Roughly 2% of agents cannot reach their goals due to mislabeling. We should make these agents static so that we don't actually have to control them. The way to do this is replay the expert trajectories and see which ones have a collision in the expert trajectory. We can overapproximate and treat all of those as experts.
Roughly 2% of agents cannot reach their goals due to mislabeling. We should make these agents static so that we don't actually have to control them. The way to do this is replay the expert trajectories and see which ones have a collision in the expert trajectory. We can overapproximate and treat all of those as experts.