Structured action interfaces anchor embodied world models

RoboInter1.5 and Masked Visual Actions both place spatially explicit signals between intent and predicted outcomes. RoboInter1.5 supplies object grounding, affordances, contact points, and motion traces across more than 230,000 episodes. Masked Visual Actions instead exposes pixel-space entity trajectories to a pretrained video model; one checkpoint can predict scene responses or infer robot motion from desired object movement. On DROID, it reports LPIPS of 0.0945 versus 0.362 for Ctrl-World. Together, the papers support visual structure as an embodiment-flexible control interface, although RoboInter1.5’s inspected excerpt does not provide downstream comparison metrics.

Two papers make physical attribution explicit rather than asking one latent transition to absorb every change. DWM separates action-driven effects from persistent environmental effects such as gravity and drift; it improves planning success by an average of 13.1 percentage points across three modified simulated benchmarks. Agentic Real2Sim reconstructs recorded interactions as physics-based episode twins with geometry, object state, alignment, and replay metrics. Its best tested backend replayed 48 of 100 DROID episodes successfully, showing both the utility and current brittleness of automated conversion.

The corpus also treats latency, hardware cost, and downstream use as part of world-model quality. ABot-World-0 reports persistent 720P generation at up to 16 FPS on one RTX 5090, with 1.2 seconds to the first action-conditioned frame and about 19 GiB peak memory. Masked Visual Actions uses imagined rollouts for policy evaluation and candidate ranking, while Agentic Real2Sim compares model cost alongside replay success. These results broaden evaluation beyond visual fidelity, but benchmark coverage remains uneven and ABot-World-0’s inspected text lacks numerical baseline comparisons.