WEBVTT 1 00:00:03.399 --> 00:00:07.719 Welcome to Bedtime Astronomy. Explore the wonders of the cosmos 2 00:00:07.759 --> 00:00:12.279 with our soothing Bedtime Astronomie podcast. Each episode offers a 3 00:00:12.359 --> 00:00:16.320 gentle journey through the stars, planets, and beyond, perfect for 4 00:00:16.399 --> 00:00:20.239 unwinding after a long day. Let's travel through the mysteries 5 00:00:20.239 --> 00:00:22.440 of the universe as you drift off into a peaceful 6 00:00:22.480 --> 00:00:26.800 slumber under the night sky. 7 00:00:26.920 --> 00:00:30.280 Welcome back to the show. It is Monday, February second, 8 00:00:30.399 --> 00:00:33.479 twenty twenty six, and I have to say, looking at 9 00:00:33.520 --> 00:00:36.759 the date today, it feels like we are finally living 10 00:00:36.799 --> 00:00:39.920 in that future we were always promised and the you know, 11 00:00:40.240 --> 00:00:42.600 the paperback sci fi novels of the nineties. 12 00:00:42.840 --> 00:00:44.920 It certainly does. Especially when you look at the news 13 00:00:44.960 --> 00:00:48.719 we're unpacking today, it feels less like a standard press release, yeah, 14 00:00:48.759 --> 00:00:52.159 and more like a lost chapter from an Isaac Asimov 15 00:00:52.159 --> 00:00:53.039 book exactly. 16 00:00:53.159 --> 00:00:55.520 I mean. Down here on Earth, we've become pretty jaded 17 00:00:55.560 --> 00:00:58.600 about artificial intelligence over the last few years. We see 18 00:00:58.640 --> 00:01:03.799 the headlines about self driving taxis navigating San Francisco. 19 00:01:03.359 --> 00:01:07.120 Fog or AI writing generic marketing emails that all sound. 20 00:01:07.000 --> 00:01:10.319 The same, right, or generating that weird, slightly hallucinated art 21 00:01:10.359 --> 00:01:13.719 with seven fingers. On one hand, it's almost become background noise. 22 00:01:13.959 --> 00:01:16.120 We sort of take it for granted. But the milestone 23 00:01:16.120 --> 00:01:19.959 we are discussing today takes that technology and well literally 24 00:01:20.000 --> 00:01:21.519 puts it on a completely different. 25 00:01:21.280 --> 00:01:23.879 Planet is a massive shift in context. We aren't talking 26 00:01:23.920 --> 00:01:26.760 about a chatbot helping you organize your calendar. We're talking 27 00:01:26.760 --> 00:01:30.879 about a nuclear powered explorer navigating the most hostile, unforgiving 28 00:01:31.000 --> 00:01:33.400 terrain imaginable entirely on its own. 29 00:01:33.719 --> 00:01:37.280 And that's the core news event. It's something that frankly 30 00:01:37.519 --> 00:01:42.680 changes the game for space exploration. NASA's Perseverance Rover, which 31 00:01:42.680 --> 00:01:45.319 has been roaming the Red Planet for years now, has 32 00:01:45.359 --> 00:01:51.480 successively completed its first ever drives planned entirely by artificial intelligence. 33 00:01:52.359 --> 00:01:55.000 And I don't mean assisted driving where a human is 34 00:01:55.040 --> 00:01:58.359 sort of hovering over a brake pedal. I mean the 35 00:01:58.439 --> 00:01:59.799 AI took the wheel right. 36 00:02:00.000 --> 00:02:01.439 And just to be clear from the outset, because I 37 00:02:01.439 --> 00:02:05.040 think people hear AI and rover and they immediately think 38 00:02:05.079 --> 00:02:07.159 of the old collision avoidance systems. 39 00:02:06.840 --> 00:02:08.680 We've had for a while, sure, like the sensors in 40 00:02:08.719 --> 00:02:09.159 your car. 41 00:02:09.360 --> 00:02:12.199 Exactly, This isn't the rover just avoiding a sharp rock 42 00:02:12.240 --> 00:02:14.319 while a human holds the steering wheel. This is the 43 00:02:14.400 --> 00:02:18.120 AI planning the route, making the decisions, and then executing 44 00:02:18.120 --> 00:02:20.240 the drive without a human looking over its shoulder in 45 00:02:20.280 --> 00:02:20.800 real time. 46 00:02:21.039 --> 00:02:23.240 That's the hook for me. I want you to imagine 47 00:02:23.280 --> 00:02:25.879 driving a car, but the steering wheel is one hundred 48 00:02:25.879 --> 00:02:28.479 and forty million miles away, and instead of trying to 49 00:02:28.520 --> 00:02:30.879 steer it yourself with the twenty minute delay, which is 50 00:02:31.680 --> 00:02:34.479 physically impossible, you just hand the keys to the onboard 51 00:02:34.520 --> 00:02:36.680 computer and say you figure it out, get me to 52 00:02:36.719 --> 00:02:37.599 that mountain range. 