Implementation-Details---VUSFA

paper - VUSFA-Variational-Universal-Successor-Features-Approximator

This repository contains implementations of USF-A3C Target Driven Visual Navigation. The discription of each model can be found in the following discription.

1. Implementation is the baseline model of Deep Siamese Actor-Critic.

2. Modified Deep Siamese Actor-Critic by replacing pre-trianed Resnet with adding 3 conv layers.

3. Modified network in the 2nd by adding an LSTM head before the policy and value function prediction heads.

4. Universal Sucessor Feature Dependant Policy (USF-DP) A3C for Target Driven Visual Navigation.

5. Modified network in the 4th by adding a Variational Siamese Bottleneck Layers to improve the tranfer learnign ability.

6. Modified 5th network by only using single frame as the input.

The AI2Thor simulator can be found in follwing link as a hdf5 data set:

Simulator Scenes