Building Self-Learning Robots For Hazardous Real-World Missions

WEBVTT
00:00:01.520 --> 00:00:02.479
Hey everyone.
00:00:02.720 --> 00:00:13.359
Really intriguing discussion today on how AI can learn and adapt in a real world environment with a true innovator in the space at Skylark Labs.
00:00:13.519 --> 00:00:14.800
Damarjat, how are you?
00:00:15.199 --> 00:00:15.839
Thank you, sir.
00:00:16.000 --> 00:00:16.719
How are you doing?
00:00:17.199 --> 00:00:18.000
I'm doing well.
00:00:18.079 --> 00:00:19.280
Thanks for joining.
00:00:19.519 --> 00:00:24.399
You have a really intriguing and distinguished background at Cambridge and Stanford.
00:00:24.640 --> 00:00:27.600
So really excited to hear what you have to say before that.
00:00:27.839 --> 00:00:29.920
Tell us more about Skylark Labs.
00:00:30.160 --> 00:00:30.879
What do you do?
00:00:30.960 --> 00:00:33.439
And what's the mission behind the company?
00:00:33.840 --> 00:00:35.920
Yeah, thank you for uh having me, Evan.
00:00:36.240 --> 00:00:59.920
So at Skylark Labs, we are building the next evolution of self-learning models for different kinds of machines, which will allow these machines to go into mainly hazardous environments and complete different kinds of tasks on their own independently without having a human in the loop.
00:01:00.320 --> 00:01:18.079
And not just do the tasks well, which they know of, but if there are new tasks, unknown conditions arise similar to humans, they should be able to then and there learn those new kinds of conditions and keep on learning and operating reliably.
00:01:18.239 --> 00:01:20.079
So that's the work we're doing.
00:01:20.959 --> 00:01:21.599
Fantastic.
00:01:21.760 --> 00:01:28.239
Of course, the buzzword is physical AI, which is generating a lot of enthusiasm, excitement.
00:01:28.400 --> 00:01:34.799
But how do you describe physical AI and your position in this world, so to speak?
00:01:35.280 --> 00:01:49.519
Yeah, I think physical AI, the way I think about it is giving end-to-end autonomy to machines so that given a vowel or maybe on their own, they are able to complete certain mission sets.
00:01:49.680 --> 00:02:02.159
And if you look at the current physical AI space, there's a lot of effort, a lot of energy, money, capital going into developing these brains for machines, which allows them to do these tasks.
00:02:02.319 --> 00:02:11.520
And main focus has been building a butler for homes so that it can make your life easy or primarily manufacturing.
00:02:11.680 --> 00:02:19.199
And in the case of homes, you're learning the skills of how to grab a plate, how to clean the floor.
00:02:19.360 --> 00:02:32.000
In the manufacturing side, you're figuring out how to move things around, how to use a screwdriver, while we focus specifically on outdoor environments, specifically of defense and public safety.
00:02:32.159 --> 00:02:51.599
So our machines know how to handle a metal detector or what to do when you get a peep so that it knows that there's a potential hazard or a bomb, what you should do, how quickly you should move away, or how to load, lead, load, reload weapons, how to move things around.
00:02:51.840 --> 00:02:53.919
So that's the world we live in.
00:02:54.159 --> 00:03:09.280
And we are trying to focus on the humanoid form factor now, although we can do other things as well, primarily because humanoid is one form factor where you can load many mission sets into that, leading to a machine which can do many things together.
00:03:10.719 --> 00:03:11.280
Interesting.
00:03:11.439 --> 00:03:14.719
So your technology can keep learning after it's deployed.
00:03:15.120 --> 00:03:17.919
Run through a scenario or an anecdote, if you would.
00:03:18.000 --> 00:03:21.120
How does it work and why is it so important?
00:03:21.599 --> 00:03:32.319
Yeah, so the most of the work we focus on very mission critical sets, very environment, difficult environments where it was not possible to do that with traditional environments.
00:03:32.479 --> 00:03:38.000
One example is picking very small debris on a military tarmac.
00:03:38.240 --> 00:03:45.759
And once the fighter jets get on, F-22s, F-35s, the engine has great suction capacity.
00:03:46.000 --> 00:03:51.439
So if you have any small debris on the tarmac, your engine would get damaged.
00:03:51.680 --> 00:03:54.719
So, and these are three kilometer long tarmacs.
00:03:54.800 --> 00:04:00.879
So people would normally walk on these to pick debris which is as small as one centimeter.
00:04:01.120 --> 00:04:07.759
And obviously you would miss things, and that leads to damage worth hundreds of millions of dollars to a single jet.
