AI Takes the Wheel: Autonomous Driving on Mars

Welcome to Bedtime Astronomy. Explore the wonders of the cosmos with our soothing Bedtime Astronomie podcast. Each episode offers a gentle journey through the stars, planets, and beyond, perfect for unwinding after a long day. Let's travel through the mysteries of the universe as you drift off into a peaceful slumber under the night sky.

Welcome back to the show. It is Monday, February second, twenty twenty six, and I have to say, looking at the date today, it feels like we are finally living in that future we were always promised and the you know, the paperback sci fi novels of the nineties.

It certainly does. Especially when you look at the news we're unpacking today, it feels less like a standard press release, yeah, and more like a lost chapter from an Isaac Asimov book exactly.

I mean. Down here on Earth, we've become pretty jaded about artificial intelligence over the last few years. We see the headlines about self driving taxis navigating San Francisco.

Fog or AI writing generic marketing emails that all sound.

The same, right, or generating that weird, slightly hallucinated art with seven fingers. On one hand, it's almost become background noise. We sort of take it for granted. But the milestone we are discussing today takes that technology and well literally puts it on a completely different.

Planet is a massive shift in context. We aren't talking about a chatbot helping you organize your calendar. We're talking about a nuclear powered explorer navigating the most hostile, unforgiving terrain imaginable entirely on its own.

And that's the core news event. It's something that frankly changes the game for space exploration. NASA's Perseverance Rover, which has been roaming the Red Planet for years now, has successively completed its first ever drives planned entirely by artificial intelligence. And I don't mean assisted driving where a human is sort of hovering over a brake pedal. I mean the AI took the wheel right.

And just to be clear from the outset, because I think people hear AI and rover and they immediately think of the old collision avoidance systems.

We've had for a while, sure, like the sensors in your car.

Exactly, This isn't the rover just avoiding a sharp rock while a human holds the steering wheel. This is the AI planning the route, making the decisions, and then executing the drive without a human looking over its shoulder in real time.

That's the hook for me. I want you to imagine driving a car, but the steering wheel is one hundred and forty million miles away, and instead of trying to steer it yourself with the twenty minute delay, which is physically impossible, you just hand the keys to the onboard computer and say you figure it out, get me to that mountain range.

And that's what NASA has effectively just done. It's incredible, it is yeah, and it's a very active way to put it. To guide us through this. We're looking at a comprehensive report that was released today February two from NASA and the Jet Propulsion Laboratory or JPL.

So this is the real deal. This isn't just speculation, Oh no, this is it.

This comes straight from the Rover Operation Center. We're analyzing data that's been verified by the Mars Connissance orbiter and there are details on a specific collaboration with the AI company Anthropic, So we have a lot of credible, high level data to sift through. This isn't a roadmap. This is operational history being made right now.

Okay, so our mission for this discussion is pretty clear. We need to unpack how we got from generative AI writing bad poetry on the Internet to navigating bedrock and craters on Mars. We need to understand the technical leap here.

We need to analyze that delta because for the last twenty eight years of Mars exploration, humans have been very very much in the loop. This represents a move from human joysticking, or at least the illusion of joysticking, to true robotic agency.

And we definitely need to talk about the safety aspect because if my computer crash is here in the studio, I reboot it and maybe lose a word document. If a rover crash is on Mars, that's a billion dollar mistake that you can't undo. Right, So we're going to talk about the digital twin technology that keeps it safe, which is a fascinating concept in itself.

It's a fascinating layer of protection. Honestly, it's probably the only reason the engineers at JPL can sleep at night.

All right, let's get into the milestone itself. Let's start by decoupling the human driver from the machine.

Right, So let's look at this specific event. This all happened very recently. The breakthrough drives occurred on December eighth and December tenth, twenty twenty five.

Okay, so just a couple of months ago. And to set the scene visually, where exactly is Perseverance located? Right now? It's situated at the rim of Jezuro Crater, which, from what I understand, is not exactly a paved parking lot. It's not the Bonneville Salt Flats where you can just you know, throttle down and go.

Far from it.

No.

Jezro is scientifically fascinating, but logistically it's a nightmare. You have steep slopes, you have patches of loose sand that act like traps. You have shirt bedrock outcrops that can chew up the wheels. It is a complex, treacherous environment for a human to navigate, let alone a machine doing it completely solo.

I love the analogy of a child taking their first steps, you know that feeling you let go of their hand and they wobble a bit and you just hold your breath, hoping they don't face plant. Except in this case, the child is a one ton nuclear powered robot and the parents are millions of miles away. Looking at a twenty minute delayed video feed.

