Open Source Self-Driving with Comma AI

0:01Welcome to the Practical AI Podcast, where we break down the real-world applications of artificial intelligence and how it's shaping the way we live, work, and create. Our goal is to help make AI technology practical, productive, and accessible to everyone. Whether you're a developer, business leader, or just curious about the tech behind the buzz, you're in the right place. Be sure to connect with us on LinkedIn, X, or Blue Sky to stay up to date with episode drops, behind-the-scenes content, and AI insights. You can learn more at practicalai.fm. Now, on to the show.

0:41Welcome to another episode of the Practical AI podcast. This is Daniel Whitenack. I'm CEO at Prediction Guard, and I'm joined as always by my co-host, Chris Benson, who is a principal AI and autonomy research engineer. How are you doing, Chris? I'm doing very well today. How's it going? It's going great. I was commenting to our guests today just before we started recording that earlier this year, I was in the car with one of our engineers, shout out to Ed, and he's like, hey, have you heard about this cool thing we're driving in the car right and he's like hey there's this cool thing you can put in your car and make it like a like a ai ai assisted driving car without it being like a specific self-driving car and uh and so he he uh forwarded me the information about comma and i'm really excited today to welcome harold schaefer who is cto at comma ai welcome harold Thank you.

1:44Thank you for having me. Excited to be here. Yeah, yeah. Well, obviously, I kind of alluded to some of what you're involved with, but maybe could you give us just a little bit of background about yourself and Kama and kind of how you ended up in this spot of working on some of the things that you're working on? Sure. So Kama makes this device, like you said, that you can install in cars and gives them autonomy features that they hadn't had before. You know, things like auto steer and better ACC. So on the highway, you kind of get some level of autonomy. And, you know, the software that runs that is OpenPilot.

2:25And that's a completely open source autonomy stack for cars. By far the most popular open source self-driving stack online. And I think it's currently even the most popular robotics project on GitHub. So that's kind of where we're at. I've been working on this a really long time. I've been at Karma for nine years now. So it's basically been my entire life. I don't think there's much to say about my professional life that's not related to Karma. But I came to the US like 10 years ago, graduated and started working here. And so I've been working on OpenPilot and this type of stuff ever since. So what was when you started that journey, what was kind of the state of both like self-driving autonomy when you started things?

3:13And also like, I guess, on the commercial and closed source side and the open source side, what did that look like kind of then compared to now if you look back on that journey? So when I joined, we didn't have a product. It was just a project. And you could kind of install the software if you went through the hassle of installing like a beefy laptop in your car and installing all the power. Or you could like retrofit a phone that you'd have to do all that yourself. So that was the state the project was in when I joined. The company was pretty young at that time. George, the founder, had I think worked on it for a little over a year, maybe two years at the time.

3:52And that's the state it was in. It was a usable ADAS system, but the product side was really not that in a greater state. So that's where we were. In terms of open source, I think there is no real genuine open source ADAS product that's useful in any way other than us. That was true then. That's true now. But the commercial side has obviously changed massively in that time. 2017, when I joined, you know, highway autonomy was bad, maybe usable in some cases, but most people probably wouldn't use it because it just makes too many mistakes. It's too uncomfortable. That's obviously not true at all anymore.

4:30You know, OpenPilot is really good. We have over 50 % of miles driven with people that have our system are driven by, you know, the system and not the human. Obviously, there's Tesla FSD, which is, you know, at an even higher level of what kind of things it can do. And you can get over 90 % engagement there if you really wanted to. And you've got Waymo, which is like supervised robo taxi. That's like an actual product. You know, unclear if they're a money making product, but it is a real product that people can use. And obviously, none of this stuff existed in 2016. As for how I got into it, I think it's worth mentioning because this is kind of the journey that I just talked about is I remember when I was, I must have been really quite young at the time, but I saw the Chris Ermsson TED talk about Waymo where they had that we had a closing statement that his kids were growing up and he was hoping they would never need a driver's license.

