Executable interfaces are becoming the common lever for robot reliability

Three systems reduce ambiguity before action generation. ReferTrack makes a vision-language-action (VLA) policy select an indexed person detection before predicting waypoints; it reaches 74.1% success on ambiguous tracking, 22.9 points above its cited single-view baseline. LENS removes or merges task-irrelevant objects before existing planners and controllers run, cutting heavy-clutter control times from roughly 1,000–4,000 seconds to about 40–135 seconds in reported settings. DEED applies the same systems logic to humanoid restocking through synchronized controls, curated demonstrations, recovery data, and latent distribution monitoring. Its hardware study is concrete, but it does not report component-level gains or comparative success rates.

World-model research is increasingly evaluated by whether predictions support action, not only whether generated observations look plausible. KineBench converts generated videos into six-degree-of-freedom end-effector trajectories and executes them in simulation; its explicit pipeline reports about 1.5–3 cm translation error on unseen trajectories, versus errors near 10 cm for an inverse-dynamics baseline. Koopman Dreamer constrains latent dynamics to control long-rollout error and raises simulated UAV navigation success from 53.8% to 73.8%. Dream rehearsal then isolates a different failure: replay preserved measurable world-model knowledge, yet the actor forgot. Cloning high-scoring imagined trajectories recovered the skill in all three seeds, while reinforcement learning over the same imagined data recovered it in none. Together, these studies make executability, stability, and component-level diagnosis central tests of model usefulness.