Block Disentanglement in CRL: Bridging Identifiability and Visual State Estimation
The paper argues that weaker intervention assumptions can still identify causal variables, but only up to meaningful blocks.
Its CRL framework recovers groups of latent causal variables when full disentanglement is not guaranteed by the available interventions. The authors then apply that block-disentangled objective to visual state estimation for robotic systems, using images and videos without labels. They present the embodied setting as a controlled bridge from identifiability theory to practical label-free state recovery. ArXiv · AI/CL/LG's note
Its CRL framework recovers groups of latent causal variables when full disentanglement is not guaranteed by the available interventions. The authors then apply that block-disentangled objective to visual state estimation for robotic systems, using images and videos without labels. They present the embodied setting as a controlled bridge from identifiability theory to practical label-free state recovery. ArXiv · AI/CL/LG's note
score 5