5:26and I think there must have been like, I don't know, less than 10 at the time and they have a driver's license now. So I think that's the timeline we're talking about is from that prediction about Waymo being able to displace the need for people to have a driver's license. That's obviously didn't manifest, but we did make a lot of progress and there are some cool stuff on the market now, even if it doesn't mean that all driving is autonomous. I'm curious, as you were getting into this, what was it about this particular problem that attracted you? What was it that caught your imagination enough to grab you for such a long period of time as well and still hold you today?

6:07What was it versus all the other things out there that you might have dived into? I mean, something that George was saying a lot of the time, and I think I really resonated with me and is still true, is that self-driving was the most interesting applied robotics problem, period. It was a place where you could make products that were essentially immediately useful. You know, we don't have 100 % reliable autonomy. You need supervision for it to be useful. But people buy our products because it adds value to their lives. This is actually quite unique in robotics. There are not many use cases where this is the case.

6:38You can have some kind of hyper-specific industrial robots are obviously useful. robot vacuum cleaners. I'm a big fan of robot vacuum cleaners. There's robot lawnmowers. And that's roughly it. There's not that many places where you can do like applied AI, applied robotics in the real world. And that's why I thought self-driving was so cool. And then the other thing is I just really like the open source nature that we have with OpenPilot that we're trying to promote. I think an open source future is just generally better for everyone. Well, yeah, I'm particularly interested in this conversation because, you know, a lot of times on the show we have really interesting people on, but sometimes they're in great discussions, but sometimes they're constrained with in terms of what level of detail they're able to talk about kind of architecture and things like that naturally, because they're maybe working on certain proprietary things.

7:33And with having so much in the open source world with Kama and OpenPilot, I'm kind of interested if maybe from a high level perspective for those out there that are listening that might not have an idea of like what sort of architecture and kind of main components are part of a self-driving or autonomous system. like OpenPilot is maybe a piece of that and fulfills a role. You have the devices. Could you just give us like a mental model for how to think about how these pieces fit together? What's where and what components are needed to make the system work, I guess? So we're talking about ours in particular, right?

8:23Or just in general? Sure, yeah. So ours in particular, we have a device that you can install in the car. The device has compute and it has some cameras. and then some other sensors like GPS and IMU that are not necessarily that needed, but they're in there. That device then runs some machine learning models by looking at the road. That tells you where roughly to drive. And concretely, that is a longitudinal acceleration and a curvature of the road. So that's like an angle of the steering wheel. I'll talk about how that works internally later. But in runtime, this is really all that happens is there is just a machine learning model that takes in the video input and outputs those actions of acceleration and curvature.

9:07And then that goes to some API that, you know, we kind of develop that interfaces with all these different car models. So we reverse engineer the canvas of the car. And so we can understand what messages we need to send to command, steering, gas and brake, and do all the auxiliary stuff of like the engagement state, you know, all those kind of things that are needed to do autonomy in a car. So that's basically what's happening at runtime. We just have a device. It runs some models. It outputs the actions to take. And then there's a car API layer that, you know, is reverse engineered for every type of car we support that sends messages on the canvas.

9:44to do the certain actions. Then those models themselves that we train, so we've been kind of talking about end-to-end training for a long time. So they're end-to-end in the sense that they take in video and they output actions. There's no intermediate space. So, you know, our cones detector, traffic lights detected, that all doesn't exist. And very high level, our training stack looks like we have hundreds of millions of miles of humans driving. We sample some of that. we teach the model okay if a human's in this situation this is the most likely trajectory they're going to take if you train that directly you get a system that doesn't really work this is like a well known machine learning problem that it's not necessarily one of them understood but you can't just do imitation learning and expect things to work in the real world you need to expose the model during training to mistakes and show how to recover from those mistakes And we do that by training in a simulator where we can introduce, you know, going off the center of the lane line and then we can supervise, OK, this is how you would recover from this.

10:46That's very high level how things are trained and how things work. I don't know if there's anything I missed there or any questions we can discuss there. Now, I was just wondering, just on that note, the open pilot project, that is part of that, I guess, the software stack that interfaces with the models and executes kind of the policy and interacts with the car API? Or does that fulfill a different role? Because it is kind of more broadly a project related to kind of autonomy and robotics generally. Or am I misunderstanding? Yes. So, I mean, the goal of OpenPilot is to be general robotics, but we're clearly focused on driving right now.

