This is a repository for studying generative pattern models for the robot.
Now, we implemented a score-conditioned generative model (by leveraging conditional VAE) and proxy score function.
Training proxy and generator
First, we train proxy score function $f_{\phi}(x) \approx f(x)$ using offline data $D_{\text{off}} = (x_i,f(x_i))_{i=1}^{N}$, where $x_i$ is pattern data (represented as image tensor) and $f(x_i)$ stands for score from robot simulation function $f$.
Second, we train conditional VAE to attain score conditioned generator which satisfies $p(x|y,z) \propto 1_{f(x)=y}$ and $z \sim \mathcal{N}(0,1)$ using offline data $D_{\text{off}}$.
Samples pattern from trained generator and proxy
First, we samples $M$ candidates of patterns $X_{\text{candidates}}$ as: $x \sim p(x|y^{\text{max}},z)$ where $z \sim \mathcal{N}(0,1)$. Note $y^{\text{max}}$ is some high value we desire to get.
Note that $X_{\text{candidates}}$ stands for the set of sampled $x$ where $|X_{\text{candidates}}| = M$.
Second, we screen the $M$ candidates as TopK patterns based on the score evaluated by the proxy function $f_{\phi}$: $X_{\text{final}}:= \text{TopK}(f_{\phi}, X_{\text{candidates}})$.
Specifically, for each sample $x \in X_{\text{candidates}}$, (1) we assign pseudo-score using $f_{\phi}(x)$, (2) sort it based on the suede-score and (3) collect Top K samples among them.