53 00:02:37.719 --> 00:02:40.840 And that's what NASA has effectively just done. It's incredible, 54 00:02:41.000 --> 00:02:42.719 it is yeah, and it's a very active way to 55 00:02:42.759 --> 00:02:44.919 put it. To guide us through this. We're looking at 56 00:02:44.919 --> 00:02:48.840 a comprehensive report that was released today February two from 57 00:02:48.919 --> 00:02:51.560 NASA and the Jet Propulsion Laboratory or JPL. 58 00:02:51.719 --> 00:02:54.000 So this is the real deal. This isn't just speculation, 59 00:02:54.080 --> 00:02:54.680 Oh no, this is it. 60 00:02:54.719 --> 00:02:58.080 This comes straight from the Rover Operation Center. We're analyzing 61 00:02:58.159 --> 00:03:01.439 data that's been verified by the Mars Connissance orbiter and 62 00:03:02.159 --> 00:03:05.159 there are details on a specific collaboration with the AI 63 00:03:05.240 --> 00:03:08.599 company Anthropic, So we have a lot of credible, high 64 00:03:08.680 --> 00:03:11.439 level data to sift through. This isn't a roadmap. This 65 00:03:11.520 --> 00:03:13.319 is operational history being made right now. 66 00:03:13.400 --> 00:03:16.319 Okay, so our mission for this discussion is pretty clear. 67 00:03:16.840 --> 00:03:19.479 We need to unpack how we got from generative AI 68 00:03:19.599 --> 00:03:23.520 writing bad poetry on the Internet to navigating bedrock and 69 00:03:23.599 --> 00:03:26.639 craters on Mars. We need to understand the technical leap here. 70 00:03:26.919 --> 00:03:29.840 We need to analyze that delta because for the last 71 00:03:29.840 --> 00:03:32.879 twenty eight years of Mars exploration, humans have been very 72 00:03:33.000 --> 00:03:35.840 very much in the loop. This represents a move from 73 00:03:35.919 --> 00:03:39.560 human joysticking, or at least the illusion of joysticking, to 74 00:03:40.360 --> 00:03:41.800 true robotic agency. 75 00:03:42.159 --> 00:03:45.400 And we definitely need to talk about the safety aspect 76 00:03:45.840 --> 00:03:48.280 because if my computer crash is here in the studio, 77 00:03:48.759 --> 00:03:51.599 I reboot it and maybe lose a word document. If 78 00:03:51.639 --> 00:03:54.319 a rover crash is on Mars, that's a billion dollar 79 00:03:54.400 --> 00:03:57.280 mistake that you can't undo. Right, So we're going to 80 00:03:57.280 --> 00:04:00.280 talk about the digital twin technology that keeps it safe, 81 00:04:00.319 --> 00:04:02.439 which is a fascinating concept in itself. 82 00:04:02.479 --> 00:04:05.919 It's a fascinating layer of protection. Honestly, it's probably the 83 00:04:06.000 --> 00:04:08.199 only reason the engineers at JPL can sleep at night. 84 00:04:08.599 --> 00:04:11.159 All right, let's get into the milestone itself. Let's start 85 00:04:11.240 --> 00:04:14.919 by decoupling the human driver from the machine. 86 00:04:15.080 --> 00:04:17.319 Right, So let's look at this specific event. This all 87 00:04:17.319 --> 00:04:20.959 happened very recently. The breakthrough drives occurred on December eighth 88 00:04:21.040 --> 00:04:23.120 and December tenth, twenty twenty five. 89 00:04:23.279 --> 00:04:25.439 Okay, so just a couple of months ago. And to 90 00:04:25.519 --> 00:04:30.240 set the scene visually, where exactly is Perseverance located? Right now? 91 00:04:30.319 --> 00:04:32.839 It's situated at the rim of Jezuro Crater, which, from 92 00:04:32.839 --> 00:04:35.319 what I understand, is not exactly a paved parking lot. 93 00:04:35.399 --> 00:04:37.839 It's not the Bonneville Salt Flats where you can just 94 00:04:37.959 --> 00:04:39.399 you know, throttle down and go. 95 00:04:39.560 --> 00:04:40.240 Far from it. 96 00:04:40.560 --> 00:04:40.639 No. 97 00:04:41.560 --> 00:04:46.279 Jezro is scientifically fascinating, but logistically it's a nightmare. You 98 00:04:46.279 --> 00:04:49.319 have steep slopes, you have patches of loose sand that 99 00:04:49.360 --> 00:04:52.480 act like traps. You have shirt bedrock outcrops that can 100 00:04:52.519 --> 00:04:56.360 chew up the wheels. It is a complex, treacherous environment 