CausalWAM
A causal world action model for robot learning. By adding sparse routing to the action head, it separates the latent variables on which action generation depends into a causally relevant subset, and provides theoretical identifiability guarantees under sparsity constraints.
Timeline (last 90 days)
CausalWAM separates action-relevant latent variables from scene-prediction variables from a causal perspective, so that action generation depends only on action-relevant variables.
The paper proves that under sparse constraints, the latent variables used to generate actions are an invertible function of the action-relevant latent variables, and the true structure can be uniquely identified up to a permutation.
CausalWAM's identifiability results do not require sufficiently strong temporal dependence, make no assumptions about the evolution of unobserved regimes, and even hold when action-relevant variables are statistically correlated with the remaining variables.
The paper validates on synthetic data that action-relevant latent variables can be correctly identified under sparse constraints.
The paper demonstrates improvements in success rate or efficiency in closed-loop tasks on real robots.
CausalWAM applies the identifiability theory of causal representation learning to a concrete architecture for robot policy learning, rather than remaining at the purely theoretical level.
All Facts (6)
CausalWAM separates action-relevant latent variables from scene-prediction variables from a causal perspective, so that action generation depends only on action-relevant variables.
50%UnverifiedThe paper proves that under sparse constraints, the latent variables used to generate actions are an invertible function of the action-relevant latent variables, and the true structure can be uniquely identified up to a permutation.
50%UnverifiedCausalWAM's identifiability results do not require sufficiently strong temporal dependence, make no assumptions about the evolution of unobserved regimes, and even hold when action-relevant variables are statistically correlated with the remaining variables.
50%UnverifiedThe paper validates on synthetic data that action-relevant latent variables can be correctly identified under sparse constraints.
50%UnverifiedThe paper demonstrates improvements in success rate or efficiency in closed-loop tasks on real robots.
50%UnverifiedCausalWAM applies the identifiability theory of causal representation learning to a concrete architecture for robot policy learning, rather than remaining at the purely theoretical level.
50%