notesum.ai

Published at October 23

physics.comp-ph

cs.AI

Released Date: October 23, 2024

Authors: Max Wilcoxson¹, Qiyang Li¹, Kevin Frans¹, Sergey Levine¹

Aff.: ¹UC Berkeley

Parameter Name	Value
Batch size	256
Optimizer	Adam
Learning rate	$3\times 10^{-4}$
GRU Hidden Size	256
GRU Layers	2 hidden layers
KL Coefficient ( $\beta)$	0.1
VAE Prior	state-conditioned isotropic Gaussian distribution over the latent
VAE Posterior	isotropic Gaussian distribution over the latent
Reconstruction Policy Decoder	isotropic Gaussian distribution over the action space
Latent Dimension	8
Trajectory Segment Length ( $H$ )	4
Image Encoder Latent Dim	50