101 00:04:56.399 --> 00:04:59.000 for a human to navigate, let alone a machine doing 102 00:04:59.040 --> 00:04:59.920 it completely solo. 103 00:05:00.360 --> 00:05:03.439 I love the analogy of a child taking their first steps, 104 00:05:03.480 --> 00:05:05.160 you know that feeling you let go of their hand 105 00:05:05.199 --> 00:05:06.959 and they wobble a bit and you just hold your breath, 106 00:05:07.000 --> 00:05:10.040 hoping they don't face plant. Except in this case, the 107 00:05:10.120 --> 00:05:13.399 child is a one ton nuclear powered robot and the 108 00:05:13.439 --> 00:05:16.040 parents are millions of miles away. Looking at a twenty 109 00:05:16.079 --> 00:05:17.480 minute delayed video feed. 110 00:05:17.720 --> 00:05:20.720 That's the tension exactly, and the steps it took were significant. 111 00:05:20.759 --> 00:05:22.879 This wasn't just inching forward to test the waters. 112 00:05:23.000 --> 00:05:24.439 This wasn't a baby step. 113 00:05:24.240 --> 00:05:26.879 Not at all. On that first drive on December eighth, 114 00:05:26.920 --> 00:05:29.759 Perseverance drove six hundred and eighty nine feet that's about 115 00:05:29.759 --> 00:05:30.959 two hundred and ten meters. 116 00:05:31.040 --> 00:05:34.040 That's more than two American football fields on its own 117 00:05:34.199 --> 00:05:35.360 in one gup exactly. 118 00:05:35.800 --> 00:05:38.439 And then just two days later, on December tenth, it 119 00:05:38.560 --> 00:05:41.040 beat that record. It drove eight hundred and seven feet, 120 00:05:41.240 --> 00:05:42.920 which is two hundred and forty six meters. 121 00:05:43.079 --> 00:05:45.879 So why are those numbers so significant? Because I can 122 00:05:45.879 --> 00:05:49.399 hear a listener saying, okay, eight hundred feet. My commute 123 00:05:49.480 --> 00:05:51.800 is twenty miles, My roomb is eight hundred feet in 124 00:05:51.800 --> 00:05:54.480 my living room. Why is this such a big deal 125 00:05:54.519 --> 00:05:55.000 for MASA. 126 00:05:55.160 --> 00:05:56.600 You have to look at it in the context of 127 00:05:56.639 --> 00:06:00.399 Martian exploration history. These aren't just inches. In the past, 128 00:06:00.439 --> 00:06:04.839 Every single meter, every foot had to be meticulously planned 129 00:06:04.879 --> 00:06:08.040 by a human to have the rover cover that kind 130 00:06:08.079 --> 00:06:13.279 of distance autonomously in complex terrain. It's a quantum leap. 131 00:06:13.480 --> 00:06:14.120 Quantum leap. 132 00:06:14.199 --> 00:06:17.879 It proves the system isn't just cautious, it's capable. It's 133 00:06:17.959 --> 00:06:21.560 traversing distances that would typically take days, sometimes weeks, of 134 00:06:21.639 --> 00:06:25.079 back and forth communication and planning. It's a fundamental shift 135 00:06:25.160 --> 00:06:26.279 in the operational tempo. 136 00:06:26.399 --> 00:06:28.360 Okay, so before we get too deep into the how, 137 00:06:28.480 --> 00:06:30.120 I think we need to bust a myth I alluded 138 00:06:30.120 --> 00:06:32.680 to it earlier, the status quo. I think when most 139 00:06:32.759 --> 00:06:35.839 people picture NASA driving a rover, they picture a guy 140 00:06:35.879 --> 00:06:38.120 in a blue polo shirt with a headset and a joystick. 141 00:06:38.160 --> 00:06:39.720 Oh yeah, video game image. 142 00:06:39.519 --> 00:06:42.079 Right, watching a screen turning left and right in real 143 00:06:42.120 --> 00:06:43.519 time like they're playing a video game. 144 00:06:43.680 --> 00:06:47.199 That is the classic Hollywood image, and it is completely 145 00:06:47.240 --> 00:06:50.279 fundamentally wrong. It is physically impossible. 146 00:06:49.720 --> 00:06:52.079 Because of the speed of light. Right, it's the ultimate 147 00:06:52.079 --> 00:06:53.680 speed limit. You just can't get around it. 148 00:06:53.759 --> 00:06:58.000 Precisely, Mars is on average about one hundred and forty 149 00:06:58.000 --> 00:07:02.399 million miles away. That's one hundred and twenty five million kilometers. 150 00:07:03.040 --> 00:07:06.199 Even at the speed of light, radio signals take anywhere 151 00:07:06.199 --> 00:07:08.639 from five to twenty minutes to get there, and then 152 00:07:08.680 --> 00:07:11.680 another five to twenty minutes to get back. It varies 153 00:07:11.720 --> 00:07:13.839 depending on where the planets are in their orbits. 