00:04:07.919 --> 00:04:18.720
So the work we did initially was to build these cars which could run up and down the tarmac, get detections, and someone would get out the car to pick this debris.
00:04:18.959 --> 00:04:28.399
Now, self-learning over there played a role that you had new kind of debris which would come up all the time because basically this is just garbage on a tarmac.
00:04:29.040 --> 00:04:36.319
Or you would have some sort of info perfections in the tarmac, which the system can sometimes flag as debris.
00:04:36.480 --> 00:04:43.759
So the self-learning thing was let's make it reliable enough so that there are no false positives or very minimum false positives.
00:04:43.920 --> 00:04:50.160
And if there are new kinds of debris, you can learn it then and there rather than having to go back to the designer.
00:04:50.480 --> 00:05:02.720
And now we are replacing the car with a robotic dog so that the same model can scan but can also collect the debris so that there's no human in the loop.
00:05:02.959 --> 00:05:10.160
But now this can not just be on the tarmac, but also in the hangover areas where they do the engine testing.
00:05:10.319 --> 00:05:12.959
So the mission set it can do is bigger.
00:05:13.120 --> 00:05:21.120
But when you're not using it for scanning debris, you can scan it for panimated monitoring, you can send it to other mission sets.
00:05:21.360 --> 00:05:27.279
So not just it can do what it was doing previously manually, it has automated that.
00:05:27.439 --> 00:05:35.759
It has increased its scope, but also single robot can now do multiple mission sets leading to multiple folds of efficiency.
00:05:36.800 --> 00:05:37.519
Fantastic.
00:05:37.680 --> 00:05:43.519
And your technology runs directly on these various devices rather than relying on the cloud.
00:05:43.600 --> 00:05:46.800
That must have been uh you know quite a challenge as well.
00:05:47.600 --> 00:05:58.959
Yes, in fact, uh here is one of our edge devices in my hand, and you can connect different sensors using this, and this goes on the cars.
00:05:59.120 --> 00:06:08.160
We have smaller versions of this which go on the robotic dogs, humanoids, and all the learning happens here itself, so there is no connection to the cloud.
00:06:08.319 --> 00:06:21.600
So a lot of our initial effort was how do we actually print the models, make them small enough so they can fit on these kind of devices, can continuously learn and keep on doing it.
00:06:21.680 --> 00:06:24.079
So a lot of our research has focused on that.
00:06:25.120 --> 00:06:25.920
Fantastic.
00:06:26.160 --> 00:06:27.040
Let's talk drones.
00:06:27.199 --> 00:06:29.360
I'm a drone enthusiast myself.
00:06:29.600 --> 00:06:39.680
They're getting pretty smart, and in the commercial world, they're even smarter, of course, being used autonomously in the war in Ukraine and and elsewhere.
00:06:40.000 --> 00:06:41.040
But where are we headed?
00:06:41.199 --> 00:06:47.680
How smart and autonomous can these drones, UAVs, kind of become?
00:06:48.160 --> 00:06:56.560
Yeah, I think like the 2018-19 world of drones primarily focused on just perception.
00:06:56.720 --> 00:07:04.399
So you will get a drone, you'll get a controller, someone flies it, it's one-to-one, and you manually see whatever you can see.
00:07:04.800 --> 00:07:09.759
And we have automated perception part of those drones quite a bit.
00:07:10.000 --> 00:07:28.000
Where now on the sensor, you can put models which can detect people from very long distances, it can count people in large music festivals, political rallies, all the ways to detecting cracks on a solar panel to cracks on, let's say, a furnace.
00:07:28.160 --> 00:07:37.439
So we have deployed them all the way from public safety to defense to critical infrastructure, meaning detecting again very small signals.
00:07:37.600 --> 00:07:39.600
And you have the same problem here as well.
00:07:39.759 --> 00:07:45.120
You would get a lot of false positives, or you would miss these new kinds of signals which have come up.
00:07:45.279 --> 00:07:50.319
So the self-learning capability has been quite beneficial here as well.
00:07:50.560 --> 00:07:58.639
But now, these drones, we are also putting robotic arms on that, small arms with hands or claws.
00:07:58.800 --> 00:08:08.079
So now it is not just detecting a crack on the solar panel, it is also spraying water on it to clean it because they're in the middle of the desert sometimes.
00:08:08.399 --> 00:08:12.319
And if they get dust on it, the efficiency falls quite a bit.
00:08:12.480 --> 00:08:19.040
Or if you get a detect uh crack on uh on a furnace, you can go and put a quick sealant.
00:08:19.439 --> 00:08:25.120
Same goes for detecting some sort of defects on insulators and then cooling them down.
