WEBVTT 00:00:00.000 --> 00:00:02.359 You know, when most people picture a robot doing 00:00:02.359 --> 00:00:06.960 its job, they usually picture a cage. Right, 00:00:07.099 --> 00:00:10.019 like a heavy plexiglass or steel wire enclosure 00:00:10.019 --> 00:00:12.759 that's just bolted to the concrete of a factory 00:00:12.759 --> 00:00:15.599 floor. Exactly. Sparks are flying everywhere 00:00:15.599 --> 00:00:18.739 and there's this massive, multi -million dollar 00:00:18.739 --> 00:00:21.420 mechanical arm performing the exact same weld 00:00:21.420 --> 00:00:24.230 on a car chassis, like... thousands of times 00:00:24.230 --> 00:00:27.070 a day. Yeah, it is the absolute definition of 00:00:27.070 --> 00:00:29.210 precision. It is. It's completely controlled 00:00:29.210 --> 00:00:31.410 and completely predictable. But the trade -off 00:00:31.410 --> 00:00:33.590 for all that perfect precision is that the robot 00:00:33.590 --> 00:00:36.070 is also completely blind to the world outside 00:00:36.070 --> 00:00:38.130 that cage. Right, because it doesn't actually 00:00:38.130 --> 00:00:40.090 know what it's doing. Exactly. If you move that 00:00:40.090 --> 00:00:42.189 car chassis, even, I mean, half an inch to the 00:00:42.189 --> 00:00:44.649 left on the assembly line... The robot doesn't 00:00:44.649 --> 00:00:47.850 know. It doesn't adapt. It just blindly performs 00:00:47.850 --> 00:00:50.189 its perfect weld on the empty air. But then you 00:00:50.189 --> 00:00:52.729 take that robot out of its cage, you strip away 00:00:52.729 --> 00:00:55.229 its perfectly timed step -by -step instructions, 00:00:55.490 --> 00:00:57.950 and you drop it in the middle of a chaotic, rainy 00:00:57.950 --> 00:01:00.770 city sidewalk. Oh, yeah. That illusion of mechanical 00:01:00.770 --> 00:01:03.100 perfection just shatters immediately. Totally. 00:01:03.259 --> 00:01:05.579 We are dealing with an entirely different beast 00:01:05.579 --> 00:01:07.819 at that point. This is a totally different paradigm 00:01:07.819 --> 00:01:10.319 for engineering. You are moving away from mere 00:01:10.319 --> 00:01:13.280 automation and stepping into, well, true autonomy. 00:01:13.519 --> 00:01:15.519 And that is exactly what we are getting into 00:01:15.519 --> 00:01:17.959 today. Welcome back to the Deep Dive. Glad to 00:01:17.959 --> 00:01:19.870 be here for this one. If you were joining us, 00:01:19.950 --> 00:01:22.230 you are probably a curious mind who wants to 00:01:22.230 --> 00:01:25.409 cut through all the sci -fi movie noise and understand 00:01:25.409 --> 00:01:28.129 what is actually happening right now in the world 00:01:28.129 --> 00:01:30.730 of autonomous robotics. Because there's a lot 00:01:30.730 --> 00:01:32.689 of noise out there. So much noise. You want to 00:01:32.689 --> 00:01:35.390 know the real mechanics, the real failures and 00:01:35.390 --> 00:01:37.870 the real trajectory of these machines. And we 00:01:37.870 --> 00:01:40.829 have a fantastic, really dense foundation for 00:01:40.829 --> 00:01:43.209 this today. We do. We are pulling exclusively 00:01:43.209 --> 00:01:45.590 from a comprehensive Wikipedia article simply 00:01:45.590 --> 00:01:49.430 titled Autonomous Robot. massive repository. 00:01:49.430 --> 00:01:52.489 It covers technical history, engineering criteria, 00:01:53.590 --> 00:01:57.109 and the societal impacts of these systems. And 00:01:57.109 --> 00:01:59.750 our mission today is to take all that technical 00:01:59.750 --> 00:02:02.409 depth and distill it down into the essential 00:02:02.409 --> 00:02:05.260 aha moments for you. We are going to look at 00:02:05.260 --> 00:02:07.620 exactly how these machines process the world, 00:02:07.859 --> 00:02:10.340 why the real world keeps breaking them, and how 00:02:10.340 --> 00:02:12.419 they are already quietly migrating into your 00:02:12.419 --> 00:02:14.400 neighborhood. It's happening faster than people 00:02:14.400 --> 00:02:16.939 realize. It really is. Okay, let's unpack this. 00:02:17.439 --> 00:02:21.039 True robotic autonomy. Our material defines an 00:02:21.039 --> 00:02:23.639 autonomous robot as one that acts... without 00:02:23.639 --> 00:02:26.780 recourse to human control. Right. So as we just 00:02:26.780 --> 00:02:29.879 established, that goes way beyond a factory arm 00:02:29.879 --> 00:02:33.180 following a script. Way beyond. Inductory alarms 00:02:33.180 --> 00:02:35.780 have their autonomy heavily restricted precisely 00:02:35.780 --> 00:02:38.139 because their environment is completely structured. 00:02:38.219 --> 00:02:40.219 They don't have to figure anything out. Exactly. 00:02:40.419 --> 00:02:42.460 And crucially, they don't locomote. They don't 00:02:42.460 --> 00:02:45.340 move from point A to point B. True autonomy means 00:02:45.340 --> 00:02:47.800 operating independently in the total chaos of 00:02:47.800 --> 00:02:50.560 the real world. Right. And to understand how 00:02:50.560 --> 00:02:53.159 a machine can act without a human whole a joystick, 00:02:53.460 --> 00:02:56.080 we first have to understand how it mimics basic 00:02:56.080 --> 00:02:58.919 biological survival. Yeah, because before a robot 00:02:58.919 --> 00:03:01.719 can deliver your groceries or weld a pipe, it 00:03:01.719 --> 00:03:05.000 has to be able to simply, you know, take care 00:03:05.000 --> 00:03:07.800 of itself. It's the very first requirement for 00:03:07.800 --> 00:03:11.000 complete physical autonomy self -maintenance. 