Density, Not Demos: Sertac Karaman on Robots That Fly and Drive

Lex Fridman asks Sertac Karaman a question the MIT professor is unusually qualified to answer. Karaman has built robots that fly and robots that drive. He is a professor in aerospace and mechanical engineering and a co-founder of the autonomous-vehicle company Optimus Ride. Which task is harder: flying, or driving without a human at the wheel?

Karaman takes a detour before he ranks them. Consumer drones, the applications people already have, are probably easier. That is one reason they took off, literally, a little earlier than autonomous cars. Look ahead to transportation and logistics, though, and he thinks autonomous flying is a lot harder than autonomous driving. A few machines have flown here and there. No one has yet deployed flying machines at massive scale.

Humanity has already deployed a lot of robots over the past fifty years. They work. They work on their own. They work on factory floors, in warehouses, and on Mars. They are either isolated or, even at scale, confined to places where they barely interact with people. The challenge of our time, Karaman says, is to take those vehicles and put them into places where humans are present.

That is not only an algorithms problem. Business models have to be redesigned. Architecture and urban planning have to move. Law and regulation have to catch up. Humans will not necessarily be trained to interact with the machines. They may not even be using the service, or know that the vehicle is autonomous. They are just living in environments designed for humans, not for robots.

Karaman can imagine tens of thousands of delivery drones filling the sky. The more interesting use case, for him, is transportation. You are in Boston and you want to go to New York, from the top of this building to the top of another building in Manhattan, in an hour and a half. Personal vehicles for four, six, or eight people would let you and a friend take that trip together and still have lunch. Traditional aviation will not vanish overnight. Plenty of people, he thinks, would spend an extra hundred dollars to cut a four-hour trip down to ninety minutes.

Fifty-year forecasts are almost a joke to him. We could get flying saucers: machines that ionize the air around them and push it down with magnets. He puts the odds at fifty-fifty over that horizon, which is his way of saying that long-range prophecy is not the useful conversation. The useful conversation is a band of airspace he calls agile airspace. It sits a little higher than the place where you can throw a stone, because a kid with a rock can take an aircraft down, and a little lower than the large aircraft flying several thousand feet up.

That band is underutilized. Most of the time, if you look up, you see nothing. Recreational flyers appear now and then. The technical challenge is building autonomy and making it as safe as we already insist aircraft be. The software will be a few orders of magnitude more complicated than what sits on aircraft today, and a lot of the intelligence will have to be trained, because some perception tasks have no other method right now besides machine learning.

When people hear simulation they jump to training. Karaman thinks one of the most important uses is development, and that is not new. The auto industry has used dynamic simulation for decades: you plug an embedded controller into another computer and let that computer pretend to be the world. What is new is how wild those environments can get. Cameras, in particular, can now be simulated far better than they used to be. Once a simulated camera sees the world the way a real camera would, you can explore the sensor's limits and train perception on the result.

Interoceptive sensors such as inertial measurement have always been easier to fake. Dynamics governed by ordinary differential equations are also tractable: how a car rolls on a road, how an aircraft flies. The hard part has been exteroceptive sensors, the ones that look out. Laser range finders are a little easier. Cameras and radars are tougher. Once those are nailed, Karaman says the next mountain is simulating human behavior, including how a human-driven car would act around you.

He has a test he sends to friends and to his mom: two pictures, which one is real? Built environments are hard to tell apart, because our eyes did not evolve to police apartments and offices. Put a human in the frame and the illusion collapses. People cannot always name the artifact. They just know something is off. Movies already render backgrounds because it is cheaper. A walking human that would pass the mom test is a different species of problem, because we do not have physics we can write down the way we have ray tracing for light.

You can collect data and try to match it, or you can show people this-or-that comparisons and have them label which clip looks more real. Anything that needs human labeling is eventually limited by how many humans you have and what else they have to do with their day. That difficulty leaks straight into self-driving. The most important variable on the road is the one we still cannot put into a processor.

