Great Wall’s Mark Crawford explains the problems with autonomous sensors that’s holding back autonomy to Louis Bedigian.

Driverless technology sounds great on paper but automakers might find it difficult to meet their pie-in-the-sky expectations for imminent deployment.

Mark Crawford, chief engineer of autonomous driving systems for China’s Great Wall Motor Company, does not think that any of the existing autonomous sensors are ready to hit the road. He said the LiDAR currently being placed on top of vehicles is not actually automotive grade. “That’s a big problem,” said Crawford. “I think that there’s a lot of different approaches when it comes to trying to achieve Level 4 technology. Some people are approaching it from a camera vision-based perspective, which I think is really challenging. It’ll be really interesting if they can accomplish it.”

LiDAR is the preferred option among automakers that are the furthest along in autonomous development, Crawford added. “I think until we actually see that there is a reasonable, cost-effective, automotive grade, production-ready LiDAR, it’s really hard to have a whole lot of confidence that this is ready right now,” he said.

The other issue is, once you have all these sensors, how accurately will the vehicle read the environment? Can it really be on par with a human driver? Said Crawford: “I could say right now that we can do SAE Level 4 in a totally closed off environment but once we start expanding the scope, I think that’s just going to constrain how we actually deploy these systems.”

LiDAR may not be enough

Driverless cars could, ultimately, require a bit more than a spinning or solid-state LiDAR sitting on the roof. Crawford referred to both LiDAR and camera technology as “necessary but not sufficient”. He explained: “I don’t think that we have all of the advanced algorithms yet in order to clearly understand everything that’s happening within the camera space but I think potentially we could get there. It could become necessary and sufficient.”

He added that the limitations of LiDAR will prevent it from being the only solution used by automakers. “LiDAR can give you very accurate sparse information about your environment,” he said. “It can provide very precise depth information and range information but it can’t really tell you the colour of the traffic light, or some other important information that’s missing because of the sparseness of the data. Cameras can give you very dense information that can overwhelm anything that’s coming out of a LiDAR system in terms of the pure number of pixels that are being generated but we have less of a semantic understanding of what’s happening in that image frame than we do with LiDAR.”

Petabytes of data

It’s believed that several petabytes of data will be needed in order to process autonomous vehicles but that presents a major challenge for those who intend to collect and store that data. There aren’t any cloud services capable of handling the amount of information that could come from millions of self-driving cars. Cellular connections might not even be fast enough to facilitate its delivery.

That could change as the technology evolves but there might be a better solution to use instead. Said Crawford: “We’re in the research phase where you probably need to collect this data and log it so you can come back and redo it in case there are issues that pop up but, in the long-term, there’s not necessarily the requirement of saving all that data. I think we could develop ways in order to compress the amount of data that we think is required and potentially discard the rest of it.”

For now automakers and tech companies are tempted to save everything, which could be useful for deep learning or regression tests. “As we progress and the systems get more mature, it becomes less important to save every single thing that the vehicle has seen in its raw form,” said Crawford. “We can be more intelligent in terms of how we compress it but, in the short-term, I think there is this big problem in terms of managing all this data. For the long-term I think we’ll get pretty smart in deciding what data needs to be saved and what can be safely discarded.”

Going global

Chinese automakers have had a difficult time exporting their cars to other regions, particularly in the United States. Crawford attributes this to the “tremendously complicated” homologation process “where you have a vehicle that was designed to work in one particular market, and then trying to go through the process of preparing it for a different market”.

“The US has a lot more restrictions in terms of selling vehicles, so all of that has to be studied and analysed,” he said. “You have to understand, clearly, what it is that needs to be done in the US and then what modifications, if any, need to be made to the vehicle and make sure that we follow through with all of the regulatory, all the admissions, all the IP issues and so forth.”


Connected & Autonomous Vehicles

14 May 2018 - 17 May 2018, Santa Clara, USA

From vehicle electrification and infrastructure to the evolution of ADAS and vehicle automation to enhanced connectivity and new mobility models, no rock will be left unturned.