MetaWorld MT50

DLAM: Distributional latent action model with temporal constraints for robot policy learning

DLAM introduces a distributional latent-action model that represents robot action transitions as diagonal Gaussians, enabling structured extraction of action priors from action-free video data. The approach uses normalized composition and reversal over equal-gap triplets to constrain both mean and variance, addressing error propagation in recursive composition that affects prior deterministic methods. A flow-matching policy jointly generates mean transition sequences and robot actions, and the method outperforms latent-action baselines on MetaWorld MT50, LIBERO, and real-world manipulation tasks under a controlled π₀ transfer protocol.