Testing the Safety of Self-driving Vehicles by Simulating Perception and Prediction

We present a novel method for testing the safety of self-driving vehicles in\nsimulation. We propose an alternative to sensor simulation, as sensor\nsimulation is expensive and has large domain gaps. Instead, we directly\nsimulate the outputs of the self-driving vehicle's perception and prediction\nsystem, enabling realistic motion planning testing. Specifically, we use paired\ndata in the form of ground truth labels and real perception and prediction\noutputs to train a model that predicts what the online system will produce.\nImportantly, the inputs to our system consists of high definition maps,\nbounding boxes, and trajectories, which can be easily sketched by a test\nengineer in a matter of minutes. This makes our approach a much more scalable\nsolution. Quantitative results on two large-scale datasets demonstrate that we\ncan realistically test motion planning using our simulations.\n