:59So imagine you need to do this across hundreds of different tasks, hundreds of different scenarios.
04:03It's hundreds of humans days for a single model evaluated.
04:07The robot wears out, the lighting changes, it's just impossible to reliably evaluate models.
04:12So that's another huge bottleneck on having general purpose models.
04:16And if you take this a step further, the way we get from Chagipity to cloud code is also millions
04:21of parallel coding environments
04:22and we need to ship this loop of interaction from the robot to the environment, to software.
04:27And that's what lets you parallelize and scale.
04:30So that's why we build simulation.
04:40Let me try to play this.
04:45So this is our Paris office.
04:50We spent a lot of time building the simulation stack from scratch, like the physics of how do objects interact
04:57with each other,
04:57the rendering, how do you generate pixels.
05:03And now you're into the matrix.
05:05This is the simulation and not the real world, but you can see it's still super realistic compared to what
05:09you would have in the real world.
05:38And you can see the real world and simulation side by side that really closely matches reality.
05:45And that's what lets us evaluate models and let them improve themselves in virtual reality through software at scale much
05:52more than is possible in real world.
05:56And so now that we have a great model that is trained through large-scale human data as well as
06:02simulation to test it at scale,
06:04you still need to put it on a robot because that's what customers expect.
06:07They just want a robot to sell their use cases.
06:09So yesterday afternoon, actually, we released our first robot, which we called Inno.
06:21So we spent a lot of time thinking about the design and it has to be human in function because
06:26you're learning from humans.
06:27So you need human arms, human hands, but you don't necessarily need the rest of the body to be human.
06:32Like you don't need a head because you don't have a brain.
06:35So we tried to simplify it to the essence of what it needs to be to fulfill its function.
06:44So I hope you like it.
06:45The idea is to have this same robot be able to be deployed in an industrial environment, in labs, in
06:51data centers, and eventually at home.
06:57So that's what we built over the last year or so.
07:00I'm really proud of the team executing across the model, the simulation, the data, the robots.
07:04The next phase for us is to bring it to the market.
07:07And so we are, because this is the way we'll get data from deployments that lets you train the better
07:11model.
07:12And so we're looking for partners across industrial use cases to deeply partner with over the next three to five
07:19years to make a plan together of how we're going to make robots coexist with humans and make everybody more
07:24productive.
07:25So if you're interested in working in like data centers or labs or industrial workflows overall, I would love to
07:31hear from you.
Comments