00:03:11.520 --> 00:03:13.800 And this relies on a biological concept called 00:03:13.800 --> 00:03:17.240 proprioception. Proprioception? Yeah. In human 00:03:17.240 --> 00:03:19.740 biology, proprioception is how you know exactly 00:03:19.740 --> 00:03:22.080 where your hand is in a dark room without having 00:03:22.080 --> 00:03:23.780 to turn on the light and look at it. You just 00:03:23.780 --> 00:03:25.939 feel your own internal state. Right. You just 00:03:25.939 --> 00:03:27.719 have that spatial awareness of your own body. 00:03:27.979 --> 00:03:31.699 Exactly. In robotics, proprioception is a machine 00:03:31.699 --> 00:03:34.300 sensing its own internal status. So what does 00:03:34.300 --> 00:03:36.800 that actually look like inside a machine? How 00:03:36.800 --> 00:03:39.280 does a robot feel its own body? Well, a great 00:03:39.280 --> 00:03:41.840 example from our research is Sony's robotic dog, 00:03:42.060 --> 00:03:44.500 Aibo. Oh, the little silver one. Yeah, exactly. 00:03:44.560 --> 00:03:47.319 It is capable of self -docking to charge its 00:03:47.319 --> 00:03:50.319 batteries, but it's not on a preset timer. Wait, 00:03:50.479 --> 00:03:52.800 really? It just decides when to charge? Yes. 00:03:53.219 --> 00:03:55.900 The robot can tell, proprioceptively, that its 00:03:55.900 --> 00:03:58.419 internal voltage is dropping. It basically feels 00:03:58.419 --> 00:04:01.039 the digital equivalent of hunger, so it actively 00:04:01.039 --> 00:04:03.580 seeks out the charger on its own. That is wild. 00:04:03.759 --> 00:04:05.520 And it goes deeper than just batteries, too. 00:04:05.789 --> 00:04:08.610 It involves thermal sensors monitoring the heat 00:04:08.610 --> 00:04:11.129 of the leg joints so the motors don't burn themselves 00:04:11.129 --> 00:04:13.550 out. Wow! So it knows if it's over -exerting 00:04:13.550 --> 00:04:16.089 itself. Right. And it uses Hall effect sensors 00:04:16.089 --> 00:04:18.649 to monitor electrical currents, making sure a 00:04:18.649 --> 00:04:20.889 limb isn't being pushed past its physical breaking 00:04:20.889 --> 00:04:24.139 point. the robot is constantly checking its own 00:04:24.139 --> 00:04:27.360 vitals. So it's this closed loop of self -awareness. 00:04:27.560 --> 00:04:29.319 But just knowing you are hungry or your joints 00:04:29.319 --> 00:04:31.500 are hot isn't enough to actually survive. You 00:04:31.500 --> 00:04:34.040 have to understand the world outside your body 00:04:34.040 --> 00:04:36.540 to do anything about it. Precisely. And that 00:04:36.540 --> 00:04:39.420 brings us to the other side of the coin, exteroception. 00:04:39.740 --> 00:04:41.680 Exteroception, OK. This is the machine sensing 00:04:41.680 --> 00:04:44.800 the environment around it. We are talking about 00:04:44.800 --> 00:04:47.399 electromagnetic spectrum sensors, sound detection, 00:04:47.980 --> 00:04:50.540 chemical odor analysis. temperature, altitude. 00:04:50.720 --> 00:04:52.860 There are some surprisingly relatable examples 00:04:52.860 --> 00:04:55.759 of this in the article. Take robotic lawn mowers. 00:04:55.879 --> 00:04:58.259 Right. They don't just drive in a random pattern 00:04:58.259 --> 00:05:01.180 for an hour. They can actually adapt their programming 00:05:01.180 --> 00:05:03.740 by detecting the physical resistance on their 00:05:03.740 --> 00:05:06.600 cutting blades, which tells them the speed at 00:05:06.600 --> 00:05:09.079 which the grass is growing. That is such a clever 00:05:09.079 --> 00:05:12.040 engineering trick. I know. Or think about high 00:05:12.040 --> 00:05:15.279 end autonomous vacuum cleaners. They use built 00:05:15.279 --> 00:05:18.060 in dirt detectors, often acoustic sensors that 00:05:18.060 --> 00:05:20.790 literally listen to the sound of particles hitting 00:05:20.790 --> 00:05:23.589 the intake. So it can hear the mess. Yeah. They 00:05:23.589 --> 00:05:25.629 sense how much dirt is coming up and use that 00:05:25.629 --> 00:05:28.110 real -time data to tell themselves to stop, turn 00:05:28.110 --> 00:05:30.490 around, and stay in one messy spot a little longer. 00:05:30.769 --> 00:05:33.050 Exactly. It's adjusting its behavior dynamically 00:05:33.050 --> 00:05:36.310 based on environmental feedback rather than just 00:05:36.310 --> 00:05:39.600 executing a blind pre -programmed routine. So 00:05:39.600 --> 00:05:41.920 proprioception is like my stomach rumbling and 00:05:41.920 --> 00:05:44.500 knowing I need a snack, while exteroception is 00:05:44.500 --> 00:05:47.120 knowing not to walk into a wall on my way to 00:05:47.120 --> 00:05:49.600 the fridge. That is a perfect framework for it. 00:05:49.689 --> 00:05:52.170 What's fascinating here is that without both 00:05:52.170 --> 00:05:55.350 of those layers working perfectly together, robust 00:05:55.350 --> 00:05:57.949 internal sensing and robust external sensing, 00:05:58.209 --> 00:06:01.149 higher level task performance is entirely impossible. 00:06:01.350 --> 00:06:03.430 So you need both. You absolutely have to have 00:06:03.430 --> 00:06:05.970 both. Think about the Amazon Astro, which launched 00:06:05.970 --> 00:06:08.689 back in 2021 for home monitoring and security. 