Waymo has now driven 10 billion autonomous miles in simulation according to TechCrunch. Why you may ask yourself is there a need to do so, in simulation?
The majority of driving is simple and incident free. You, as a human driver, are basically on autopilot and by sticking to the rules and driving between the lines, in your lane, you nearly always arrive safe, at your intended destination. There are rarely any exceptions in your driving and edge cases, like ethical the baby & grandma scenario that people talk about are rare exceptions too.
A self driving vehicle will use an AI model, classical robotics and computer visions approaches to drive. What we are talking here is about training, the AI model components.
What an AI needs to learn from, is not only the normal driving cases, that it might encounter but also the edge cases. An edge case where a human, animal or expensive infrastructure is put at risk, is not feasible, in the real world, for acquiring data for the purposes of training an AI.
The AI needs to learn from data, and keep repeating experiments until it gets the scenario correct. This is best achieved in simulation to enable data to be acquired safely.
To that end, sdcar.ai has been working away in stealh mode in related technology, to assist with acquiring data, to create realistic simulations… we are still early in that journey.
Recently R&D work has been done on using Unreal Engine as a simulator. Unreal Engine is a popular set of tools, Software Development Kit (SDK) and runtime for creating near photo realistic 3D worlds for games. It is a powerful creation engine. The initial hypothesis, was that through using Unreal Engine, a vehicle simulator could be created, data acquired, an AI model trained, and that AI model used to control a vehicle in the simulator.
That hypothesis was indeed correct, and a GitHub repo simtrack1-ai-drive was released with details on how to use the simtrack1 simulator.
This simtrack1 simulator is in its early stages, and is not fully decorated. It has sufficient detail to enable the capturing of vehicle telemetry data & camera images to train an AI model that can in turn be used to drive the vehicle.
If you are so inclined, there are instructions in the repos README.md to try running it experimentally, if you have a recent graphics card.
Some ideas for future work, include:
– creating a leaders board to capture best times around the track;
– setting up waypoints, to ensure the track is actually traversed;
– installing traffic signals;
– populating road signs with speed limits;
– speeding up the game speed, with multiple instance, to enable deep reinforcement learning;
– creating zombie people and animals that freely roam to facilitate obstacle avoidance; and
– having multiple vehicles on the road at any one time to simulate traffic.
Would love feedback but please do remember that this is an early public release.
Nick Hortovanyi's blog 