Learning-latent-space-dynamics

Implementation of a Variational Autoencoder for learning latent space dynamics from images to push

Data collection

The actions that are uniformly random sampled within the action space limits and the corresponding state are collected

Action-Space

Each action

$\mathbf u = \begin{bmatrix} p & \phi & \ell\end{bmatrix}^\top\in \mathbb R^3$ is composed by:
* $p \in [-1,1]$: pushing location along the lower block edge.
* $\phi \in [-\frac{\pi}{2},\frac{\pi}{2}]$ pushing angle.
* $\ell\in [0,1]$ pushing length as a fraction of the maximum pushing length. The maximum pushing length is is 0.1 m

VAE

The VAE Encoder, which maps images to a Gaussian distribution over latent vectors is implemented. The encoder outputs $\mu$ and $\log\sigma^2$ which parameterize the latent distribution.The architecture is shown below:

State Decoder

State Encoder

RESULTS

Reconstructed image

Image-based controller