00:06:08.850 --> 00:06:12.029 Right, the little rolling screen guy. Yeah. That 00:06:12.029 --> 00:06:14.910 robot performs what we call conditional tasks, 00:06:15.449 --> 00:06:18.899 meaning its response entirely depends on a complex 00:06:18.899 --> 00:06:22.040 logitry. If it detects an intruder, it does X. 00:06:22.339 --> 00:06:24.980 If it detects the homeowner, it does Y. Makes 00:06:24.980 --> 00:06:28.699 sense. But a machine absolutely cannot monitor 00:06:28.699 --> 00:06:31.379 our safety or execute a conditional security 00:06:31.379 --> 00:06:34.600 protocol if it doesn't first possess the proprioception 00:06:34.600 --> 00:06:37.660 to know its own physical limits and the exteroception 00:06:37.660 --> 00:06:40.310 to map out the hallways of your house. It has 00:06:40.310 --> 00:06:42.509 to fully know itself before it can begin to help 00:06:42.509 --> 00:06:45.430 us. Exactly. So once a robot can sense its body 00:06:45.430 --> 00:06:48.129 and its environment, the immediate next hurdle 00:06:48.129 --> 00:06:50.149 is physically moving through that environment. 00:06:50.509 --> 00:06:53.009 And moving through a sterile, flat laboratory 00:06:53.009 --> 00:06:55.490 is very, very different from navigating the actual 00:06:55.490 --> 00:06:58.410 world. Oh, absolutely. The jump from indoor navigation 00:06:58.410 --> 00:07:00.870 to outdoor navigation is arguably one of the 00:07:00.870 --> 00:07:02.870 biggest dividing lines in the entire field of 00:07:02.870 --> 00:07:04.910 robotics. Let's start with the indoors, because 00:07:04.910 --> 00:07:07.370 the history of how we solved this is fascinating. 00:07:07.870 --> 00:07:10.310 Indoor navigation started back in the 1970s with 00:07:10.310 --> 00:07:12.329 simple wire guidance. Yeah, the early days. It 00:07:12.329 --> 00:07:14.889 was basically a robot following a magnetic wire 00:07:14.889 --> 00:07:17.110 buried in the floor, almost like a slot car track. 00:07:17.329 --> 00:07:20.870 It was reliable, but incredibly rigid. If you 00:07:20.870 --> 00:07:23.250 wanted the robot to go somewhere new, you had 00:07:23.250 --> 00:07:25.670 to literally tear up the concrete and lay a new 00:07:25.670 --> 00:07:28.790 wire. But by the early 2000s, the technology 00:07:28.790 --> 00:07:32.209 evolved dramatically. You had robots like Mobile 00:07:32.209 --> 00:07:35.449 Robots Patrol Bot in 2004, which introduced dynamic 00:07:35.449 --> 00:07:37.680 mapping. Right, that was a huge leap. It could 00:07:37.680 --> 00:07:39.980 roll into an unknown building and create its 00:07:39.980 --> 00:07:42.379 own laser -based map of the environment on the 00:07:42.379 --> 00:07:45.959 fly. If someone left a box in its path, its control 00:07:45.959 --> 00:07:49.180 system didn't just stop and give up. It recalculated 00:07:49.180 --> 00:07:52.019 a new route around the box instantly. And today, 00:07:52.699 --> 00:07:55.439 advanced indoor systems fuse data from LIDAR, 00:07:55.800 --> 00:07:58.319 cameras, and wheel encoders to pinpoint their 00:07:58.319 --> 00:08:00.959 exact location within millimeters. And what I 00:08:00.959 --> 00:08:03.379 found so brilliant about indoor navigation is 00:08:03.379 --> 00:08:06.800 how engineers handle complex obstacles like multistore 00:08:06.540 --> 00:08:09.519 Oh, the elevator problem. Exactly. Instead of 00:08:09.519 --> 00:08:11.939 building insanely complicated top -heavy hardware 00:08:11.939 --> 00:08:14.639 with mechanical legs to climb stairs, which is 00:08:14.639 --> 00:08:16.620 an engineering nightmare, most indoor robots 00:08:16.620 --> 00:08:18.959 just use handicapped accessible infrastructure. 00:08:19.199 --> 00:08:21.300 Yeah, they use wireless interfaces to communicate 00:08:21.300 --> 00:08:23.839 with the building's own electronics. Right, telling 00:08:23.839 --> 00:08:26.560 the automatic doors to open or commanding the 00:08:26.560 --> 00:08:28.899 elevator to go to the third floor. They adapt 00:08:28.899 --> 00:08:31.339 the building to the robot rather than the robot 00:08:31.339 --> 00:08:34.340 to the building. It's an elegant, highly efficient 00:08:34.340 --> 00:08:36.990 workaround. But the moment you take that robot 00:08:36.990 --> 00:08:39.809 out of the building and put it outdoors, the 00:08:39.809 --> 00:08:42.590 complexity of navigation just skyrockets. It's 00:08:42.590 --> 00:08:44.950 a whole different ballgame. Interestingly though, 00:08:45.230 --> 00:08:47.250 outdoor autonomy is actually easiest in the sky. 00:08:47.580 --> 00:08:50.340 Obstacles are incredibly rare at 10 ,000 feet. 00:08:50.740 --> 00:08:53.980 Right. The documentation details pilotless drone 00:08:53.980 --> 00:08:56.620 aircraft executing entire reconnaissance missions 00:08:56.620 --> 00:09:00.399 autonomously. We also see SpaceX operating massive 00:09:00.399 --> 00:09:03.720 autonomous spaceport drone ships. Those are incredible 00:09:03.720 --> 00:09:05.759 feats of engineering. They navigate the ocean 00:09:05.759 --> 00:09:08.139 and dynamically position themselves to safely 00:09:08.139 --> 00:09:11.559 catch Falcon 9 rockets falling from space. The 00:09:11.559 --> 00:09:14.080 ocean surface and the sky are largely empty space. 