Offline navigation demonstrations are inherently multimodal: the same observation can justify several different actions, such as steering around an obstacle on the left or on the right. Standard behavior cloning treats these demonstrations as if they were unimodal and averages the actions. The authors of a new arXiv preprint argue that this averaging can produce unsafe behavior, because the mean of two valid actions may point directly into the obstacle.
To address this, they propose a method that identifies geometric candidate modes from the demonstrations. Rather than collapsing multiple valid actions into a single average, the model keeps the distinct modes as separate options. This preserves the structure of the demonstration data and avoids the unsafe trajectories that arise from naive averaging.
The paper is available on arXiv under the identifier 2610.20241. The abstract does not include experimental results, so the reported benefits are limited to the conceptual argument presented here.