11:29But a lot of things that are written in OpenPilot are things like, OK, well, there's a UI and then there's a localizer, which is written in classical code that takes in all these input sensors and makes the best estimate of the current motion of the device. You know, that's robotics for driving. Then there's, like you said, there's this whole layer of interfacing with the car. So that's also part of OpenPilot is how you communicate with the car and manage like the state machines. And then in general, it's also an operating system, right? So there's many different threads and processes running and those intercommunicate and those need to get managed.

12:01And so that's OpenPilot. But a lot of the decision making about what happens and how to control the car essentially happens inside the neural network. It's pretty cool. I'm curious, I think the space that you've chosen in terms of addressing the need and, you know, where there is, you know, vehicles without autonomy capability altogether, traditional, below you and then kind of above you, there are the fully integrated, you know, vehicles. You know, we talk about things like Tesla's and their competitors, where they were the whole, you know, the hardware software is all working together for the whole vehicle.

12:37And that gives you another level of capability. How do you target the level of capability that that Kama is addressing in terms of when you're doing this? this, you know, what is kind of an add-on to an existing vehicle that doesn't have any autonomy capability. Like, how do you think about the problem of what you can add value to by bringing autonomy capability? And what also, like, what's the constraint? What's too much without having a fully integrated platform from the get-go, you know, that you're manufacturing out of a factory? So just to be clear, our mission is to solve self-driving cars while shipping intermediaries.

13:20And now that, you know, we're kind of, we're kind of evolving that to solve robotics. So that is what we see everything as we're trying to build solutions that, you know, genuinely do solve the AI and the applied robotics problem in a real and genuine way. That's the starting point of how we think about it. It's not that we think like, oh, we don't really start from the product or the user side. But given that starting point, we then think, okay, if we want to make progress on the robotics problem, on self-driving cars, how do we have intermediaries that make sense? We don't want to be doing research in a lab and saying like, oh, we're never going to release everything until it's done.

13:58We want to make progress and at the meantime, be able to ship useful features. So we really focus on how do we create some kind of end-to-end solution that's scalable and can be the most intelligent driving agent that can drive better than a human. But then we just kind of see, okay, given this system, can we apply this in a way that's useful to people? And then we kind of figure out, you know, which cars does this work well enough on that can be supported? Because you were talking about constraints. Like there are some cars that have certain constraints that make them unusable for open pilot.

14:31And so then that's not as interesting. Does that kind of answer your question? I'm not sure. I think so. Yeah, definitely. And I'm glad you expanded on that. I may have had too narrow a vision in my understanding of what you were addressing there. So I like the way that you are kind of iteratively solving that problem in the large. Yeah, so I mean, the distinction is, I think we're trying to make incremental steps towards some late stage solution that also have incremental usefulness to people. And I think that's kind of where we differ from a lot of other people. I think a lot of other people are like, you know, everything or nothing kind of approach.

15:10And I think that's just generally a bad approach. which I think we want to make money along the way. We want to pay our own bills. We run our own data center. And, you know, those limitations obviously mean that we run with 100x less compute in our data center than, you know, Waymo or Tesla. And that obviously has some side effects. But I think in the long run, this isn't really a big deal. As long as we're making progress towards some solution and shipping useful products, you know, I think the long-term future looks bright. And I'm not sure, you mentioned a couple times this kind of idea of end-to-end, which you mentioned was, you know, part of what you've been talking about for some time.

15:47Some listeners might not totally get kind of the implications of that. So if you could maybe talk through like an end-to-end model, as in your case, and how it might compare or contrast to other approaches that maybe have been used in autonomy that wouldn't be considered end-to-end solutions? Sure. So, yeah, end-to-end, it kind of is a matter of interpretation. When I say end-to-end, what I mean is we have examples of humans driving competently by just recording them. That data contains the information about how to drive, and we want to let a machine learning model distill that information and learn how to drive.

16:30So directly taking in just raw sensor data and outputting a policy that looks like that or some really good subset of that. In contrast, something that would not be end-to-end at all, for example, is something like some kind of segmentation, semantic segmentation network that detects where all the lane lines are, detects where all the traffic lights are, then builds this huge grid. And then you write some algorithms like, okay, well, don't touch lane lines, don't hit kids and don't crash. And then that goes through some kind of optimization, you know, that's again hand-tuned by someone and produces some trajectory.

