Putting Self-driving Cars on the Road: Virtual Worlds and Unpredictability in Human Behaviour
This is the second post in a four-part series on the winners of the Porsche ideas competition “Mobility for a Better World”. In our last post, we covered Hamburg-based start-up Breeze Technologies with their idea for environmental sensors and analysis software for better air quality in cities.
This post is about virtual worlds and autonomous vehicles.
Our second winner, Phantasma Labs, is a Berlin-based start-up that focuses on complex simulations of urban spaces. The company was founded in 2018 by Ramakrishna Nanjundaiah and Maria Meier. With their company, they want to take autonomous driving to the next level and to do so, they have built a simulation platform that helps cars better understand humans.
The challenge: teaching autonomous vehicles how to deal with people
Self-driving cars are already here — sort of. Self-driving technology company Waymo recently announced that its vehicle fleet has driven more than 30 million kilometres on public roads.
But today’s self-driving cars are far from perfect. While they can navigate difficult terrains, they still struggle with dense urban environments and cannot handle the full range of driving scenarios. Especially the unpredictability of human behaviour is an intractable problem that machines find hard to anticipate and process. Cruising on highways is much easier for them than driving in crowded cities because they haven’t learnt yet how to properly interact with people in edge-case scenarios. Indeed, people who take liberties with the rules of the road, chaotic traffic jams and pedestrians suddenly running on the road expose the limitations of autonomous driving.
As with human drivers, there is one way for self-driving vehicles to quickly become better — practice. The deep neural networks running self-driving vehicles must be trained with very large datasets. In other words, the more kilometres the vehicles put under their wheels, the better they become. Not every kilometre is the same, however, with some more densely packed than others. Monotonous highway kilometres, for example, are typically less useful than intersection scenarios involving multiple vehicles and pedestrians. Only the latter are the “quality” kilometres in which autonomous cars learn something really new.
However, the problem here is availability. One of the reasons why autonomous vehicles have stayed in beta phase is that, according to the founders of Phantasma Labs, there is not enough quality training data. Obtaining this data is time-consuming and expensive. It would take billions of kilometres worth of driving data to test self-driving systems in a robust way on the road. And that’s where virtual worlds come in. The idea is to transfer the “quality” kilometres to the lab and the computer so that developers can run their digital cars through countless edge-case test scenarios.
The solution: generating human behaviour datasets with virtual worlds
Phantasma Labs specializes in building virtual worlds, including fully modelled versions of London and Manhattan, with which they can easily recreate various real-world scenarios. Vehicles, cyclists, pedestrians, curbs or traffic lights can be added with the push of a button. Leveraging the power of high-performance computing, they create life-like virtual environments and crowdsource real human behaviour.
Users interact both as motorists and pedestrians in these life-like simulations and Phantasma then captures their behaviours, generating behaviour models and annotated synthetic datasets. Their virtual datasets are cheaper and faster to produce than real-world datasets and encompass various edge-case scenarios. The virtual simulation also helps take vehicles off the road, thereby reducing greenhouse gas emissions.
With their datasets, which can be used in existing workflows, Phantasma Labs wants to help automakers develop self-driving systems that are capable of full autonomy. The jury praised the team for identifying a problem in need of a solution, noting that autonomous vehicles have the potential to make road traffic safer.
Many congratulations to Phantasma Labs. Well done! Next week, we’ll take a look at Ducktrain, a fully electric, automatized logistics system. Stay tuned!
About this publication: Where innovation meets tradition. There’s more to Porsche than sports cars — we’re tackling new challenges, develop digital products and think digital with focus on the customer. On our Medium blog, we tell these stories. It´s about our #nextvisions, smart technologies and the people that drive our digital journey. Please follow us on Twitter (Porsche Digital, Next Visions), Instagram (Porsche Digital, Next Visions, Porsche Newsroom) and LinkedIn (Porsche AG, Porsche Digital) for more.