That's the tension exactly, and the steps it took were significant. This wasn't just inching forward to test the waters.

This wasn't a baby step.

Not at all. On that first drive on December eighth, Perseverance drove six hundred and eighty nine feet that's about two hundred and ten meters.

That's more than two American football fields on its own in one gup exactly.

And then just two days later, on December tenth, it beat that record. It drove eight hundred and seven feet, which is two hundred and forty six meters.

So why are those numbers so significant? Because I can hear a listener saying, okay, eight hundred feet. My commute is twenty miles, My roomb is eight hundred feet in my living room. Why is this such a big deal for MASA.

You have to look at it in the context of Martian exploration history. These aren't just inches. In the past, Every single meter, every foot had to be meticulously planned by a human to have the rover cover that kind of distance autonomously in complex terrain. It's a quantum leap.

Quantum leap.

It proves the system isn't just cautious, it's capable. It's traversing distances that would typically take days, sometimes weeks, of back and forth communication and planning. It's a fundamental shift in the operational tempo.

Okay, so before we get too deep into the how, I think we need to bust a myth I alluded to it earlier, the status quo. I think when most people picture NASA driving a rover, they picture a guy in a blue polo shirt with a headset and a joystick.

Oh yeah, video game image.

Right, watching a screen turning left and right in real time like they're playing a video game.

That is the classic Hollywood image, and it is completely fundamentally wrong. It is physically impossible.

Because of the speed of light. Right, it's the ultimate speed limit. You just can't get around it.

Precisely, Mars is on average about one hundred and forty million miles away. That's one hundred and twenty five million kilometers. Even at the speed of light, radio signals take anywhere from five to twenty minutes to get there, and then another five to twenty minutes to get back. It varies depending on where the planets are in their orbits.

So let's play that out. If you're the driver and you saw a cliff coming up on your screen and you hit the brakes.

The rover would have fallen off that clip twenty minutes ago. You cannot drive live. It's like trying to drive a car while looking at a photo of the road that was taken ten minutes ago. If you see a pedestrian in the photo, well, you've already hit them.

So how have we been doing it for the last twenty eight years since Sojourner back in the nineties.

It's a painstaking process. We call it the human in the loop workflow. Typically, human rover planners sitting at JPL in California download all the latest images from the rover and from satellites. They analyze the terrain manually. They look at the rocks, the sand, traps, the slopes. They put on three D glasses literally stereoscopic glasses, and stare at serial images to judge depth and distance.

They're basically staring at foot photos and drawing a line on a map like go here, then turn left thirty degrees, then stop.

Essentially, yes, they sketch a route using waypoints. These are specific coordinates the rover travels to, and because humans are cautious and because they can only see so much from static images, these waypoints are usually spaced pretty close together.

How close are we talking.

Usually know more than about one hundred feet to maybe three hundred and thirty feet apart on a really good day, with clear terrain, that's thirty to one hundred meters. They have to be short mops to ensure they aren't sending the rover into a hazard they didn't catch in the photos.

So the old way is move a little bit, stop, take a picture, send it all the way to Earth, wait for a whole team of humans to look at it, sleep on it, draw new lines, send the command back, and then it moves a little bit more. It sounds excruciatingly slow it is.

It's incredibly stop and go. It relies entirely on the Earth based cycle. If the humans are sleeping, the rover is sleeping. If the data takes a while to download via the deep space network, the rover sits there. The rover spends far more time I'm waiting than it does driving.

And when you have a limited mission lifespan, that's just lost time, wasted opportunity.

It's terribly inefficient. And this new system, this new AI, it blows that limitation completely out of the water.

It decouples the driving from the human schedule.

It decouples the rover's progress from Earth's rotation. The rover can make decisions on the fly. It doesn't need to ask for permission for every meter. It just needs a destination.

Let's get into the technology, because this is the part that I think will really surprise people. We aren't just talking about a basic collision avoidance system like you have in a modern car.

Right.

This isn't just a sensor beeping if you get too close to a wall.

No, no, this is much more sophisticated. The report identifies the specific technology as generative AI using vision language models.

Okay, stop right there, vision language models. When I hear generative AI, I think of chatbots. I think of asking a computer to write me a recipe for lasagna or a sonnet about a toaster.

Right, And that's what most of us think of. It's the same underlying architecture, it's the same kind of neural network, but it's applied in a completely different way.