00:09:14.600 --> 00:09:17.159 But ground navigation outdoors, that is an entire 00:09:17.039 --> 00:09:20.919 It is undeniably the most difficult domain for 00:09:20.919 --> 00:09:22.679 vehicles. You are dealing with three -dimensional 00:09:22.679 --> 00:09:25.340 terrain, massive disparities in surface friction, 00:09:25.840 --> 00:09:27.879 like transitioning smoothly from dry concrete 00:09:27.879 --> 00:09:31.519 to wet mud. And weather. Exactly. Unpredictable 00:09:31.519 --> 00:09:34.460 weather patterns, that blind sensors, and just 00:09:34.460 --> 00:09:37.980 general constant instability. Okay, let me push 00:09:37.980 --> 00:09:39.539 back on this for a second, because I think a 00:09:39.539 --> 00:09:41.440 lot of people listening might be... wondering 00:09:41.440 --> 00:09:43.860 this exact thing. Sure, go for it. If my smart 00:09:43.860 --> 00:09:46.419 vacuum can perfectly map my living room and avoid 00:09:46.419 --> 00:09:49.779 the stairs, why is it so incredibly difficult 00:09:49.779 --> 00:09:53.059 for a robot to drive across my bumpy sunlit backyard? 00:09:53.360 --> 00:09:55.820 That is the million dollar question in the industry. 00:09:56.029 --> 00:09:58.590 If we connect this to the bigger picture, robotics 00:09:58.590 --> 00:10:01.110 is what we call an inherently systems problem. 00:10:01.210 --> 00:10:03.870 A systems problem. Yeah. That means a robot isn't 00:10:03.870 --> 00:10:07.350 just one smart brain. It is a chain of interconnected 00:10:07.350 --> 00:10:09.830 modules. You have the perception module, the 00:10:09.830 --> 00:10:12.049 planning module, the actuation module controlling 00:10:12.049 --> 00:10:14.070 the motors. Got it. I'm tracking. If just one 00:10:14.070 --> 00:10:16.649 of those modules fails, the entire machine is 00:10:16.649 --> 00:10:18.950 paralyzed. So if the camera gets confused, the 00:10:18.950 --> 00:10:22.009 wheels stop turning. Exactly. Our research highlights 00:10:22.009 --> 00:10:24.570 the massive challenge of systemic robustness 00:10:24.570 --> 00:10:28.070 versus real -world brittleness. Real -world brittleness. 00:10:28.269 --> 00:10:31.070 That's a great term. A sudden, harsh beam of 00:10:31.070 --> 00:10:33.470 sunlight hitting a camera lens can completely 00:10:33.470 --> 00:10:36.669 overexpose the image, temporarily blinding the 00:10:36.669 --> 00:10:39.879 vision system. Or a patch of unexpected wet leaves 00:10:39.879 --> 00:10:42.620 can cause the wheels to slip, ruining the actuation 00:10:42.620 --> 00:10:45.440 math. Which leads to the massive open problem 00:10:45.440 --> 00:10:48.820 of open world scene understanding. Yes. And the 00:10:48.820 --> 00:10:50.399 root of that problem is something called the 00:10:50.399 --> 00:10:53.620 reality gap, right? This concept blew my mind. 00:10:54.080 --> 00:10:56.379 It is foundational to understanding why robots 00:10:56.379 --> 00:10:59.299 fail. Researchers today use deep reinforcement 00:10:59.299 --> 00:11:01.899 learning to teach robots how to walk or drive. 00:11:02.000 --> 00:11:04.039 By running them through millions of simulated 00:11:04.039 --> 00:11:06.919 environments on a computer. Right. But a simulation, 00:11:07.019 --> 00:11:09.879 no matter how good, is perfectly logical and 00:11:09.879 --> 00:11:12.600 bounded. The real world is not. There's just 00:11:12.600 --> 00:11:15.639 too much chaos. When you take an artificial brain... 00:11:15.230 --> 00:11:17.350 that has been trained in a pristine computer 00:11:17.350 --> 00:11:19.610 simulation and downloaded into a physical body 00:11:19.610 --> 00:11:23.070 outside. The tiny chaotic discrepancies between 00:11:23.070 --> 00:11:26.950 the simulation and that reality gap caused the 00:11:26.950 --> 00:11:29.090 logic systems to collapse. I was trying to picture 00:11:29.090 --> 00:11:30.970 what that feels like. It's like practicing your 00:11:30.970 --> 00:11:33.470 driving skills on a perfectly coded racing video 00:11:33.470 --> 00:11:35.889 game for 10 ,000 hours. Sure. Yeah. You know 00:11:35.889 --> 00:11:39.269 the physics perfectly. But then your very first 00:11:39.269 --> 00:11:41.649 real world driving lesson is in the middle of 00:11:41.649 --> 00:11:44.379 a blizzard in downtown Manhattan. The rules of 00:11:44.379 --> 00:11:47.240 driving are technically the same, but the sheer 00:11:47.240 --> 00:11:49.639 chaos and unpredictability of the environment 00:11:49.639 --> 00:11:52.259 just breaks your brain. That's a highly accurate 00:11:52.259 --> 00:11:54.820 way to visualize it. And because the chaotic, 00:11:55.000 --> 00:11:57.440 unstructured, everyday world is just so difficult 00:11:57.440 --> 00:12:00.279 for a general purpose robot to handle, what we've 00:12:00.279 --> 00:12:03.879 seen historically is autonomous robots evolving 00:12:03.879 --> 00:12:06.419 into highly specialized niches. They sort of 00:12:06.419 --> 00:12:09.179 retreat to where it's safe. Yeah, they thrive 00:12:09.179 --> 00:12:12.039 where their environments or their specific tasks 00:12:12.039 --> 00:12:14.929 are. very tightly constrained. We can trace the 00:12:14.929 --> 00:12:17.789 evolutionary tree back quite a ways. The very 00:12:17.789 --> 00:12:20.129 first autonomous robots were built in the late 00:12:20.129 --> 00:12:23.230 1940s by a neurologist named W. Gray Walter. 00:12:23.389 --> 00:12:26.049 The tortoises. Yes. He built these little machines 00:12:26.049 --> 00:12:28.529 named Elmer and Elsie, which people call tortoises 00:12:28.529 --> 00:12:30.190 because of their shape. They didn't have digital 00:12:30.190 --> 00:12:32.450 computers. They were entirely analog. Right. 