17:04we've been focused on the end-to-end approach for a long time you know end-to-end has generally made really fast progress over the last several years so basically everyone's interested in end-to-end to some extent what people have actually shipped is more some kind of mixture now Waymo is I think slightly less end-to-end they're doing end-to-end research but they still have a lot of classical detection I think that's relevant Tesla I think has a little bit on the back end, but they're also trying to shift fully towards end to end. But yeah, I think that's kind of the state of things. As you look at the kind of the trajectory that you're currently on and both, you know, in terms of where you've come from and kind of where you're at now and into the near term future with the recognition that the vision is to solve, you know, autonomous driving kind of in the large to paraphrase you where do you see yourself now on that and kind of and also like in the way that you're approaching it why are you at where you're at now um you know you know on that path how are you envisioning your approach to this uh to a solution uh compared to the waymos and the teslas and the others of the world and the fact that they have different uh different approaches there.

18:22Yeah, I mean, the difference is a lot of it's just constraints, right? The reason we're so passionate about this end-to-end thing is because it requires less human effort to leverage the same levels of capability. Waymo and Tesla for a long time were able to get capability by doing things like having humans hand label things in scenes, by, you know, piping back data that showed a lot of uncertainty and then humans would label them. They had like this whole data engine thing. You know, they all have, they've used many different approaches to patch the holes in what is, I think, everyone's kind of vision of this end-to-end thing.

18:57Whereas, you know, we're constrained. We don't have that sort of, we want to be profitable. We don't have that sort of money. So we've been focusing on this idea since the beginning because we think this is the right long-term strategy. I think most people see this as the right long-term strategy. But I think because we're so focused on it in just strictly the strategy sense, We're just, you know, slightly more sophisticated, but, you know, in the capability sense, I think not quite. I think so one big innovation that we have, and I think we're the first people to do this at this point, is our models are trained in simulation, but they're not just they're trained in a machine learning simulation.

19:30So, you know, those like video generations you see like Sora, all those types of things, you know, we train our models now in a diffusion simulator. So the videos are all generated. Again, Waymo and Tesla are exploring these things, but they haven't been as focused on it. So I think they've not quite shipped this yet. And they also have a higher bar to meet before they can ship that. But for us, this is the most efficient way to make progress. So I think we're just we're slightly closer to this kind of envision of the strategy. Could you dig into that a little bit? Because I think that is a really interesting point.

20:01And I think some people have maybe on the hype side of things heard mention of, you know, a world model or or other things like that. You know, you talk about your world model in the most recent release blog post and this fact of how, you know, the agent that you've released is maybe the first to your knowledge that is kind of fully trained in this learn simulation. Could you could you just pick apart that a little bit for, you know, maybe more on the practical side for for our listeners in the sense of like, what what what do you mean by a world model? How is that problem actually, like what are the challenges associated with that problem actually creating that world model?

20:49That might be a first place to start because even before training the policy model or the model that you end up wanting to use, right? Somehow you have to, if you're going to train it in that simulation, you have to have a really good simulation. So, yeah. Yeah, so, I mean, I talked about in the beginning, this is just an assumption that we have to make. is that it's impossible to just do imitation learning. You have to do some tricks on top of that to get something to recover from mistakes and not drift out of the lane line. And everyone has kind of different strategies on how to deal with that.

21:23Like I said, I think we think the long-term strategy is you just need a really good simulator. And so our solution has always involved training things in simulation. It's just our previous simulators were like classical. They would estimate depth and then reproject depth. This has artifacts. You make some assumptions that don't hold. So you want a really good simulator that can capture the world completely. And obviously, if we think about end-to-end again, the end-to-end way to do is just tell a machine learning model, okay, try to simulate the world. The difficulty here is, okay, first of all, you need to make video that looks somewhat realistic because otherwise you have all these artifacts that can be exploited.

