00:00So let's talk a little bit about Google. Why humanoids?
00:03Yeah, I mean, humanoids are super interesting.
00:06First of all, robotics in general is a super interesting area.
00:08It combines all of this different expertise, like machine learning and physics simulation and AI in general.
00:15Humanoids are particularly interesting for many reasons.
00:18One is that they can actually go into regular everyday environments, which were designed for humans.
00:23The other reason, from the machine learning perspective, is that humanoids have the shape of humans.
00:28So perhaps we can learn from humans doing different tasks and transfer that onto humanoids.
00:35So we have seen the incredible progress in artificial intelligence, but embodied intelligence has felt like it's lagged behind.
00:42Is that an opportunity for Google DeepMind?
00:44Absolutely. I think embodied intelligence is the next frontier on AI.
00:48And we've been working on bringing Gemini into the physical world.
00:51And what that does is that it brings all of Gemini's world understanding, multimodality, in order to enable robots to
00:57understand their environment, to reason, and to be able to take action to the level of precision of a human
01:03expert.
01:03I just mentioned that partnership with Boston Dynamics.
01:06So Gemini will go into Atlas. We're going to spot the little dog as well.
01:11Tell us a little bit about what the future looks like for DeepMind and Gemini Robotics.
01:16Yeah, so we're really excited about Gemini Robotics bringing all of that intelligence from Gemini into the physical world.
01:21But there's still a lot of work to do.
01:23Gemini Robotics is able to give you that reasoning.
01:26It's able to give you that interactivity.
01:28It's able to give you multimodality.
01:29But it's not yet able to – what we're pushing the boundary on is on doing highly dexterous tasks, like,
01:35for example, folding origami or packing a lunchbox.
01:39That requires a lot of dexterity that humans have, and we don't really realize it, but it's incredibly important to
01:44make robots useful.
01:45When it comes to scalability of the industry itself and competing also with dozens of these new companies, not only
01:54in the U.S., but dozens in China as well, where is the edge?
01:58So there is a really exciting time right now for robotics all over the world.
02:03I think that the really hard problem that people don't realize is that the edge is understanding the nuance and
02:11complexity of the human world.
02:13Actually, a lot of what you see out there is predefined sequences, memorized sequences that the robots are doing.
02:19The actual intelligence needs to be there in order for robots to operate in all of our environments.
02:24Our environments are constantly changing.
02:26There's humans in them.
02:27They're unstructured.
02:28That is what's needed in order to get robots to be really helpful in the physical world.
02:32And that progress has been invents, right?
02:34I mean, you had the launch of LLMs, and then now you have robotics transformers.
02:38Tell us about the evolution here.
02:40Yeah, I mean, it's been phenomenal to see.
02:42We've introduced over the last few years a few advancements that have now become table stakes in the field of
02:48general-purpose robots.
02:49So we introduced LLMs and VLMs to robots.
02:52That was the first time that we enabled robots to understand natural language, understand their environment.
02:56We introduced transformers, and that shifted robotics into the era of data-driven robotics.
03:01It also introduced a new type of foundation model called a VLA, which is a vision language action model.
03:07And it enables robots to understand their environment, understand language, and then transfer that into very general behaviors.
03:13And then we've also introduced things like reinforcement learning in simulation.
03:17And that's actually what enables robots now to imitate humans and to operate in very unstructured spaces.
03:24So that's actually what gets used in a lot of these demonstrations where you see robots doing really cool acrobatics.
03:30It's reinforcement learning for whole-body control.
03:32And then lastly, we've shown what is possible when it comes to learning from imitation, highly dexterous behaviors.
03:37So we have examples of robots tying shoelaces, which is something we never thought we would be able to do
03:42in our lifetime.
03:43And you can take this, if you can combine high dexterity with precise whole-body control and high-level intelligence
03:49to understand the complexity of the environment,
03:51you can start to see how you can have robots that are actually quite useful.
03:54The complexity of the environment itself, of course, not just a challenge, but also something that needs to resate the
04:01framework for how you see safety when it comes to humanoids and interactions with people.
04:06How is Google DeepMind seeing this?
04:08Yeah, I mean, we're incredibly excited about the progress, and we want to make sure we're doing this in a
04:12responsible way.
04:13So we actually have a very important and deep program around safety.
04:18So we think safety as like a multi-layer approach.
04:21So safety includes, for example, just making sure that the robot can handle any changes in its own physicality, like
04:28if something goes wrong with the system.
04:29That's functional safety.
04:30We also think about control safety, make sure the robot is stable and doesn't fall or doesn't hit anything.
04:35We also have semantic physical safety, which is sort of giving the robot common sense.
04:40So you and I think about, for example, walking around a puddle or trying to not put an object too
04:46close to the edge so it doesn't fall.
04:48That's semantic physical safety.
04:49That's something that can only be brought through intelligence.
04:52You're here in Tokyo.
04:54When it comes to the advancement and progress of this technology, we're seeing also the geopolitical backdrop between the rivalry
05:00in the U.S. and China as well.
05:02How do you see this complexity of the landscape as you see the future of humanoids?
05:08Yeah, I mean, I think it's a really exciting time.
05:10And we're actually, I would say, still quite early.
05:13The progress is happening extremely fast.
05:15I think what's going to really unblock this is really being able to bring that accelerated hardware trajectory that we're
05:22seeing along with accelerated intelligence.
05:25So really, that's what's going to get us from just fixed sequences of robots, like fixed behaviors, to actually something
05:32that thinks in the moment, understands the situation, is able to do something useful for us.
05:37And that could be extremely valuable for all of this, all the opportunities that exist in Asia, for example, with
05:44aging population.
05:44I think there's a huge opportunity for robotics there.
05:47I think there's something that's going to be something that's going to be very important.
Comments