I've been working with applying SkewFit over the sawyer_push_multiobj.py environment with 3 differently colored pucks and goals with only one moved object. But I'm finding the VAE reconstructions to be very poor, even after running it over different latent representation dimensionalities, betas, num of steps before VAE training, etc. Would you be willing to share your run-time parameters to the multi-colored push environment? Or is getting SkewFit to work on that environment sort of a work in progress?
Hello,
I've been working with applying SkewFit over the sawyer_push_multiobj.py environment with 3 differently colored pucks and goals with only one moved object. But I'm finding the VAE reconstructions to be very poor, even after running it over different latent representation dimensionalities, betas, num of steps before VAE training, etc. Would you be willing to share your run-time parameters to the multi-colored push environment? Or is getting SkewFit to work on that environment sort of a work in progress?