Temporally Centered SIGReg Improves LeWorldModel Representations for Robot Policy Learning
技术动态
来源：arXiv cs.RO发布时间待核实
arXiv:2607.26924v3 Announce Type: replace-cross Abstract: Recent work on LeWorldModel (LeWM) has shown that the Sketched Isotropic Gaussian Regularizer (SIGReg) enables stable end-to-end world model learning from pixels by regularizing the latent representation toward an isotropic Gaussian. While effective for latent-space planning, the representations learned by Raw LeWM are poorly suited for downstream robot policy learning.