So how does a language model drive a rover. Is it talking to the rocks? Hello?

Rock?

Are you friendly? What's happening here?

It's a great question, and it creates a bit of cognitive dissonance. But think about what those models are actually doing. They're processing vast amounts of information. In the case of a chatbot, that's text and finding patterns.

Okay.

In this case, the initiative was led by JPL's Rover Operations Center, the ROC, in collaboration with Anthropic.

The makers of claud Ai.

Correct, they're using claude Ai models. Now, a vision language model doesn't just process text. It processes images as if they were a language. You can look at a photo and understand the context of what is in it.

So it's not just seeing pixels. It's seeing rock, sand cliff. It's assigning meaning to the visual data exactly.

It analyzes imagery. And the crucial part here is that it uses the exact same visual data that the human planners use. It's not using some secret new sensor that we don't know about. It's looking at the same maps, the same photos and deciding independently where to put the wait points.

You used to term there semantic understanding explain that distinction for us because that sounds important.

It's the key difference. Traditional computer vision, like in older systems, looks at geometry. It sees a bump, it sees a drop, it sees an obstacle, but it doesn't necessarily know what that obstacle is. Is it a soft bush or a granite boulder. A vision language model can look at a patch of ground and say that is a sand ribble. Sand ripples are dangerous avoid, or that is bedrock. Bedrock

It's a huge difference. It's the difference between seeing a shape and knowing it's a stop sign, and you have to break let's break down the inputs. What is the AI actually looking at to make these decisions.

There are three main buckets of data that report highlights. First, it's using high resolution orbital image. This comes from the high rise camera.

That's on the Mars reconnaissance orbiter flying overhead right so.

It has the bird's eye view the macromap. It sees the layout of the land from space, the big picture. Second, it uses terrain slope data from digital elevation models, so it knows the three D shape of the ground, where it's flat, where it's deep, where the cliffs.

Are, got it the lay of the land.

And the third is JPL's own surface mission data set. This is the historical knowledge, the context of the mission itself. It's effectively the memory of where the rover has been, what kind of terrain it has encountered before, and what the overall mission parameters are.

So it takes all that data, the map from above, the three D model of the ground, and its own memories and does what.

It makes decisions. The report specifically lists the features the AI is trained to identify. It looks for bedrock, it looks for outcrops, it looks for hazardous boulder fields, and very importantly, it looks for sand ripples.

Sand ripples sounds so harmless, but on Mars they are basically quicksand traps for rovers right, they are deadly.

The Spirit rover, one of the previous generation, eventually met its end, effectively getting stuck in soft soil that looks solid. If you get stuck in a sandtrap on Mars mission over, you can't call a tow truck wow. So the AI identifies all these features, weighs the risks, and then generates a continuous path. It places those wait points, those fixed locations where the rover can stop for new instructions completely on its own.

That is just wild. It's basically doing the job of a highly trained, very cautious NASA engineer, but it's doing it right there on the processor in real time.

Well, the processing is fascinating. Yeah, but I know what you're thinking, and I know what the listener is probably thinking right now.

That trust oh absolutely. I mean we've all seen AI make mistakes. We've seen them hallucinate facts. We've seen chatbots and vent court cases that never happen. What if this AI hallucinates a road where there's actually a crater.

That is the billion dollar question. It's the thing that keeps engineers up at night. And this brings us to the next critical piece, the safety net.

Right because NASA isn't known for being reckless. They don't just upload code and hope for the best. The Silicon Valley motto of move fast and break things doesn't really work when breaking things ends the entire space program.

No, they are incredibly and rightly risk averse. The solution they developed is something they call the digital twin.

I love this concept. It sounds very cyberpunk. Explain what a digital twin is in this context.

It's exactly what it sounds like. Yeah, it is a virtual replica of the Perseverance Rover, but it lives inside a supercomputer at JPL in California. It is a physics perfect simulation of the rover and the Martian environment it's in.

So before the real rover moves an inch on Mars, the.

AI generates the plan. It says I want to drive here, here, and here. But instead of just beaming that command straight to Mars, the engineers feed it into the digital twin.

First they simulate the drive.

They do more than just simulate it, They stress test it to an unbelievable degree. The report states they verify over five hundred thousand telemetry.

Variables five hundred thousand half.

A million data points for every single proposed drive. They are checking everything wheel traction, suspension tilt, the power being drawn by each motor, the thermal limits on the electronics, the currents. They run the AI's plan through this digital gauntlet.

