Accurate Replaying of Simulations in Non-Deterministic Physics Engines

Sammanfattning

Modern game engines provide tools for game development. One such tool is simulations through physics engines. Game engines can be used for racing games, where people can race against other people's ghosts or in other areas entirely, such as autonomous vehicle testing. Both of these require reliable results, which becomes hard when the physics engines used are non-deterministic. This thesis investigated using prediction and reconciliation, used in online games, to increase the accuracy of replayed simulations and keep recording sizes low, in the non-deterministic physics engine Bullet. The game Drive Mad Versus, developed by Fancade, was used to collect recordings. When these recordings were replayed, dead reckoning with sample points was used for reconciliation and entity interpolation was used for prediction between sample points. The replayed simulation showed high accuracy through a low average divergence with an error between 10-3 and 10-1. However, replays showed high maximum divergence of around 100 to 102 for recordings which contained discontinuities in the original simulation. Levels which contained no moving objects showed a decrease of around 90% in file size. Levels with moving objects showed varying file size decreases, with the lowest decrease being 26.1% the size of the baseline recording format.