DLAM: Distributional Latent Actions with Temporal Constraints
DLAM turns action-free robot videos into probabilistic latent transitions that compose more cleanly over time.
The paper represents each visual transition as a diagonal Gaussian, then constrains those transitions through composition and reversal across equal-gap frame triplets. Its policy setup freezes the encoder and trains a flow-matching model to generate both latent transition means and robot actions. The authors report stronger held-out video reconstruction than latent-action baselines, plus better transfer results on MetaWorld MT50, LIBERO, and real-world manipulation tasks. ArXiv · AI/CL/LG's note
The paper represents each visual transition as a diagonal Gaussian, then constrains those transitions through composition and reversal across equal-gap frame triplets. Its policy setup freezes the encoder and trains a flow-matching model to generate both latent transition means and robot actions. The authors report stronger held-out video reconstruction than latent-action baselines, plus better transfer results on MetaWorld MT50, LIBERO, and real-world manipulation tasks. ArXiv · AI/CL/LG's note
score 4