00:12:32.629 --> 00:12:35.149 Meant to mimic basic biological brain functions 00:12:35.149 --> 00:12:37.789 using photo taxes, which is just a hardwired 00:12:37.789 --> 00:12:39.850 movement in response to a light stimulus. They 00:12:39.850 --> 00:12:42.360 just chased light around the room. And if we 00:12:42.360 --> 00:12:44.360 follow that evolutionary branch from those simple 00:12:44.360 --> 00:12:47.000 constrained tortoises, we arrive at environments 00:12:47.000 --> 00:12:50.460 that are so extreme, they actually force autonomy, 00:12:51.179 --> 00:12:54.299 like space exploration. Space is the ultimate 00:12:54.299 --> 00:12:57.259 harsh environment. Space rovers are arguably 00:12:57.259 --> 00:12:59.460 the most successful autonomous ground robots 00:12:59.460 --> 00:13:02.059 we have ever built purely out of necessity. Right, 00:13:02.179 --> 00:13:04.240 because of the speed of light. The communication 00:13:04.240 --> 00:13:06.460 delay between Earth and Mars can be up to 20 00:13:06.460 --> 00:13:08.980 minutes each way. You literally cannot drive 00:13:08.980 --> 00:13:11.659 a Mars rover with a joystick. Right. If you see 00:13:11.659 --> 00:13:13.580 it driving toward a cliff, by the time your stop 00:13:13.580 --> 00:13:15.639 command reaches Mars, the rover fell off the 00:13:15.639 --> 00:13:18.720 cliff 20 minutes ago. Exactly. The engineering 00:13:18.720 --> 00:13:21.919 breakdown of the Mars rovers like Spirit, Opportunity, 00:13:22.000 --> 00:13:24.379 and the planned Rosalind Franklin rover shows 00:13:24.379 --> 00:13:27.460 a constant closed -loop autonomous cycle. How 00:13:27.460 --> 00:13:30.639 does that loop work? The rover stops. uses its 00:13:30.639 --> 00:13:32.899 stereo cameras to reconstruct the 3D model of 00:13:32.899 --> 00:13:35.559 the terrain in front of it, computes which specific 00:13:35.559 --> 00:13:37.779 grids are safe and which have hazardous rocks, 00:13:38.399 --> 00:13:40.720 plots an optimal path exclusively across the 00:13:40.720 --> 00:13:43.220 safe zones, drives that short route, and then 00:13:43.220 --> 00:13:45.299 stops to repeat the entire cycle all over again. 00:13:45.399 --> 00:13:48.419 And the navigation tech gets even more impressive 00:13:48.419 --> 00:13:52.279 closer to home. In 2016, during a NASA centennial 00:13:52.279 --> 00:13:55.740 challenge, a specialized rover named Cataglyphus 00:13:55.740 --> 00:13:58.399 accomplished something wild. Oh, the Cataglyphus 00:13:58.399 --> 00:14:01.379 run. Over a two -hour period, it autonomously 00:14:01.379 --> 00:14:04.879 navigated 2 .6 kilometers across an unknown desert, 00:14:05.440 --> 00:14:08.440 detected specific samples, mechanically retrieved 00:14:08.440 --> 00:14:10.759 them, and returned precisely to its starting 00:14:10.759 --> 00:14:14.179 point. And here's the kicker. It did all of that 00:14:14.179 --> 00:14:16.700 without a GPS signal or a magnetometer. That 00:14:16.700 --> 00:14:18.559 is a staggering achievement in dead reckoning. 00:14:18.860 --> 00:14:21.460 How is that even possible without GPS? To navigate 00:14:21.460 --> 00:14:23.799 without GPS, Cataglyphus had to rely entirely 00:14:23.799 --> 00:14:27.200 on fusing data from internal sensors. It used 00:14:27.200 --> 00:14:29.259 wheeling coders, literally counting the exact 00:14:29.259 --> 00:14:31.500 number of rotations its wheels made to calculate 00:14:31.500 --> 00:14:33.580 distance. But wheels slip in the sand, don't 00:14:33.580 --> 00:14:36.320 they? Yes. So combine that data with an inertial 00:14:36.320 --> 00:14:38.259 measurement unit, like an electronic inner ear, 00:14:38.600 --> 00:14:40.879 to feel its own tilt and acceleration, while 00:14:40.879 --> 00:14:43.000 simultaneously mapping the terrain with LIDAR. 00:14:43.399 --> 00:14:45.100 It's a master class in specialized autonomy. 00:14:45.289 --> 00:14:47.850 And speaking of environments where traditional 00:14:47.850 --> 00:14:51.250 navigation like GPS might be denied, we see those 00:14:51.250 --> 00:14:54.509 exact same extreme engineering constraints driving 00:14:54.509 --> 00:14:57.750 massive investment in the military sector. Battlefields 00:14:57.750 --> 00:15:00.730 are environments where GPS is routinely jammed 00:15:00.730 --> 00:15:03.509 and radio signals are frequently lost. Exactly. 00:15:04.110 --> 00:15:06.879 Now, Looking objectively at the military sector, 00:15:07.399 --> 00:15:09.379 and to be clear, we are just neutrally reporting 00:15:09.379 --> 00:15:12.580 what the source material lays out here, the engineering 00:15:12.580 --> 00:15:15.299 constraints change drastically. The data shows 00:15:15.299 --> 00:15:17.960 an increasing reliance on lethal autonomous weapons, 00:15:18.340 --> 00:15:21.200 or laws. Or Ws. Right. These are systems engineered 00:15:21.200 --> 00:15:23.779 to independently search for and engage targets 00:15:23.779 --> 00:15:27.389 based on strictly programmed constraints. It 00:15:27.389 --> 00:15:29.769 is worth noting that current systems, based on 00:15:29.769 --> 00:15:32.330 the documentation up through 2018, generally 00:15:32.330 --> 00:15:34.590 still require a human in the loop to issue the 00:15:34.590 --> 00:15:36.629 final attack command, though there are certain 