habitat_annawan_nav
---------------------

This folder contains videos of image based navigation trials inside the Annawan apartment from the Habitat 
simulation environment. This apartment was not used for training the perceptual representations. Therefore,
these videos are intended to illustrate that the perception model traied on a single apartment generalizes 
to test apartments sufficiently well to allow perceptual memory retrieval and retrospective planning. 


# Explanations for the videos:

- The leftmost image on the videos show the current-state seen by the agent. 
The image has an FOV of 150 degrees, and a resolution of 256 x 256.

- The middle static image on the videos show the goal-state the agent is trying to reach. 
The image has an FOV of 150 degrees, and a resolution of 256 x 256.

- The layout on the right of the videos provides a top-down visualization of the navigation trial.
We emphasize that the agent only receives current and goal images as input, and the top-down layout 
is solely for visualization purposes.


# Additional explanations for the top-down layout visualizations in the videos annawan_nav_detailed_vis_1.mp4
and annawan_nav_detailed_vis_2.mp4:

- The white dot marks the position of the goal-state.

- A navigation trial ends if the agent reaches within 2-step proximity of the goal position.

- The moving dark-blue dot marks the current position of the agent, and the trailing yellow dots mark the 
past positions visited by the agent.

- The light blue trajectories that blip in and out of view visualize the restitched trajectories resulting
from retrospective planning. As the agent replans at each timestep in a model predictive control loop,
there is a different restitched trajectory in each frame.


# Additional explanations for the top-down layout visualizations in the remaining videos:

- These videos are directly taken from the monte-carlo evaluations that produced Fig.8 of the supplementary 
pdf. During these monte-carlo evaluations, visualizations weren't logged with as much detail as the previous 
two videos, but we still include them for completeness.

- The pale green dots visualize the navigation trajectory.

- A navigation trial ends if the agent reaches within 2-step proximity of the goal position.
