This repository contains implementations of USF-A3C Target Driven Visual Navigation. The discription of each model can be found in the following discription.
-
Implementation is the baseline model of Deep Siamese Actor-Critic.
-
Modified Deep Siamese Actor-Critic by replacing pre-trianed Resnet with adding 3 conv layers.
-
Modified network in the 2nd by adding an LSTM head before the policy and value function prediction heads.
-
Universal Sucessor Feature Dependant Policy (USF-DP) A3C for Target Driven Visual Navigation.
-
Modified network in the 4th by adding a Variational Siamese Bottleneck Layers to improve the tranfer learnign ability.
-
Modified 5th network by only using single frame as the input.