search

jkulhanek / robot-visual-navigation

Visual Navigation in Real-World Indoor Environments Using End-to-End Deep Reinforcement Learning Official Implementation
MIT License
58 stars 11 forks source link

[Paper] Clarification on auxiliary tasks #1

Closed gwendalp closed 2 years ago

gwendalp commented 3 years ago

Hello, Thank you for making your work public. Concerning the paper (Visual Navigation in Real-World Indoor Environments Using End-to-End Deep Reinforcement Learning), I don't understand well the auxiliary task consisting in reconstructing the input image and the target image, its interest? I understood it as the function materialized by the convolution network must deviate as little as possible from the input image and the target image. I may be totally wrong. Thank you for your answer.

jkulhanek commented 3 years ago

Dear @gwendalp, The goal of these auxiliary tasks is to work essentially as an autoencoder, where the latent space is shared with the critic and the policy. The generated images should, therefore, be as close as possible to the input images (distance measured by MSE).