VAE-Model-based-RL-for-navigation

Model-based Reinforcement Learning algorithm for navigation using VAE and DDPG. This is a model-based RL algorithm using a VAE with an angular latent representation of the environment to perform a navigation task for autonomous vehicles.

The algorithm is inspired by model-based RL neuroscience experiments, where mice can pre-acquire a latent representation of a the map of a maze, that will accelerate the learning process when placed again in the maze. In general, for a naigation task an angular latent representation (compass) can be pre-aquired by having a look to the map. In our environment the vehcile should arrive at a target by reducing Euclidean distance and alignment with respect to the target.

In our algorithm the VAE with angular latent representation is combined with a policy network updated using Deep Deterministic Policy Gradient (DDPG).

Our algorithm outperformed state-of-the-art algorithms such as Twin Delayed Deep Deterministic policy gradient (TD3), and DDPG in performing this navigation task.

• images folder contains the enviroments, workflow diagram, and the results.
• models folder contains the different files for training the models, and the file to test them.
• weights folder contains the weight files for each network used by the test file.