00:15:36.629 --> 00:15:39.149 automated defensive exceptions. The physical 00:15:39.149 --> 00:15:41.750 scale of these machines is incredible. Consider 00:15:41.750 --> 00:15:44.700 the crusher. An autonomous off -road combat vehicle 00:15:44.700 --> 00:15:46.879 developed for DARPA that weighs over 13 ,000 00:15:46.879 --> 00:15:50.039 pounds. Or the Ripsaw M5, which was unveiled 00:15:50.039 --> 00:15:53.879 in 2019. It has a combat weight of 10 .5 tons, 00:15:54.500 --> 00:15:57.019 can aggressively navigate rough terrain at speeds 00:15:57.019 --> 00:15:59.919 over 40 miles per hour, and is armed with a 30 00:15:59.919 --> 00:16:02.620 millimeter autocannon. From a purely systems 00:16:02.620 --> 00:16:05.019 engineering perspective, the Ripsaw M5 features 00:16:05.019 --> 00:16:07.879 a brilliant logic loop for its fail safes. How 00:16:07.879 --> 00:16:10.419 so? If the vehicle takes damage and is physically 00:16:10.419 --> 00:16:13.179 disabled, its autonomy engine is programmed to 00:16:13.179 --> 00:16:16.120 instantly re -prioritize. It reroutes power to 00:16:16.120 --> 00:16:18.799 its sensors and radio uplink, ensuring it can 00:16:18.799 --> 00:16:21.240 continue to transmit vital battlefield data as 00:16:21.240 --> 00:16:23.860 its primary function while retaining its defensive 00:16:23.860 --> 00:16:26.960 capabilities. It's an incredibly robust redundancy. 00:16:27.360 --> 00:16:30.399 There is also the SGR -A1, an autonomous stationary 00:16:30.399 --> 00:16:32.799 sentry gun developed to assist South Korean troops. 00:16:33.320 --> 00:16:35.779 It operates in the demilitarized zone and integrates 00:16:35.779 --> 00:16:38.289 surveillance, tracking and firing with recognition 00:16:38.289 --> 00:16:40.509 capabilities to challenge targets. So we have 00:16:40.509 --> 00:16:42.669 space and we have the military both environments 00:16:42.669 --> 00:16:45.009 that absolutely demand rugged specialized machines. 00:16:45.029 --> 00:16:47.289 Right. But then we have robots designed to work 00:16:47.289 --> 00:16:49.830 in our homes and hospitals. And this brings us 00:16:49.830 --> 00:16:53.490 to a wildly different approach humanoids. Machines 00:16:53.490 --> 00:16:56.590 explicitly engineered to mimic the human form. 00:16:56.870 --> 00:16:58.690 There are quite a few of these in development. 00:16:58.769 --> 00:17:01.450 There's the Tesla robot. There's Sophia created 00:17:01.450 --> 00:17:04.369 by Hanson Robotics, which utilizes an AI architecture 00:17:04.369 --> 00:17:07.430 called OpenCog to handle general reason. Sophia 00:17:07.430 --> 00:17:10.230 actually uses mechanical actuators under artificial 00:17:10.230 --> 00:17:12.930 skin to imitate human facial expressions too. 00:17:13.180 --> 00:17:15.220 Right, there's even a smaller companion version 00:17:15.220 --> 00:17:17.579 called Little Sophia designed to teach children 00:17:17.579 --> 00:17:19.799 programming. But then the history highlights 00:17:19.799 --> 00:17:22.380 another humanoid named Pepper, which was highly 00:17:22.380 --> 00:17:25.200 publicized but ultimately failed to achieve widespread 00:17:25.200 --> 00:17:27.960 commercial use. Yeah, Pepper ran into some hard 00:17:27.960 --> 00:17:30.440 limits. Here's where it gets really interesting. 00:17:30.599 --> 00:17:32.940 Let me push back on this entire humanoid trend. 00:17:33.380 --> 00:17:35.480 Given that Pepper failed because it was too complex 00:17:35.480 --> 00:17:38.400 and expensive, is making robots look human like 00:17:38.400 --> 00:17:41.680 Sophia actually practical for real -world adoption, 00:17:41.920 --> 00:17:44.579 or is it largely a PR exercise? It is one of 00:17:44.579 --> 00:17:46.480 the most hotly debated topics in robotics right 00:17:46.480 --> 00:17:49.039 now. I can imagine. The rationale behind humanoids 00:17:49.039 --> 00:17:52.579 addresses the open problem of human -robot interaction 00:17:52.579 --> 00:17:55.539 in unstructured settings. Meaning places where 00:17:55.539 --> 00:17:57.759 humans and robots are mixed together. Exactly. 00:17:57.980 --> 00:18:00.220 When robots share intimate spaces with humans, 00:18:00.519 --> 00:18:02.940 like a hospital ward or a living room, they need 00:18:02.940 --> 00:18:05.500 behaviors that humans can instinctively interpret. 00:18:05.960 --> 00:18:08.819 A humanoid face or a recognizable hand gesture 00:18:08.819 --> 00:18:12.240 provides immense psychological comfort and predictability 00:18:12.240 --> 00:18:14.559 for the humans around it. That makes sense. We 00:18:14.559 --> 00:18:16.359 know how to read a face. We don't know how to 00:18:16.359 --> 00:18:19.380 read a blinking LED light on a metal box. Exactly. 00:18:19.880 --> 00:18:21.779 But you hit the nail on the head regarding the 00:18:21.779 --> 00:18:24.880 downsides. The hardware and economic constraints 00:18:24.880 --> 00:18:27.839 of embodied A .I. are brutal. Because you're 00:18:27.839 --> 00:18:29.579 trying to fit everything into a human shape. 