154 00:07:14.040 --> 00:07:16.040 So let's play that out. If you're the driver and 155 00:07:16.079 --> 00:07:18.120 you saw a cliff coming up on your screen and 156 00:07:18.120 --> 00:07:19.240 you hit the brakes. 157 00:07:18.920 --> 00:07:21.240 The rover would have fallen off that clip twenty minutes ago. 158 00:07:21.480 --> 00:07:24.040 You cannot drive live. It's like trying to drive a 159 00:07:24.079 --> 00:07:25.839 car while looking at a photo of the road that 160 00:07:25.879 --> 00:07:28.439 was taken ten minutes ago. If you see a pedestrian 161 00:07:28.439 --> 00:07:30.240 in the photo, well, you've already hit them. 162 00:07:30.399 --> 00:07:32.560 So how have we been doing it for the last 163 00:07:32.600 --> 00:07:35.800 twenty eight years since Sojourner back in the nineties. 164 00:07:35.839 --> 00:07:38.560 It's a painstaking process. We call it the human in 165 00:07:38.600 --> 00:07:42.879 the loop workflow. Typically, human rover planners sitting at JPL 166 00:07:42.920 --> 00:07:45.560 in California download all the latest images from the rover 167 00:07:45.680 --> 00:07:49.040 and from satellites. They analyze the terrain manually. They look 168 00:07:49.040 --> 00:07:52.000 at the rocks, the sand, traps, the slopes. They put 169 00:07:52.040 --> 00:07:56.439 on three D glasses literally stereoscopic glasses, and stare at 170 00:07:56.480 --> 00:07:58.720 serial images to judge depth and distance. 171 00:07:58.800 --> 00:08:01.079 They're basically staring at foot photos and drawing a line 172 00:08:01.079 --> 00:08:04.480 on a map like go here, then turn left thirty degrees, 173 00:08:04.560 --> 00:08:05.160 then stop. 174 00:08:05.480 --> 00:08:09.560 Essentially, yes, they sketch a route using waypoints. These are 175 00:08:09.600 --> 00:08:12.959 specific coordinates the rover travels to, and because humans are 176 00:08:12.959 --> 00:08:15.399 cautious and because they can only see so much from 177 00:08:15.439 --> 00:08:18.959 static images, these waypoints are usually spaced pretty close together. 178 00:08:19.040 --> 00:08:20.360 How close are we talking. 179 00:08:20.519 --> 00:08:22.600 Usually know more than about one hundred feet to maybe 180 00:08:22.600 --> 00:08:24.959 three hundred and thirty feet apart on a really good day, 181 00:08:25.000 --> 00:08:28.519 with clear terrain, that's thirty to one hundred meters. They 182 00:08:28.519 --> 00:08:30.439 have to be short mops to ensure they aren't sending 183 00:08:30.439 --> 00:08:32.679 the rover into a hazard they didn't catch in the photos. 184 00:08:32.759 --> 00:08:35.320 So the old way is move a little bit, stop, 185 00:08:35.399 --> 00:08:37.320 take a picture, send it all the way to Earth, 186 00:08:37.600 --> 00:08:39.759 wait for a whole team of humans to look at it, 187 00:08:39.919 --> 00:08:42.240 sleep on it, draw new lines, send the command back, 188 00:08:42.279 --> 00:08:44.159 and then it moves a little bit more. It sounds 189 00:08:44.240 --> 00:08:46.159 excruciatingly slow it is. 190 00:08:46.399 --> 00:08:49.399 It's incredibly stop and go. It relies entirely on the 191 00:08:49.440 --> 00:08:52.720 Earth based cycle. If the humans are sleeping, the rover 192 00:08:52.879 --> 00:08:55.559 is sleeping. If the data takes a while to download 193 00:08:55.679 --> 00:08:58.679 via the deep space network, the rover sits there. The 194 00:08:58.759 --> 00:09:01.519 rover spends far more time I'm waiting than it does driving. 195 00:09:01.600 --> 00:09:03.759 And when you have a limited mission lifespan, that's just 196 00:09:03.919 --> 00:09:06.080 lost time, wasted opportunity. 197 00:09:06.120 --> 00:09:09.879 It's terribly inefficient. And this new system, this new AI, 198 00:09:10.480 --> 00:09:12.960 it blows that limitation completely out of the water. 199 00:09:13.200 --> 00:09:16.159 It decouples the driving from the human schedule. 200 00:09:16.279 --> 00:09:20.320 It decouples the rover's progress from Earth's rotation. The rover 201 00:09:20.399 --> 00:09:22.799 can make decisions on the fly. It doesn't need to 202 00:09:22.840 --> 00:09:26.120 ask for permission for every meter. It just needs a destination. 