Karaman splits autonomous driving into three waves. After the DARPA Urban Challenge, the field poured itself into mapping and localization: knowing where you are. Google was doing the same. Around 2012 and 2013 the AI revolution started telling the car what everybody else is. The missing piece is what everybody else is going to do next, and that last bit has been a real challenge.

Fridman adds a fourth layer that the industry mostly skips. The ego vehicle is not a spectator. Drive more aggressively, drive more gently, or signal in some way, and you change what other people do. It is a game-theoretic dance. Karaman agrees it is totally ignored. Predicting others is already hard. Folding in how you yourself shape those predictions makes the model explode. Without that layer the car can only adapt. It cannot negotiate.

The social texture of the road makes the dance worse. See a vehicle that is completely empty and it becomes a thing, the way a table is a thing. You can throw a backpack on a table. You can put your feet on it. Face to face with a human, people are civil. An autonomous car has to be as smart as a human and also survive being a thing. "You not only need to be as smart as humans are, but you're a thing." People may abuse it a little, and it still has to complete the trip.

Karaman has written papers on game theory and on social value orientation: some drivers are aggressive, some are not, and a robot can often tell by watching how the other car approaches. The hard question is whether the robot should answer aggression with aggression. Right now that looks dangerous, because the societal cost is high and people are not very accepting of aggressive robots. Abuse may even be the wrong word. People feel they own an environment designed for humans, and the robot has to understand that and respond.

Once robots deploy at large scale, Karaman says, you can design the system to be very efficient or very sustainable. The tradeoff will sit in the open. You can be aggressive, cut transportation delays, and raise capacity. You can be a lot nicer, let other people own the environment, and watch efficiency drop. Sustainability usually gets attached to energy and emissions. Livability is a sustainability impact too. If robots go around being aggressive and you do not want to live there, the system has failed a different test.

That choice has always been in transportation. Drivers' personalities and habits used to swallow it. As autonomy arrives, the choice becomes explicit, and once it is explicit you can start to optimize it. Then society has to answer what it values more, efficiency or the feeling of living in a nice place.

Testing sits inside the same frame. You test autonomous vehicles in human environments. The risk may be very small, and it is still strictly greater than zero. There is an innovation-versus-risk tradeoff, and if you do not test, development at least gets slower. Those trades are already being made. As the systems become ubiquitous, Karaman thinks they will really hit.

Fridman offers the usual spectrum: Waymo as the cautious Google self-driving project, Tesla as the aggressive, innovative one under Elon Musk. Karaman does not love the ranking. What matters most, he says, is an informed public. He would personally take a certain amount of risk; other people would not. Ideally they can choose. Unanimous choice is hard. Informing people is not actually that hard.

Turn the table and the two companies sit on orthogonal dimensions. Waymo, as he reads it, is a little bit of a research project. They talk about a product because there is pressure to talk about a product. In the back of their minds they are building a new engine: a system that knows where the ego vehicle is, where everything else is, what everything else will do, and how to interact with humans the right way. Autonomous vehicles are a very interesting embodiment of that engine. They want to understand it.

Tesla is doing something different. People use Autopilot. They still have to pay attention, and it works for something. Karaman figures Elon Musk as a visionary adds maybe five thousand dollars to the value of the car, and Autopilot another five. People are willing to pay five to ten thousand for that novelty, and they do not try it twice and put it away. They use it a lot of the time. The company collects money, data, and iteration in one loop.

They may end up using a similar engine, the way two companies might share something like an internal combustion engine while one builds a car and the other builds a truck. The use cases diverge. Talking about who is ahead on a single axis misses the point.

Karaman never bought the 2014 talk that fully autonomous vehicles were two or three years away. He had been on MIT's Urban Challenge team. He had seen the limits of computers and the difficulty of understanding human behavior. Optimus Ride took a market-based path: pick a certain kind of market and build a system for it. Full autonomy goes into the vehicles. Humans still operate them, just not by sitting in the seat.

Today one person operates one vehicle, whether it is a forklift, a truck, or a car. The goal is ten people operating fifty. That is not a person in VR goggles driving from far away. Karaman pictures a room more like air traffic control. Some people see the whole map. They know which vehicles are confident and which need a little help. The help is for efficiency, never for safety. With zero operators the vehicles should still be very safe, crawling at five miles an hour. If you want twenty-five, people come in. A car at an intersection might wait, inch forward to show intent, or turn left when it is clear, and it will not go until it gets the go.