00:18:29.920 --> 00:18:32.200 Right. Trying to pack an advanced computing brain, 00:18:32.579 --> 00:18:35.339 dozens of complex joint actuators, and sufficient 00:18:35.339 --> 00:18:37.839 battery power into a frame that resembles a slender 00:18:37.839 --> 00:18:40.519 human body limits the machine's endurance and 00:18:40.519 --> 00:18:43.039 strength drastically. So it's a trade -off. For 00:18:43.039 --> 00:18:45.519 practical industrial adoption, purpose -built 00:18:45.519 --> 00:18:48.700 machines almost always win out. Look at the Satan 00:18:48.700 --> 00:18:51.160 AMR transfer cards used in manufacturing. Oh, 00:18:51.220 --> 00:18:53.940 the ones that carry heavy loads. Yeah. They don't 00:18:53.940 --> 00:18:56.460 look human at all. They look like giant rolling 00:18:56.460 --> 00:19:00.299 flatbed platforms. But they autonomously, safely, 00:19:00.539 --> 00:19:03.880 and efficiently transfer loads of up to 1 ,500 00:19:03.880 --> 00:19:07.160 kilograms around a busy factory floor all day 00:19:07.160 --> 00:19:10.200 long. Form follows function. A robot doesn't 00:19:10.200 --> 00:19:13.039 need a face to carry a ton of steel. Form follows 00:19:13.039 --> 00:19:15.640 function, exactly. So if a robot's job is to 00:19:15.640 --> 00:19:18.079 deliver a pizza, it shouldn't look like a tiny 00:19:18.079 --> 00:19:20.539 metal person carrying a box. It should just look 00:19:20.539 --> 00:19:24.380 like a rolling cooler. Precisely. And those specialized 00:19:24.380 --> 00:19:26.740 delivery robots aren't just staying inside factories 00:19:26.740 --> 00:19:29.220 anymore. They are finally trickling down into 00:19:29.220 --> 00:19:32.079 our civic spaces, rolling onto our sidewalks 00:19:32.079 --> 00:19:35.000 and forcing society to figure out how to actually 00:19:35.000 --> 00:19:37.200 live with them. The current legislative landscape 00:19:37.200 --> 00:19:39.900 for sidewalk delivery robots is a perfect case 00:19:39.900 --> 00:19:42.359 study of how society wrestles with the friction 00:19:42.359 --> 00:19:45.039 of integrating new technology. It's totally completely 00:19:45.039 --> 00:19:46.940 fragmented right now. The regulatory approaches 00:19:46.940 --> 00:19:49.079 across different cities are starkly different. 00:19:49.359 --> 00:19:52.700 We see that friction clearly. Back in 2016 and 00:19:52.700 --> 00:19:55.559 2017, places like Washington D .C. and the state 00:19:55.559 --> 00:19:58.339 of Virginia took a very welcoming, proactive 00:19:58.339 --> 00:20:01.819 approach. They passed legislation explicitly 00:20:01.819 --> 00:20:04.240 allowing autonomous delivery robots to travel 00:20:04.240 --> 00:20:06.960 on public sidewalks and use pedestrian crosswalks. 00:20:07.099 --> 00:20:09.000 They just put some basic safety parameters on 00:20:09.000 --> 00:20:10.900 them, right? Yeah, the robots couldn't weigh 00:20:10.900 --> 00:20:13.220 more than 50 pounds unloaded, and their software 00:20:13.220 --> 00:20:15.640 had to cap their speed at 10 miles per hour. 00:20:15.900 --> 00:20:17.359 But if you look at the other side of the country, 00:20:17.440 --> 00:20:19.680 San Francisco took a much more restrictive stance 00:20:19.680 --> 00:20:22.539 in late 2017. They were not having it. Their 00:20:22.539 --> 00:20:25.059 board of supervisors recognized the reality gap 00:20:25.059 --> 00:20:27.700 we discussed earlier. Sidewalks are chaotic. 00:20:28.329 --> 00:20:30.990 A robot freezing up because a sensor got confused 00:20:30.990 --> 00:20:34.609 by a puddle creates a massive hazard for pedestrians 00:20:34.609 --> 00:20:37.450 and wheelchair users. So they clamped down. Yeah, 00:20:37.509 --> 00:20:39.609 they required companies to obtain specific city 00:20:39.609 --> 00:20:43.309 permits just to test the robots, and they actively 00:20:43.309 --> 00:20:45.430 banned them from making non -researched commercial 00:20:45.430 --> 00:20:48.069 deliveries on city sidewalks. It's a real -time 00:20:48.069 --> 00:20:50.769 negotiation of public space. And this friction 00:20:50.769 --> 00:20:53.009 has elevated to the highest levels of global 00:20:53.009 --> 00:20:56.210 awareness. Elon Musk has been very vocal, publicly 00:20:56.210 --> 00:20:58.890 warning about the potential existential hazards 00:20:58.890 --> 00:21:02.269 of autonomous robotics and AI, even as his own 00:21:02.269 --> 00:21:04.630 company races to develop the Tesla humanoid. 00:21:04.849 --> 00:21:07.289 It's a complex dynamic. And again, just objectively 00:21:07.289 --> 00:21:10.650 reporting the facts here in 2021, a United Nations 00:21:10.650 --> 00:21:13.210 group of government experts convened specifically 00:21:13.210 --> 00:21:16.089 to address the accelerating technological capabilities 00:21:16.089 --> 00:21:19.519 of aut... weapons and the profound ethical frameworks 00:21:19.519 --> 00:21:22.039 needed to manage them. This raises an important 00:21:22.039 --> 00:21:24.039 question and perhaps the most critical hurdle 00:21:24.039 --> 00:21:26.400 for the next decade of this technology, the issue 00:21:26.400 --> 00:21:29.509 of ethics and liability. Who takes the blame 00:21:29.509 --> 00:21:32.710 when things go wrong? Exactly. Our legal frameworks 00:21:32.710 --> 00:21:35.630 are built around human agency. So who is legally 00:21:35.630 --> 00:21:37.930 and financially responsible when an autonomous 00:21:37.930 --> 00:21:40.809 system causes harm or property damage in a public 00:21:40.809 --> 00:21:43.210 space? Right. If an autonomous delivery robot 00:21:43.210 --> 00:21:46.009 isn't being directly controlled by a human operator 00:21:46.009 --> 00:21:49.329 and its path planning module makes a slight mathematical 00:21:49.329 --> 