203 00:09:26.879 --> 00:09:28.759 Let's get into the technology, because this is the part 204 00:09:28.799 --> 00:09:30.879 that I think will really surprise people. We aren't just 205 00:09:30.919 --> 00:09:34.159 talking about a basic collision avoidance system like you have 206 00:09:34.200 --> 00:09:35.080 in a modern car. 207 00:09:35.159 --> 00:09:35.360 Right. 208 00:09:35.600 --> 00:09:37.600 This isn't just a sensor beeping if you get too 209 00:09:37.600 --> 00:09:38.360 close to a wall. 210 00:09:38.480 --> 00:09:42.039 No, no, this is much more sophisticated. The report identifies 211 00:09:42.080 --> 00:09:46.240 the specific technology as generative AI using vision language models. 212 00:09:46.519 --> 00:09:50.399 Okay, stop right there, vision language models. When I hear 213 00:09:50.559 --> 00:09:54.200 generative AI, I think of chatbots. I think of asking 214 00:09:54.200 --> 00:09:56.840 a computer to write me a recipe for lasagna or 215 00:09:56.840 --> 00:09:58.480 a sonnet about a toaster. 216 00:09:58.480 --> 00:10:00.279 Right, And that's what most of us think of. It's 217 00:10:00.320 --> 00:10:03.639 the same underlying architecture, it's the same kind of neural network, 218 00:10:04.080 --> 00:10:05.879 but it's applied in a completely different way. 219 00:10:06.039 --> 00:10:08.480 So how does a language model drive a rover. Is 220 00:10:08.559 --> 00:10:11.360 it talking to the rocks? Hello? 221 00:10:11.440 --> 00:10:11.639 Rock? 222 00:10:11.679 --> 00:10:13.200 Are you friendly? What's happening here? 223 00:10:13.360 --> 00:10:16.399 It's a great question, and it creates a bit of 224 00:10:16.440 --> 00:10:20.279 cognitive dissonance. But think about what those models are actually doing. 225 00:10:20.559 --> 00:10:24.039 They're processing vast amounts of information. In the case of 226 00:10:24.039 --> 00:10:26.720 a chatbot, that's text and finding patterns. 227 00:10:26.759 --> 00:10:27.039 Okay. 228 00:10:27.679 --> 00:10:30.440 In this case, the initiative was led by JPL's Rover 229 00:10:30.600 --> 00:10:34.919 Operations Center, the ROC, in collaboration with Anthropic. 230 00:10:34.480 --> 00:10:36.000 The makers of claud Ai. 231 00:10:36.279 --> 00:10:39.399 Correct, they're using claude Ai models. Now, a vision language 232 00:10:39.440 --> 00:10:43.200 model doesn't just process text. It processes images as if 233 00:10:43.240 --> 00:10:45.279 they were a language. You can look at a photo 234 00:10:45.360 --> 00:10:47.240 and understand the context of what is in it. 235 00:10:47.320 --> 00:10:50.960 So it's not just seeing pixels. It's seeing rock, sand cliff. 236 00:10:51.399 --> 00:10:54.480 It's assigning meaning to the visual data exactly. 237 00:10:54.519 --> 00:10:57.720 It analyzes imagery. And the crucial part here is that 238 00:10:57.759 --> 00:11:00.960 it uses the exact same visual data that the human 239 00:11:01.039 --> 00:11:04.120 planners use. It's not using some secret new sensor that 240 00:11:04.159 --> 00:11:06.840 we don't know about. It's looking at the same maps, 241 00:11:06.960 --> 00:11:10.360 the same photos and deciding independently where to put the 242 00:11:10.399 --> 00:11:10.919 wait points. 243 00:11:11.279 --> 00:11:14.720 You used to term there semantic understanding explain that distinction 244 00:11:14.879 --> 00:11:16.480 for us because that sounds important. 245 00:11:16.720 --> 00:11:20.879 It's the key difference. Traditional computer vision, like in older systems, 246 00:11:21.279 --> 00:11:23.960 looks at geometry. It sees a bump, it sees a drop, 247 00:11:24.000 --> 00:11:26.840 it sees an obstacle, but it doesn't necessarily know what 248 00:11:26.879 --> 00:11:29.279 that obstacle is. Is it a soft bush or a 249 00:11:29.279 --> 00:11:31.840 granite boulder. A vision language model can look at a 250 00:11:31.840 --> 00:11:33.519 patch of ground and say that is a sand ribble. 251 00:11:33.600 --> 00:11:37.840 Sand ripples are dangerous avoid, or that is bedrock. Bedrock 252 00:11:37.919 --> 00:11:40.840 is stable and safe to drive on drive. It brings 253 00:11:40.840 --> 00:11:44.600 a layer of reasoning, of geologic understanding to the image processing. 