The first places he likes to target are geo-fenced environments, roughly two miles by two miles, that are transportation deprived. Brooklyn Navy Yard is the example that had been publicized. Uber is too expensive there compared with denser markets, so people drive in, park, and use the car to go to lunch. An efficient, safe, sustainable system could reclaim some of that parking. Even in a small environment you are talking tens of millions of dollars. Smack in the middle of New York, billions. The deeper return is livability. A hundred years ago cities were not built this way. Close a road for a celebration and people still will not walk in the middle of it.

Those places already run shuttles, and people hate shuttles, because the driver is expensive so you glue twenty to thirty seats to one person. Customers ask Optimus Ride for a twenty-seater. Karaman's answer is the opposite: split the seats into four- and six-seaters so they travel faster. A big vehicle is slow to speed up and slow down, hard to climb into, and sparse on a route. Split the seats and the gaps shrink from a mile to half a mile. You may not even need an app. Walk in, get recognized, pick one of the places you usually go. "Take that shuttle experience that no one likes and tilt it into something that everybody loves."

Fridman asks how many Optimus Ride vehicles you would need so that anyone stepping outside in Boston or New York sees two within line of sight. Karaman treats that as a density question. If you see one vehicle and look around to see another, that is the target. Walk around looking for Teslas and you probably will not see that density except in a couple of zip codes in California. Those zip codes matter.

People call it data density or data velocity. Drive ten thousand miles around the country and you see each intersection once. Drive ten thousand miles in a confined environment and you see the same intersection hundreds of times. Predicting what people will do at that intersection gets much better. Tesla's huge dataset is still concentrated. Volume is not the same as repetition in place.

In good times in Singapore, taxis can be ten to twenty percent of traffic, and you can just grab a cab. Autonomous fleets need that kind of fraction before the benefits kick in by orders of magnitude. People will not wait half an hour for themselves. They will wait half an hour for goods. At peak the system can focus on humans; off-peak it can move packages. Once a fraction of a person is operating each car, the economics of today's vehicles start to look absurd. We build them to go on highways and pass crash tests. They can be twenty-five times the weight of their passengers. Purpose-built urban cars, in the models he runs, can get an A-to-B trip under a dollar, ninety-nine cents if you do not change destination.

Fridman puts Waymo, Optimus Ride, and Tesla back on the table and asks the question nobody likes: timelines. Musk has talked about some version of full autonomy by the end of the year. Karaman says it is really hard to guess. He does not know what chart Tesla can look at and read off December. For a machine-learning approach, extrapolating the end state without knowing what will happen next is very hard.

He has a name for the distortion: time dilation in technology prediction. A lot of things that are far ahead, people think are close. A lot of things that are close, people think are far. They look at a chaotic landscape where anything can happen in any order, clamp it, and put it into the next three years. Work that could land this year or next gets pushed a few years out because it is hard to explain. Twenty-year, maybe-in-my-lifetime problems get pulled in because nobody knows how hard they are, or even whether they are solvable. Optimists then say end of year. Realists say five years. Karaman thinks nobody actually knows.

Tesla could do this with electric cars. There was a Moore's-law-like chart to stand on. Autonomous driving does not have that chart. When there is a technology gap and you cannot even say what you are quantifying, prediction mostly fails. After five minutes he admits he has skillfully avoided the question. What he does believe in is iterated learning: being out there, running experiments, learning. Silicon Valley did that with business models. These companies are trying to do it across a gap they cannot plot.

As a leader he would rather run a one- or two-year plan, where the technology gets you close and investors can compare bets, than announce a five- or ten-year date that means nothing. Fridman throws the communications critique back: Optimus Ride is doing awesome tracking of objects and pedestrians, and the internet does not know. Karaman takes the hit. They chose stealth for a while. The company is bigger now, hitting a new era, and he wants to put the technology out so the best people get excited before they walk in for an interview.

