Manufacturers of autonomous driverless vehicles (ADVs) are very much in a race. 1st prize? Open access to a market without competition. Testing is the final hurdle – or rather a series of hurdles – and may yet trip up a number of the front runners. Manufacturers have looked to simulations as a means of accelerating their testing processes, with Waymo recently estimating that a single day in simulation is the equivalent of more than 100 years in the real world.
Simulating cities and transport networks provides a cost efficient, safe way to test the behaviour of ADVs before deploying to the road. Accurate simulation reveals emergent variables, providing accurate insight into performance and helping to train new models and software updates.
But the complexity of the world’s road network is difficult to recreate.
From the densely populated cityscapes and winding roads of London, complete with cycles lanes, bus lanes, taxi ranks and complex one way systems, through to the uncurtailed speed of Germany’s vast autobahn network, there are myriad topographies, circumstances and complexities for a vehicle to learn. All the while they must be able to track and monitor one another, and be prepared for the unpredictability of human behaviour.
Current best-in-class technology can simulate a single 10-20 minute trip for one vehicle. Up to twenty of these simulations can be carried out concurrently – depending on the complexity of the route and events that take place – before computing costs become prohibitive. Repeating the process across an entire city and then country is a painfully laborious task. At best it results in a slow time to insight, and at worst, suboptimal validation of real-world driving.
Drawn out testing processes take months to reveal an issue. However, circumnavigating the issue by sacrificing fidelity or complexity in the simulation leads to a flawed output and doesn’t guarantee the vehicle will perform as expected or required when deployed to the real-world. Ultimately, inaccuracies in calibrating rates of different incidents/events render it difficult to assess if ADV models are safe enough for the real world.
Mapping Multiplayer Frameworks to Autonomous Vehicles
The solution may already exist, living in the seemingly distant world (or worlds) of massive multiplayer online games. Creating complex, realistic and coherent worlds that span across vast cities, planets and galaxies sits at the heart of game design. Gameworlds exist in perpetuity and serve to create a deeply immersive experience often shared by many hundreds or even thousands of players, all of whom interact with a typical human capriciousness and unpredictability. The same model can be applied to autonomous vehicle testing, where creating the realistic environments, 3D structures, dynamic populations and high-definition maps is key to accurately testing the vehicles and modelling.
So, what if instead of needing to run the same simulation thousands upon thousands of times, it was possible to run a single coherent simulation that connected a full active fleet of ADVs? One that minimised discrepancy between the simulated and real-worlds and demonstrated the complete distribution of emergent behaviours?
Implementing a multiplayer framework opens up the possibility of a long-running city-level sim engine with concurrent ADVs – potentially even a full active fleet. Highly scalable parallelised simulations will enable ADVs to drive more frequently, cover a broad operational design domain and related traffic more densely, learn faster, and all the while consume less computing resources.
Human controlled characters could be replaced by AI controlled vehicles, and the many entities that inhabit games could be replaced with complex representative traffic. Data could be synthesised from sensors in near real-time to update the surrounding environment as the vehicle moves through the simulation. Outputs will be aggregated and presented back to reveal any insight, issues or results in a significantly shorter time frame.
Modular game architectures lend themselves well to ADV training, enabling an open and extensible framework that integrates COTS/proprietary software, AI models and sensor data within an interoperable cloud-native environment. Moreover, integration to data sources such as LIDAR would not require a vast overhaul of the current systems.
Ultimately, concurrent simulations will accelerate time to insight, uncovering issues quickly and presenting key information such as number of collisions, stuck vehicles and so on in near real-time. Mapping multiplayer frameworks to autonomous vehicle testing is critical to validate different vehicle platforms and cities at scale, which has far reaching consequences for GTM and first-mover market share, proving key in accelerating the operational rollout of ADV fleets and determining who crosses the line to production first.