254 00:11:44.639 --> 00:11:46.840 It's a huge difference. It's the difference between seeing a 255 00:11:46.879 --> 00:11:48.559 shape and knowing it's a stop sign, and you have 256 00:11:48.639 --> 00:11:51.480 to break let's break down the inputs. What is the 257 00:11:51.519 --> 00:11:54.240 AI actually looking at to make these decisions. 258 00:11:54.480 --> 00:11:58.039 There are three main buckets of data that report highlights. First, 259 00:11:58.080 --> 00:12:01.240 it's using high resolution orbital image. This comes from the 260 00:12:01.320 --> 00:12:02.519 high rise camera. 261 00:12:02.360 --> 00:12:05.200 That's on the Mars reconnaissance orbiter flying overhead right so. 262 00:12:05.200 --> 00:12:07.799 It has the bird's eye view the macromap. It sees 263 00:12:07.840 --> 00:12:11.000 the layout of the land from space, the big picture. Second, 264 00:12:11.480 --> 00:12:14.960 it uses terrain slope data from digital elevation models, so 265 00:12:14.960 --> 00:12:16.919 it knows the three D shape of the ground, where 266 00:12:16.960 --> 00:12:19.000 it's flat, where it's deep, where the cliffs. 267 00:12:18.639 --> 00:12:20.440 Are, got it the lay of the land. 268 00:12:21.279 --> 00:12:24.679 And the third is JPL's own surface mission data set. 269 00:12:24.960 --> 00:12:28.080 This is the historical knowledge, the context of the mission itself. 270 00:12:28.360 --> 00:12:31.320 It's effectively the memory of where the rover has been, 271 00:12:31.799 --> 00:12:34.360 what kind of terrain it has encountered before, and what 272 00:12:34.399 --> 00:12:35.840 the overall mission parameters are. 273 00:12:35.879 --> 00:12:38.159 So it takes all that data, the map from above, 274 00:12:38.320 --> 00:12:40.440 the three D model of the ground, and its own 275 00:12:40.480 --> 00:12:42.279 memories and does what. 276 00:12:42.399 --> 00:12:46.080 It makes decisions. The report specifically lists the features the 277 00:12:46.120 --> 00:12:49.120 AI is trained to identify. It looks for bedrock, it 278 00:12:49.120 --> 00:12:52.000 looks for outcrops, it looks for hazardous boulder fields, and 279 00:12:52.320 --> 00:12:54.000 very importantly, it looks for sand ripples. 280 00:12:54.279 --> 00:12:56.960 Sand ripples sounds so harmless, but on Mars they are 281 00:12:56.960 --> 00:12:59.919 basically quicksand traps for rovers right, they are deadly. 282 00:13:00.720 --> 00:13:04.720 The Spirit rover, one of the previous generation, eventually met 283 00:13:04.759 --> 00:13:08.519 its end, effectively getting stuck in soft soil that looks solid. 284 00:13:09.039 --> 00:13:11.559 If you get stuck in a sandtrap on Mars mission over, 285 00:13:11.840 --> 00:13:14.600 you can't call a tow truck wow. So the AI 286 00:13:14.679 --> 00:13:18.519 identifies all these features, weighs the risks, and then generates 287 00:13:18.559 --> 00:13:21.919 a continuous path. It places those wait points, those fixed 288 00:13:21.960 --> 00:13:26.039 locations where the rover can stop for new instructions completely 289 00:13:26.080 --> 00:13:26.600 on its own. 290 00:13:26.679 --> 00:13:29.159 That is just wild. It's basically doing the job of 291 00:13:29.240 --> 00:13:33.639 a highly trained, very cautious NASA engineer, but it's doing 292 00:13:33.679 --> 00:13:35.840 it right there on the processor in real time. 293 00:13:35.919 --> 00:13:38.240 Well, the processing is fascinating. Yeah, but I know what 294 00:13:38.279 --> 00:13:40.879 you're thinking, and I know what the listener is probably 295 00:13:40.879 --> 00:13:41.480 thinking right now. 296 00:13:41.639 --> 00:13:44.759 That trust oh absolutely. I mean we've all seen AI 297 00:13:44.840 --> 00:13:48.360 make mistakes. We've seen them hallucinate facts. We've seen chatbots 298 00:13:48.360 --> 00:13:50.840 and vent court cases that never happen. What if this 299 00:13:50.919 --> 00:13:53.320 AI hallucinates a road where there's actually a crater. 