Elon Musk has said lidar is a crutch, and that camera-based systems will define the future. Karaman is a big believer in camera-only autonomy. He will put his name on it: there will be a time when you can use only cameras and be fine. At that same time, lidar may be so cheap, or so good at making the software less compute-heavy and easier to certify, that it is stupid not to put it there.

A lot of autonomous-vehicle companies came and went with the same recipe: slap a lidar on a car, code it up, hit the button, do a demo. Lidar is easier to build with. Cameras can do the job. It is just harder, and you need real expertise. Optimus Ride has believed in computer vision from day one and also in sensor fusion. They use lidar relatively minimally. They do use it.

You can pay money for the lidar or pay money for the computer. Leave the lidar off and the system can get more expensive because you have to pour in compute. Short-range or low-resolution solid-state units, MEMS-style scanners, are not that hard to build affordably. The physics still bites. When you scan the environment you hit the speed of light, and you do not get both good resolution and range. Even so, Karaman thinks early deployments will involve lidar.

Karaman has a role in Lockheed Martin's Alpha Pilot Innovation Challenge, where teams race drones fully autonomously. Drone racing already exists as a sport: a human in goggles flies a real aircraft, and there is no AI. Alpha Pilot asks for aggressive flight with nobody on the sticks. What he wants is vehicles that go far faster than any human possibly can. Humans have biological limits on how fast they process information. Eyes to brain to hands: the delays a robot does not have to have.

His group builds systems that see at a kilohertz. A human would barely hit a hundred hertz. That is seeing the world at something like ten times slow motion. About a hundred lives are lost every day in U.S. traffic accidents, many of them familiar: you come onto a highway ramp, hit somebody, get confused, swerve, and go off the road. If the car had enough compute and a very fast camera at that instant, it could shut down infotainment, throw the silicon at autonomy, and either stop the car or finish the motor skill the human was trying and failing to complete.

The bottleneck is hardware and software together, codeveloped. On visual-inertial state estimation they sometimes hit the camera's frame-rate limit, because you want consecutive images to overlap a lot if you are tracking. The camera dumps data onto a serial link, USB or Camera Link, and copper wires can only carry so much. You hit another universal limit before you even start to compute. The fix is to put processing next to the pixels and go heavily parallel.

CPUs clock at several gigahertz and stop, partly from heat. At three gigahertz, light travels on the order of an inch, about the size of a chip. The clock signal is still walking around the die when the next cycle arrives. Transistors cannot shrink forever either; electrons start to tunnel. Karaman says we have hit fundamental limits of the universe in a lot of what we design. Rearrange the silicon so information moves in a way that matches high-speed cameras. He hopes that engine will one day be as impactful as seatbelts, and that it will land a fixed-wing air taxi on a rooftop with maneuvers no human could time. On a middling race course, a machine loses slightly over one lap and wins slightly but consistently over ten, because humans get tired and the code does not. Strategy and deception are still the human side.

Fridman's last question is the most beautiful idea Karaman has come across in robotics: an equation, an experiment, a demo, a piece of software. Karaman has worked more on control and decision making than on perception. The equation that has always baffled him is Bellman's. More than half a century ago, Richard Bellman asked how you jointly choose among several variables. There is one beautiful equation. People who do reinforcement learning still use it. You can teach it in a first course on decision making.

The same line also tells you how computationally hard the problem is. Karaman feels that a lot of his MIT research has been a fight against that cost, all the way to redesigning the chip. People now call it the curse of dimensionality. If you have a hundred variables and each takes ten values, the possible assignments outnumber the atoms in the universe. That thinking is embodied in one equation: theoretically optimal, and in the worst case hopeless, and in practice it nevertheless works.

In a lot of fields we admit how little we know. We do not know if the axioms of mathematics are consistent. We do not know the beginning of time, or our own future. Decision making has the same shape. Worst case, things are pretty hard. In practice, generally, things work. Fridman offers a metaphor: we focus on the worst case, freak out about the traumatic stuff, and the average case of a life still seems to work out. Karaman says that seems like a good metaphor.