LADFN: Learning Actions for Drift-Free Navigation in Highly Dynamic Scenes

With LADFN, the ego is able to perform complex behaviors that lead to drift-minimisation in dynamic as well as static scenes. The following is a non-exhaustive list of the behaviors learned by the model. These clips do not show the entire trajectories of the scenes mentioned in the paper, but only the places where interesting behaviors were observed.

Behavior-1

The ego observes traffic on both sides, while the left side of the road is feature-rich. The ego quickly speeds up and moves left at the same time. This ensures that it moves away from the traffic, while at the same time maintaining proximity with the feature-rich regions.

Behavior-2

The situation here is similar to the Behavior-1 situation, but both the traffic vehicles are on the same side. Further, the ego cannot move left directly since the traffic is blocking it. We see that the ego decides to overtake the traffic first, and then it moves closer to the feature-rich regions.

Behavior-3

This is a static scenario, i.e, there is no traffic around. The right side of the road is feature-rich in this case. We observe that the ego moves closer to the feature-rich side. However, unlike Behavior-1 and Behavior-2, there is no unnecessary acceleration/deceleration to move away from traffic.