So if the AI says, drive over that sketchy looking rock, the digital twin simulates it and flags a warning like warning suspension damage is ninety percent likely or warning tilt exceeds safe threshold by fifteen degrees.

Precisely. It catches the hallucinations, it catches the risky maneuvers, it catches the small miscalculations, and only after the digital twin confirms the drive is safe green lights Across all half a million variables are the plans packaged up and sent via NASA's Deep space network to the physical rover on Mars.

So it's a hybrid model. It's not fully autonomous in the sense that the AI can do whatever it wants.

It is. It's AI planning, digital verification by humans and their simulators, and then physical execution. It's not blind trust. It's trust, but verify on a massive, massive scale.

That makes me feel a lot better about the whole thing. It's not just a robot running wild. It's a robot proposing a plan and a very sophisticated simulation proving it works before anything real.

Happens, and that verification step is what allows them to be bold. It allows them to let the AI try things, to plan these longer, more aggressive routes, knowing the safety net will catch a bad decision before it becomes a disaster. It essentially lets the rover think creatively about a path while the digital twin acts as its conscience.

That's a great way to put it. So we know how it works. But let's talk about the why. Why do we even need this? Is it just because it's cool tech? Or is there a more practical reason for this push towards autonomy.

There is a very practical, very urgent reason. It all comes down to two things, efficiency and scientific return.

Okay, we have a quote here from the NASA administrator, Jared Isaacman.

Yes, he said this broad how we will explore other worlds. He pointed out that autonomous technologies are absolutely essential for operating more efficiently and for responding to challenging terrain without waiting for.

Earth responding to challenging terrain. That's an interesting phrase because right now, if the terrain gets tough, the humans on Earth slow everything way way down. They take smaller steps, they get more cautious.

Exactly, and that slows down the science. If you have to wait twenty four hours for every thirty meters of progress because the ground is tricky, it takes years to get to the interesting geology over the next ridge. The rover has a finite lifespan. It's nuclear power source, the RTG slowly decays over time, every day wasted waiting for commands as a day of science loss.

Forever Vandy Verma, who is a space roboticist at JPL, she broke this down into three pillars. I think this is a great way to visualize what the AI is actually doing for the mission.

Yes, she outlined the three core functions of off planet driving that AI enhances. First, there is perception seeing the world, seeing the rocks and riffles, but not just taking a picture, understanding what is in the picture. This is where that vision language model shines. It perceives the environment more like a human geologist would with context.

Okay, that's pillar One.

Second is localization, which is knowing where you are, knowing exactly where you are on the map. And that sounds simple, but you have to remember on Mars you don't have GPS. There are no satellites pinging your phone. The rover has to figure out where it is by looking at the landmarks around it, the hills, the craters, and matching them to the orbital maps. It's an incredibly complex calculation.

That is a good point. I never even thought about the lack of GPS. It's true dead reckoning old school navigation, it is.

And the third pillar she mentions is planning in control. That's deciding the safest, most efficient path and then executing the commands to follow it. Firma's point is that AI streamlines all three of these pillars simultaneously. It perceives faster, it localizes more accurately, and it plans more aggressive paths more safely than the old stopping go method.

And there's a bonus here too. It's not just about driving, right. The report mentions that the AI is also helping with the science itself.

This is where it gets really exciting for the researchers back on Earth. The rover takes thousands upon thousands of images. A whole team of humans can't possibly look at every pixel of every image with the same level of scrutiny.

We get tired, we get distracted, we miss things, we need coffee breaks, we blink exactly.

But the AI doesn't get tired. The report says it can scour huge volumes of rover images. It can flag interesting surface features, maybe a strange rock formation or a discolored patch of soil that might indicate a specific mineral deposit that a human might have just scrolled past.

So the AI acts as a scout, a science scout.

It acts as a primary filter. It filters all the noise so the human scientists can focus on the discoveries. It basically taps them on the shoulder and says, hey, you should really look at this weird rock over here.

That changes the dynamic completely. Instead of humans telling the rover what to look at, the rover is not telling the humans what they should be looking at.

It's a partnership. It's a genuine collaboration, and it makes the science return per day, per dollar much much higher. You aren't wasting time looking at boring dust. You are jumping straight to the anomalies.

So let's look to the future beyond Jesero Crater. Because if this works here, surely they aren't going to stop with just one rover. This has to be the plan for everything going forward.

Oh absolutely, this is just the proof of concept. The vision is to scale this up significantly.