300 00:13:53.639 --> 00:13:56.399 That is the billion dollar question. It's the thing that 301 00:13:56.480 --> 00:13:59.000 keeps engineers up at night. And this brings us to 302 00:13:59.039 --> 00:14:01.759 the next critical piece, the safety net. 303 00:14:02.039 --> 00:14:05.360 Right because NASA isn't known for being reckless. They don't 304 00:14:05.360 --> 00:14:08.200 just upload code and hope for the best. The Silicon 305 00:14:08.279 --> 00:14:12.360 Valley motto of move fast and break things doesn't really 306 00:14:12.440 --> 00:14:15.480 work when breaking things ends the entire space program. 307 00:14:15.600 --> 00:14:20.120 No, they are incredibly and rightly risk averse. The solution 308 00:14:20.200 --> 00:14:22.960 they developed is something they call the digital twin. 309 00:14:23.360 --> 00:14:26.399 I love this concept. It sounds very cyberpunk. Explain what 310 00:14:26.480 --> 00:14:28.240 a digital twin is in this context. 311 00:14:28.279 --> 00:14:30.320 It's exactly what it sounds like. Yeah, it is a 312 00:14:30.399 --> 00:14:34.200 virtual replica of the Perseverance Rover, but it lives inside 313 00:14:34.240 --> 00:14:37.799 a supercomputer at JPL in California. It is a physics 314 00:14:37.799 --> 00:14:40.960 perfect simulation of the rover and the Martian environment it's in. 315 00:14:41.240 --> 00:14:44.559 So before the real rover moves an inch on Mars, the. 316 00:14:44.519 --> 00:14:47.360 AI generates the plan. It says I want to drive here, here, 317 00:14:47.480 --> 00:14:50.240 and here. But instead of just beaming that command straight 318 00:14:50.279 --> 00:14:53.480 to Mars, the engineers feed it into the digital twin. 319 00:14:53.600 --> 00:14:55.240 First they simulate the drive. 320 00:14:55.559 --> 00:14:58.399 They do more than just simulate it, They stress test 321 00:14:58.440 --> 00:15:01.879 it to an unbelievable degree. The report states they verify 322 00:15:01.960 --> 00:15:04.360 over five hundred thousand telemetry. 323 00:15:04.000 --> 00:15:06.440 Variables five hundred thousand half. 324 00:15:06.240 --> 00:15:09.200 A million data points for every single proposed drive. They 325 00:15:09.240 --> 00:15:13.360 are checking everything wheel traction, suspension tilt, the power being 326 00:15:13.440 --> 00:15:16.399 drawn by each motor, the thermal limits on the electronics, 327 00:15:16.399 --> 00:15:20.399 the currents. They run the AI's plan through this digital gauntlet. 328 00:15:20.519 --> 00:15:23.600 So if the AI says, drive over that sketchy looking rock, 329 00:15:23.840 --> 00:15:26.559 the digital twin simulates it and flags a warning like 330 00:15:26.720 --> 00:15:31.080 warning suspension damage is ninety percent likely or warning tilt 331 00:15:31.159 --> 00:15:33.519 exceeds safe threshold by fifteen degrees. 332 00:15:33.600 --> 00:15:37.440 Precisely. It catches the hallucinations, it catches the risky maneuvers, 333 00:15:37.639 --> 00:15:41.120 it catches the small miscalculations, and only after the digital 334 00:15:41.200 --> 00:15:44.440 twin confirms the drive is safe green lights Across all 335 00:15:44.519 --> 00:15:46.919 half a million variables are the plans packaged up and 336 00:15:47.000 --> 00:15:50.120 sent via NASA's Deep space network to the physical rover 337 00:15:50.240 --> 00:15:50.679 on Mars. 338 00:15:50.720 --> 00:15:53.799 So it's a hybrid model. It's not fully autonomous in 339 00:15:53.840 --> 00:15:55.720 the sense that the AI can do whatever it wants. 340 00:15:55.879 --> 00:15:59.720 It is. It's AI planning, digital verification by humans and 341 00:15:59.759 --> 00:16:03.960 their simulators, and then physical execution. It's not blind trust. 342 00:16:04.279 --> 00:16:07.639 It's trust, but verify on a massive, massive scale. 343 00:16:07.720 --> 00:16:09.600 That makes me feel a lot better about the whole thing. 344 00:16:09.639 --> 00:16:11.879 It's not just a robot running wild. It's a robot 345 00:16:11.879 --> 00:16:15.559 proposing a plan and a very sophisticated simulation proving it 346 00:16:15.600 --> 00:16:16.919 works before anything real. 347 00:16:16.840 --> 00:16:19.960 Happens, and that verification step is what allows them to 348 00:16:19.960 --> 00:16:22.759 be bold. It allows them to let the AI try things, 349 00:16:23.080 --> 00:16:26.639 to plan these longer, more aggressive routes, knowing the safety 350 00:16:26.679 --> 00:16:29.159 net will catch a bad decision before it becomes a disaster. 