22:00And I mean, it's only in the last couple of years we've gotten anywhere close to that. And then the other big challenge, which is where we differ from most video generation models is if you want to make like a robotic simulator with this approach, you need it to be accurate in terms of responding to inputs. If you tell your simulator, okay, the car turns left 10 degrees, the simulator actually has to then produce video that reflects that left turn 10 degrees. It's not enough for the video to just look realistic if it doesn't respond accurately to the inputs. Let's say those are the two challenges in making a simulator.

22:37It has to look photorealistic, has to be obviously as diverse as the real world, and then has to respond accurate to inputs. So that's the challenges we were working on.

22:48The photorealistics kind of solved for us, right? A lot of people are working on this. And then I think where we kind of had to, where our stuff kind of deviates from all the published papers and stuff is trying to make it respond accurately. I'm curious, as you implement the notion of a worldview, how are you approaching it? Is it more of a training mechanism or is it something where you're doing real-time planning, you know, in route? How does it fit into the overall architecture and how are you guys using it? The world model? Yeah, yeah. Yeah, so the world model acts as a, in our case, it acts as a simulator.

23:26So you can just initiate, instantiate it on some real video. And then basically you take control in simulation and you give to the world model actions like turn left 10 degrees and then the world model will produce the next image. So that's one aspect of how it's used. Additionally, the world model also supervises the recoveries during training. So the world model that's producing these images also sees the future. Like there is some real future attached to this scenario that we're making. And so we can then introduce deviations and the world model who sees the future can also say, here is a likely trajectory to get from this deviated state that we are now to this future.

24:10Sorry, I'm trying to explain this in a way that's not extremely confusing, but it is just extremely confusing. but that is how our training training stack works is that the world model simulator produces the images and supervises the recovery trajectories that then are eventually what gets into the car and on that point of kind of the trajectory to what gets in the car one of the things that might be helpful to highlight like you mentioned like we mentioned different models the world model the model that's on the car the kind of harness around that, if you will. Am I correct? You also mentioned having a data center.

24:52Am I correct that that data center and centralized infrastructure is mainly geared towards that training and simulation? And then a lot of the real-time inference for individual cars and decision-making, does that happen on device? or? Yeah, exactly. So everything at runtime is strictly on the device. The device doesn't need any connectivity other than updates or to send data back so we can use it to train. And the data center, we don't run any user-facing services there. It's just the training data center that runs all these different training things. And yeah, just to be clear, so there are two main models involved, which is the world model that acts as the simulator that we train inside of.

25:40And then there's the agent policy, which is in comparison, a really tiny model. That's the one then that trains inside the simulation and gets shipped to the devices. Yeah, that's interesting. What have you, I guess, devices have obviously advanced. So like in addition to some of these things, like your ability to create more realistic imagery for like the simulations and that sort of thing, obviously hardware or maybe capabilities of running models within kind of edge environments has advanced over the years since you started this. Could you give us a little bit of a picture of that world and maybe what's different now versus when you started this, both in terms of what you need to do to run this sort of model in that kind of edge environment, maybe the tooling or ease of doing that now versus years ago?

26:34Sure. Actually, on our product and specifically, there's not been as much progress as in other places. We still run a relatively old chip. Because our device sits on the windshield, there's quite a limitation to how much heat we can generate. So we haven't followed really the trend in how much these other systems have increased their compute power. That's something we're now trying to solve by we're going to sell an external GPU that you can plug in that can, you know be under your seat or something that way we can kind of match the compute that things like FSD and Waymo use right now we use probably less than a hundred x a hundredth of of what what an FSD computer has there's definitely been increase in efficiency so you can you those those cars have better chips than they used to but a big change has just been that they put in more power hungry chips than they used to I think there's been more of a recognition that it's worth spending watts there.

27:26I'm curious as you're looking at that as a possibility in terms of upgrading the hardware, the processing capability on board, what is an immediate term type of capability that you would add to your stack? So compared to like, if I'm looking at the Coma website and seeing the demos that are listed there, if you could like have a quick hit on it by having extra compute, What are the things that you guys are talking about, at least in the open, that you can share with the public that would be good for that? I mean, so we just started working on this, but our tests currently indicate that it recognizes green lights in nuanced situations twice as much if we have a 10x bigger model.