The reality of autonomous cars: Ask Me Anything with Ian Baldwin

How far away are we from seeing self-driving cars in our cities? In this AMA with Pulse, CTO Ian Baldwin of Moss AI fills us in on autonomous cars and recognizing unsuccessful companies from more successful ones. 

What is the thing people misunderstand most about self-driving vehicles?

That’s a great question. We have really gone to the 18-20 games in terms of what technology can support in autonomous performance. Obviously, I stand to be corrected on a lot of this but most companies are very secretive in terms of how they formulate and publish metrics but we really are in the long tail situation now in terms of trying to drive.

For example, if you take some metric like disengagements per mile, if you want to drive that down to a realistic value where you can actually operate a self-driving taxi fleet, your reliability has to be multiples in terms of performance and we’re just not there yet. 

From my perspective, there has been huge strides in terms of what the general autonomous capabilities are of vehicles. But in terms of getting something that is truly at a service level, where the end consumer is not going to be frustrated–because their vehicle’s pulling up every single time they get into the car–that still is going to take some work. 

I think the overall enthusiasm is still very high, but I just worry about how that’s going to translate over the coming years when, we attack the kind of “plateau of productivity” in terms of the Gartner hype cycle, in terms of actually delivering a product.

So self-driving cars can mostly operate now as freight in less populated areas? 

Convoy based trucking makes a lot more sense than self-driving cars. In terms of delivering goods at a higher cadence, being able to work with a semi-supervised approach makes a lot of sense. The problem is fairly constrained in that aspect and in terms of doing that over long distances, in general where you’re doing that on highways and roads that are in open sky conditions.You have very good performance from a localization system, for example. 

The lane tracking and localization systems under open sky conditions (without signal clutter) are really good in those instances. The most unstable condition is when you get into dense urban scenes where there’s a lot of different actors and there’s inadequacies in terms of what your localization capabilities are, because a lot of systems to this point still rely on being able to co-register where they are to an existing map of the area. So there’s very few companies that can on day zero deploy autonomy into a new area. 

So what that means is that in terms of expense, in general, you have to have an advanced team that will go out and map the area to some level of fidelity to enable autonomous operations. 

In addition to that, you have to figure in map decay. So, in general, there’s going to be lots of places where, due to temporal changes, the map is going to look very different on various cadences that might be seasonal, that might be daily, whatever the case is, and how can you be robust to that? 

So most companies are relying on maps to some extent. That makes life easier because you have this robust data structure but it does add a lot of complexities.

How are autonomous vehicles able to predict traffic decisions made by others? 

There was a great piece published recently about the next challenges in terms of autonomy with respect to self-driving. One of those is prediction. At some point, we believe that there’s known unknowns and then there’s the unknown unknown at some level. So we believe that we can solve the localization problem. We can get our perception system to a level of robustness that can deliver great value in terms of the economy system. 

But then the next component is that you are not somebody who just acts irrespective of what else is going on in the scene. So being able to predict what other actors are doing over a short to medium timescale such that you can plan appropriately to interact with the scene, but also take into account that your actions influence how other people react. For this, there is a very strong push in terms of trying to understand how that works. 

With advances in deep learning and massive amounts of data, we hope at some point we’ll be able to reason about not just what other people are doing, but what other people do when we take a certain action and how we can influence the environment around it. It’s a very thorny problem and the actual space, as you can imagine, in terms of trying to determine what actions we should take.

So trying to come up with concise learnable models that relate your actions on other people’s actions and how that scales is a thorny problem. 

Driving in San Francisco, in a rush hour for example, you want to get onto the freeway. If you’re very conservative the end result is that you’ll never get on the freeway. Driving in San Francisco, even as a human, I feel my blood pressure really skyrocketing because it’s such a complex environment. So you have to have some things that you are allowed to or you have to break rules or conventions at some point to be able to deliver the service and to teach the system to be cautious at some points and to be aggressive at other points.

Do autonomous cars have to live up to a higher standard?

In terms of the kind of infrastructure built into safety for commercial aviation, for example, it took many horrendous accidents, lots of introspection, and lots of procedural changes for those standards to improve. But in terms of manufacturing operations to bring about the safety level it was astounding. In the US over the last 10 years there hasn’t been a single fatal accident from the commercial sector in the lower 48, which is incredible. 

Keeping that in mind in terms of what people will expect, obviously driving has a much higher incident record than commercial aviation. But I think people are looking at it from the perspective that it has to be basically completely safe. 

In terms of your question, I think the vehicle to vehicle the V2V, and the V2I components have been drastically undertreated. The one approach is to try and retrofit autonomy to existing cities. So we teach an autonomous machine and we deploy it wherever we want.  That’s great, because that’s incredibly scalable if the conditions are the same. The alternative approach is to develop cities in conjunction with autonomy. 

The question here will be if we can legislatively and technically build better infrastructure in terms of our cities to enable, for example, superior localization for vehicles and test environments. There’s been less emphasis on this just because it’s very difficult to scale. The V2V side, because being able to communicate on an agent to agent level is going to be fundamentally important. I think there’s been good strides in terms of trying to come up with a platform for autonomous vehicle companies to be able to interact.

The one concern that they would have is proprietary knowledge and what they gain in terms of their own systems. They wouldn’t really be incredibly motivated to share that with a competitor. But at the same time people need to have access to data, particularly in terms of incidents as well as unexpected things that people do on the roads. 

What distinguishes a successful company in this space from an unsuccessful one?

Particularly with respect to Zippy, one thing that we discovered is that you have to be able to iterate quickly. You have a concept of what the product market but until the metal hits the road you have really no idea. For us it was very important to iterate quickly but at the same time we were trying to build soup-to-nuts systems. So hardware and software.

When you give to developing software in conjunction with hardware as well as trying to iterate on what your current market fit is, it means you burn through a lot of money and you do that really quickly. So, for us the lesson I learned there was that it is much better to try and get something maybe subpar, but an MVP out there, get it interacting with people and start testing your hypothesis before taking large amounts of VC capital and making strategic bets. 

From the technical side, I am not going to start another company with a hardware and software product in combination. Without a crystal clear goal in terms of product and customer, it is basically a recipe for disaster.

You May Also Like
kaj van de loo
Read More

Retaining your best team

Starting out as a developer, Kaj van de Loo always knew he had change things up to keep…