00:21:52.109 error that causes it to veer into traffic and 00:21:52.109 --> 00:21:55.019 cause an accident, Our current regulatory and 00:21:55.019 --> 00:21:57.539 insurance frameworks are simply not equipped 00:21:57.539 --> 00:22:01.140 to definitively assign that liability. So what 00:22:01.140 --> 00:22:03.940 does this all mean? Let's ground this entirely 00:22:03.940 --> 00:22:06.420 in your own reality. Think about your daily commute 00:22:06.420 --> 00:22:08.019 tomorrow morning. It's coming to your street 00:22:08.019 --> 00:22:10.940 soon. Exactly. Picture the layout of your own 00:22:10.940 --> 00:22:13.569 neighborhood. How would you feel sharing the 00:22:13.569 --> 00:22:16.410 crosswalk with a 50 pound robot doing a grocery 00:22:16.410 --> 00:22:19.789 run? How do we adapt our social norms when navigating 00:22:19.789 --> 00:22:22.490 a crowded sidewalk means sidestepping a machine 00:22:22.490 --> 00:22:24.710 that is constantly calculating your trajectory? 00:22:24.809 --> 00:22:26.730 It's an adjustment for sure. It's not a thought 00:22:26.730 --> 00:22:28.650 experiment about the distant future anymore. 00:22:28.690 --> 00:22:31.509 It's the present reality of our civil infrastructure. 00:22:31.750 --> 00:22:34.049 We have covered an immense amount of ground today. 00:22:34.130 --> 00:22:36.569 We really have. We started at the very foundation 00:22:36.569 --> 00:22:39.890 of biological mimicry, exploring how proprioception 00:22:39.890 --> 00:22:42.759 allows a robot to monitor its own voltage and 00:22:42.759 --> 00:22:45.039 how extra reception lets a vacuum hear the dirt 00:22:45.039 --> 00:22:48.220 on your rug. We broke down the reality gap, illuminating 00:22:48.220 --> 00:22:51.039 just how incredibly difficult it is to take a 00:22:51.039 --> 00:22:53.720 machine out of a perfectly mapped indoor hallway 00:22:53.720 --> 00:22:56.380 and force it to navigate the sheer chaos of an 00:22:56.380 --> 00:22:58.500 outdoor environment. And we examined how those 00:22:58.500 --> 00:23:01.220 engineering failures drove robots into highly 00:23:01.220 --> 00:23:04.400 specialized extreme niches like navigating the 00:23:04.400 --> 00:23:07.279 deserts of Mars without GPS or maintaining radio 00:23:07.279 --> 00:23:10.160 uplinks on a battlefield. And finally, how those 00:23:10.160 --> 00:23:12.220 machines are now sparking legislative battles 00:23:12.220 --> 00:23:15.140 on our local sidewalks. It is a massive ongoing 00:23:15.079 --> 00:23:17.720 shift in how we interact with the physical world. 00:23:18.039 --> 00:23:21.220 But before we sign off, there is one final, almost 00:23:21.220 --> 00:23:23.759 unnerving concept buried deep in the research 00:23:23.759 --> 00:23:25.920 that we haven't touched on yet. And it is something 00:23:25.920 --> 00:23:28.000 you should really chew on after you finish listening 00:23:28.000 --> 00:23:30.700 to this. It's a behavioral concept in advanced 00:23:30.700 --> 00:23:33.740 robotics known as energy autonomy and foraging. 00:23:34.000 --> 00:23:36.420 Foraging. When researchers talk about creating 00:23:36.420 --> 00:23:38.779 true artificial life, they aren't just looking 00:23:38.779 --> 00:23:40.920 at intelligent navigation. They are looking at 00:23:40.920 --> 00:23:44.740 giving robots the capacity to actively find and 00:23:44.740 --> 00:23:47.599 secure their own survival resources. Meaning, 00:23:48.039 --> 00:23:50.279 the robot doesn't just passively return to a 00:23:50.279 --> 00:23:52.799 pre -programmed wall charger in your kitchen 00:23:52.799 --> 00:23:55.940 when its battery is low. Right. It is programmed 00:23:55.940 --> 00:23:58.619 to actively hunt its environment for energy. 00:23:58.799 --> 00:24:02.319 It scavenges. Exactly. Imagine an autonomous 00:24:02.319 --> 00:24:05.200 machine deployed in a remote area making complex 00:24:05.200 --> 00:24:07.640 decisions about energy expenditure versus energy 00:24:07.640 --> 00:24:10.740 intake. It might actively hunt for a solar patch 00:24:10.740 --> 00:24:13.539 to recharge or even scavenge a scrap yard for 00:24:13.539 --> 00:24:16.359 usable spare parts and compatible batteries to 00:24:16.359 --> 00:24:18.420 physically repair itself in order to survive. 00:24:18.740 --> 00:24:21.240 A machine hunting for its own survival resources 00:24:21.240 --> 00:24:23.839 in the wild. It completely upends the traditional 00:24:23.839 --> 00:24:26.859 relationship between creator and tool. It forces 00:24:26.859 --> 00:24:29.309 you to ask. What happens when the machines we 00:24:29.309 --> 00:24:31.829 build no longer need us to plug them in? It is 00:24:31.829 --> 00:24:34.829 a profound, slightly unsettling question about 00:24:34.829 --> 00:24:37.250 the ultimate trajectory of artificial life. It 00:24:37.250 --> 00:24:39.609 really is. Thank you for joining us on this deep 00:24:39.609 --> 00:24:41.470 dive. Keep questioning the world around you. 00:24:41.509 --> 00:24:43.869 Keep exploring. And the next time you see a machine 00:24:43.869 --> 00:24:46.170 doing a job, you know, remember, it's not just 00:24:46.170 --> 00:24:48.289 a mechanical arm trapped safely in a cage anymore. 00:24:48.349 --> 00:24:50.410 It's out here on the sidewalk with us, trying 00:24:50.410 --> 00:24:53.250 to figure out the chaotic real world one sensor 00:24:53.250 --> 00:24:53.829 at a time.