Robot world models are moving into the action loop, with stronger results than structure

DIAL treats future visual features as the control interface. The vision-language model (VLM) predicts a latent future at horizon 16, and a separate policy turns that forecast plus the current observation into a 16-step action chunk. The concrete payoff is data efficiency: the paper reports state-of-the-art results on RoboCasa GR1 Tabletop with 2,400 trajectories where prior full-data runs used 24,000. It also extends the recipe beyond one robot setup, using 27,419 EgoDex human trajectories for zero-shot generalization tests and reporting real-world transfer on the IRON-R01-1.11 humanoid.

HCLSM pushes world modeling toward object and event structure, but the evidence is mixed. The model splits scenes into slots, runs separate dynamics for continuous motion, event frames, and higher-level goals, and adds learned interaction edges between objects. Its two-stage training matters: first enforce slot specialization, then turn on future prediction. On PushT, the reported next-state prediction error is 0.008 with the staged method, and throughput reaches 2.9 steps per second. The paper also shows the current limits. Slots stay diffuse, the learned causal graph does not become useful, and only 2 of 4 runs finish because of bf16 NaNs.