Context-sensitive data-driven crowd simulation

In terms of computation, steering through a crowd of pedestrians is a challenging task. The problem space is inherently high-dimensional, with each added agent giving yet another set of parameters to consider while finding a solution. Yet in the real world, navigating through a crowd of people is very similar regardless of the population size. The closest people have the most impact while those distant set a more general strategy. To this end, we propose a data-driven system for steering in crowd simulations by splitting the problem space into coarse features for the general world, and fine features for other agents nearby. The system is comprised of a collection of steering contexts, which are qualitatively different overall traffic patterns. Due to their similarity, the scenarios within these contexts have a machine-learned model fit to the data of an offline planner which serves as an oracle for generating synthetic training data. An additional layer of machine-learning is used to select the current context at runtime, and the context's policy consulted for the agent's next step. We experienced speedup from hours per scenario with the offline planner and 10 agents to an interactive framerate of 10FPS for 3,000 agents using our data-driven technique.