We talk about driving two hundred maybe two hundred and fifty meters what's the next goal. What are they aiming for?

Kilometer scale drives. Vanni Varma talks about minimizing operator workloads so the rovers can handle long all distances completely on their own. Imagine a rover that you tell go to that mountain range five miles away, and you don't talk to it again for a week. It just reports in when it gets there.

That would exponentially increase the amount of Mars we can explore, cover entire regions, not just single craters in a single mission.

It would, but it's not just for rovers. Matt Wallace, who is the manager of JPL's Exploration System's office, talks about edge applications.

Edge applications, what does that mean?

Computing at the edge, meaning right there on the device itself, not in a server back on Earth. He's talking about expanding this tech to helicopters, to drones, and to other surface elements.

We've seen the Ingenuity helicopter, which was absolutely amazing, but imagine a whole swarm of autonomous drones mapping a canyon without any human input.

Precisely flying on Mars is incredibly difficult. You don't have time for a human pilot to correct for a gust of wind. The atmosphere is less than one percent as dense as Earth's, the winds are unpredictable. The AI needs to handle stability and navigation instantly. You need that split second reaction time that only an onboard AI can provide.

And Wallace brings up this beautiful concept. He calls it collective wisdom.

This is fascinating. He talks about t t's training these AI systems with the knowledge of NASA's best engineers, scientists, and even astronauts.

What does that mean in practice? How do you do that?

You are essentially taking the brain power, the experience, the intuition of the best human explorers and baking it directly into the AI model. You feed the model every drive decision ever made by a human planner. You show it every rock a geologist ever flagged is interesting. You teach it what a safe slope looks like according to the most experienced driver.

So the AI isn't starting from scratch. It's standing on the shoulders of well human giants.

It creates a system where the rover drives with the caution of a veteran engineer and the curiosity of a lead scientist. It democratizes that expertise and puts it inside the robot. It's preserving the institutional knowledge of NASA and exporting it to another planet.

And the ultimate goal of all this where does this road lead.

It leads to establishing the infrastructure for a permanent human presence beyond.

Earth, first on the Moon with the Artemis.

Program on the Moon, and eventually taking the US to Mars and beyond.

This is the key takeaway for me. This AI isn't about replacing humans, It's about building the road for us to get there exactly.

We can't send humans to Mars safely if we don't have autonomous systems that can maintain the habitat, scout the terrain, and ensure safety when we aren't looking. That communication lag makes manual control not just inefficient, but dangerous for life support systems. If an oxygen generator fails, you need an AI to fix it now, not twenty minutes from now, when the alarm finally reaches Houston.

We need machines that can think for themselves to keep us safe.

That's the only way it works.

So let's wrap this up. We've covered a lot of ground pun absolutely intended we have. Let's just summarize the key points for everyone listening.

First, the headline Perseverance has successfully used generative AI, specifically anthropics claud to navigate the Jazuro Crater on Mars completely on its own.

It drove over fourteen hundred feet total across two in December twenty twenty five. This is a massive leap from the inch by inch crawls of past missions.

And it did it safely. The entire system is backstock by a digital twin at JPL that verifies over five hundred thousand variables before a single wheel turns on the Red planet.

This effectively signals the end of the joystick era, which was really a myth anyway, and the beginning of true, meaningful autonomy in deep space exploration.

It's a turning point. We are no longer micromanaging our robox from one hundred and forty million miles away. We are empowering them to be our proxies.

And that leads me to my final thought. And I want you to chew on this, and I want everyone listening to chew on this as well.

Oh fork.

We talked about that concept of collective wisdom. We talked about the rover flagging its own science targets. So if the rover can now perceive the world, localize itself, plan its own path, and decide for itself what is scientifically interesting enough to show us? At what point does it stop being just a tool?

That is the question, isn't it?

At what point does the rovers start being a remote controlled car and start being a partner exploring with us? And if it flags a discovery, if it's the one that finds the fossil or the evidence of past water, or you know, life, whose discovery is it? Is it the scientist back in Pasadena who looks at the photo, or is it Perseverance's discovery?

It blurs the line between creator and creation. As the machines get smarter, the credit gets harder to assign. But perhaps, you know, maybe that's the point. We are extending our consciousness to another world. We are building our successors, our partners in expiation.

It's a fascinating time to be alive and a fascinating time to be watching the stars. Thank you for joining us on this exploration of the future of Mars.

It was a pleasure.

We'll see on the next one.

Keep looking up.

The school system