351 00:16:29.919 --> 00:16:33.000 It essentially lets the rover think creatively about a path 352 00:16:33.320 --> 00:16:36.440 while the digital twin acts as its conscience. 353 00:16:36.720 --> 00:16:38.200 That's a great way to put it. So we know 354 00:16:38.240 --> 00:16:40.440 how it works. But let's talk about the why. Why 355 00:16:40.440 --> 00:16:42.159 do we even need this? Is it just because it's 356 00:16:42.200 --> 00:16:45.120 cool tech? Or is there a more practical reason for 357 00:16:45.159 --> 00:16:46.639 this push towards autonomy. 358 00:16:46.960 --> 00:16:50.039 There is a very practical, very urgent reason. It all 359 00:16:50.039 --> 00:16:53.480 comes down to two things, efficiency and scientific return. 360 00:16:53.679 --> 00:16:56.759 Okay, we have a quote here from the NASA administrator, 361 00:16:56.840 --> 00:16:57.639 Jared Isaacman. 362 00:16:58.080 --> 00:17:01.879 Yes, he said this broad how we will explore other worlds. 363 00:17:02.360 --> 00:17:06.200 He pointed out that autonomous technologies are absolutely essential for 364 00:17:06.319 --> 00:17:12.559 operating more efficiently and for responding to challenging terrain without 365 00:17:12.559 --> 00:17:13.039 waiting for. 366 00:17:13.000 --> 00:17:16.480 Earth responding to challenging terrain. That's an interesting phrase because 367 00:17:16.559 --> 00:17:19.400 right now, if the terrain gets tough, the humans on 368 00:17:19.440 --> 00:17:22.200 Earth slow everything way way down. They take smaller steps, 369 00:17:22.240 --> 00:17:23.519 they get more cautious. 370 00:17:23.279 --> 00:17:26.000 Exactly, and that slows down the science. If you have 371 00:17:26.000 --> 00:17:28.240 to wait twenty four hours for every thirty meters of 372 00:17:28.279 --> 00:17:31.799 progress because the ground is tricky, it takes years to 373 00:17:31.839 --> 00:17:34.759 get to the interesting geology over the next ridge. The 374 00:17:34.880 --> 00:17:38.039 rover has a finite lifespan. It's nuclear power source, the 375 00:17:38.160 --> 00:17:42.200 RTG slowly decays over time, every day wasted waiting for 376 00:17:42.279 --> 00:17:43.880 commands as a day of science loss. 377 00:17:43.920 --> 00:17:47.359 Forever Vandy Verma, who is a space roboticist at JPL, 378 00:17:47.440 --> 00:17:49.920 she broke this down into three pillars. I think this 379 00:17:49.960 --> 00:17:52.039 is a great way to visualize what the AI is 380 00:17:52.079 --> 00:17:53.359 actually doing for the mission. 381 00:17:53.519 --> 00:17:56.240 Yes, she outlined the three core functions of off planet 382 00:17:56.319 --> 00:17:59.920 driving that AI enhances. First, there is perception seeing the world, 383 00:18:00.400 --> 00:18:03.880 seeing the rocks and riffles, but not just taking a picture, 384 00:18:03.960 --> 00:18:06.920 understanding what is in the picture. This is where that 385 00:18:07.000 --> 00:18:10.680 vision language model shines. It perceives the environment more like 386 00:18:10.680 --> 00:18:12.960 a human geologist would with context. 387 00:18:13.000 --> 00:18:14.079 Okay, that's pillar One. 388 00:18:14.200 --> 00:18:18.599 Second is localization, which is knowing where you are, knowing 389 00:18:18.680 --> 00:18:21.160 exactly where you are on the map. And that sounds simple, 390 00:18:21.400 --> 00:18:24.480 but you have to remember on Mars you don't have GPS. 391 00:18:24.480 --> 00:18:27.359 There are no satellites pinging your phone. The rover has 392 00:18:27.400 --> 00:18:29.640 to figure out where it is by looking at the 393 00:18:29.720 --> 00:18:33.119 landmarks around it, the hills, the craters, and matching them 394 00:18:33.160 --> 00:18:36.240 to the orbital maps. It's an incredibly complex calculation. 395 00:18:36.400 --> 00:18:38.119 That is a good point. I never even thought about 396 00:18:38.119 --> 00:18:42.000 the lack of GPS. It's true dead reckoning old school navigation, 397 00:18:42.200 --> 00:18:42.480 it is. 398 00:18:42.839 --> 00:18:45.319 And the third pillar she mentions is planning in control. 399 00:18:45.519 --> 00:18:49.079 That's deciding the safest, most efficient path and then executing 400 00:18:49.079 --> 00:18:51.880 the commands to follow it. Firma's point is that AI 401 00:18:51.960 --> 00:18:55.759