00:08:25.279 --> 00:08:36.320
So there, even on the drone side now, we have moved quite a bit from autonomous perception-only operations to autonomous end-to-end autonomy, which can also be some sort of actions.
00:08:37.120 --> 00:08:37.600
Amazing.
00:08:37.679 --> 00:08:43.200
And what's what's the most difficult part about getting AI to understand and react to the world around it?
00:08:43.360 --> 00:08:51.519
You make it sound so simple, but obviously there's a a lot of work, a lot of intelligence, one intended, to make it happen.
00:08:52.000 --> 00:08:57.919
Yeah, I think like on the perception side, the biggest challenge is to main shift.
00:08:58.159 --> 00:09:03.360
You can get things which look like a human in the middle of a tree, so false positive problem.
00:09:03.519 --> 00:09:05.039
So you need to figure that out.
00:09:05.279 --> 00:09:18.639
But now the robotic side, the control side, when you have an arm which is touching, let's say, the side of a furnace, if not totally spraying it, that is producing a force on the drone.
00:09:18.799 --> 00:09:30.960
So you need to put AI, which is not just managing how to fly the drone, but also now to account for this extra force which is coming when you're interacting with the environment.
00:09:31.120 --> 00:09:37.679
So you need to have an AI model or a control policy which can jointly model all those things.
00:09:37.919 --> 00:09:42.399
Also, we are flying a lot of these drones very fast in the middle of the desert.
00:09:42.480 --> 00:09:44.399
Sometimes they can be very high winds.
00:09:44.480 --> 00:09:48.480
So your control policy needs to account for those kind of things.
00:09:48.639 --> 00:09:52.720
And that is basically the specialization we have.
00:09:53.039 --> 00:09:57.759
But a lot of these environments can also be unguided or uncharted.
00:09:58.000 --> 00:10:04.879
For example, just the day before yesterday, we were trying to do something near a furnace, but they had scaffolding.
00:10:05.039 --> 00:10:13.679
So now the drone has to maneuver around the scaffolding, which it did not see in the in the training sets before, and that is where the adaption comes in.
00:10:13.840 --> 00:10:24.480
So I think it goes all the way from modeling different uh flying dynamics together all the way to accounting for new kind of variations.
00:10:25.600 --> 00:10:26.320
Interesting.
00:10:26.399 --> 00:10:31.440
And so there is this big push for adoption, physical AI, and lots of great use cases.
00:10:31.600 --> 00:10:33.519
What's holding the industry back?
00:10:33.759 --> 00:10:35.039
Is it still the technology?
00:10:35.200 --> 00:10:41.759
Is the cost coming down, or is it trust regulations, or maybe some mixes, all of them?
00:10:42.639 --> 00:10:53.919
I think it's a function of the end-to-end technology not being there right now, or mainly being there in a lot of the experimental settings.
00:10:54.159 --> 00:11:08.240
I think if you look at the autonomy for the drone use case, which we just mentioned, you would need to have models which are continuously learning in the field for new kinds of conditions.
00:11:08.480 --> 00:11:17.440
So I think technology which can learn on the fly, on edge, can adapt as a big bottleneck because new things are gonna come all the time.
00:11:17.600 --> 00:11:22.480
And a lot of the things which we had before kind of work well in very specific settings.
00:11:22.559 --> 00:11:26.320
So I think that was a challenge, or it need to have a human loop.
00:11:26.559 --> 00:11:35.440
The other thing is if you want to build this into an autonomy, data is a big challenge, and you can get data for some settings and some variations, but new variations come up.
00:11:35.600 --> 00:11:38.000
So I think data is a big bottleneck.
00:11:38.240 --> 00:11:43.039
And at least on the humanoid side, the other big bottleneck is hardware.
00:11:43.279 --> 00:11:49.200
You need many specialized hardware which can work in the field, it should be waterproof.
00:11:49.279 --> 00:11:53.279
The battery should not catch fire for a lot of our oil and gas use cases.
00:11:53.519 --> 00:11:59.679
So a lot of the humanoids we have, they're not waterproof, they don't have a battery which has these kind of capacities.
00:11:59.919 --> 00:12:04.080
Also, the hands, they need to be able to lift heavy stuff.
00:12:04.320 --> 00:12:11.840
But right now, the capacity with which they can lift stuff is maybe like eight pounds, ten pounds, not more than that.
00:12:12.000 --> 00:12:21.519
So I think at least on the drone side where the hardware is commercialized, you need an intelligent brain.
00:12:21.759 --> 00:12:26.639
But on the other side, where even hardware is not there, that's a bigger challenge.
00:12:27.519 --> 00:12:28.159
Interesting.
00:12:28.320 --> 00:12:30.080
Well, a lot of work ahead for you.
00:12:30.240 --> 00:12:33.840
What are you excited about short term at Skylark Labs?
00:12:34.080 --> 00:12:36.960
Are there any particular insights you can share?
00:12:37.039 --> 00:12:40.080
Uh, big things that are on your radar the next few months.
00:12:40.799 --> 00:12:49.840
Yeah, I think like the first thing that you are working on is improving the AI brain or the physical AI models.
00:12:49.919 --> 00:13:01.919
So we just recently released our continual field adaptive models C FAMS paper, and that's the model class which we have been building over the past three to four years.
00:13:02.080 --> 00:13:06.320
And we published that paper with Carnegie Mellon and UC Berkeley.
00:13:06.480 --> 00:13:12.799
So that is out, so that kind of starts to show the world as to what are the kind of things we have been building.
00:13:13.039 --> 00:13:26.799
But that same model class now is being extended to new kind of novel signatures, deploying them on humanoids, robotic dogs, and also showing those technologies to the end customers.
00:13:26.960 --> 00:13:46.320
So we are working with the US Army, Special Operations Command to build mission sets of them of humanoids breaching into buildings so that they are in danger rather than a human being in danger, or for bomb disposal applications of IED.
00:13:46.559 --> 00:14:01.200
So those are the kind of things we are excited about where we are improving our model class, building it for specific mission sets with specific customers in mind, and then scaling that further.
00:14:02.080 --> 00:14:04.000
Wow, super interesting.
00:14:04.240 --> 00:14:05.360
What about longer term?
00:14:05.440 --> 00:14:07.679
It's it's tough to have a crystal ball in this space.
00:14:07.759 --> 00:14:09.519
Everything is moving so quickly.
00:14:09.679 --> 00:14:13.840
I daren't ask you about the next few years, because who knows?
00:14:14.000 --> 00:14:20.639
But can you look ahead one, two, three years, and tell us what you're hoping to see in terms of breakthroughs and innovation?
00:14:21.279 --> 00:14:36.559
Yeah, I think the first thing at least we are trying to do is to become a hardware humanoid form factor so that it is able to sustain the ruggedized environment of minefields or oil and gas.
00:14:36.720 --> 00:14:47.360
So we are working on building our own humanoid, which has those properties of waterproofing, as I mentioned, or the battery being more fireproof.
00:14:47.600 --> 00:14:53.360
We are building our own humanoid hand so that it is able to pick heavy things.
00:14:53.519 --> 00:15:07.120
And if we are able to achieve that, we will get to a point so that we start to become the category leader for mission-critical humanoids, which is able to go and then learn mission critical skills, which other people are not doing.
00:15:07.200 --> 00:15:11.759
So that is something we wanted to go over the next one year or so.
00:15:12.000 --> 00:15:19.200
But then after that, we are trying to layer our humanoid with a brain so that it can have many skills.
00:15:19.440 --> 00:15:32.320
Skills of bomb disposal to parameter securities, to picking luggage, to firing weapons, to loading weapons, ultimately, hopefully, getting to a point so that you don't need different kinds of robotic systems.
00:15:32.559 --> 00:15:37.120
This one robotic platform will be able to do everything.
00:15:37.279 --> 00:15:39.679
And that is the other thing we want to do.
00:15:39.919 --> 00:15:44.799
And then finally, we want to insert them into groups.
00:15:45.120 --> 00:16:09.360
We are teaching a lot of group dynamics where the machine is able to work alongside a human, learn from them, teach other humanoids, hopefully getting to a world where the machine starts to acquire all the skills of humans, bringing digital life into the real world because of its adaptive capabilities.
00:16:09.519 --> 00:16:11.759
And that would be a very interesting world.
00:16:12.159 --> 00:16:13.519
An interesting world indeed.
00:16:13.679 --> 00:16:16.159
Well, congratulations on the amazing work.
00:16:16.559 --> 00:16:22.799
Really incredible use cases and opportunities for good and to save lives and more.
00:16:23.120 --> 00:16:24.240
Thanks for joining.
00:16:24.639 --> 00:16:26.000
Thank you very much for having me.
00:16:26.320 --> 00:16:28.080
And thanks everyone for listening and watching.
00:16:28.240 --> 00:16:32.879
Also, check out our TV show, Tech Impact TV, on Bloomberg and Fox Business.
00:16:33.039 --> 00:16:33.519
Thanks, everyone.
00:16:33.679 --> 00:16:34.799
Thanks, Hammer Jan.
00:16:35.039 --> 00:16:35.679
Thank you.