[Tiny Paper] Toward Pixel-Grounded World Models for Powered Descent: A Rocket Landing Benchmark and Expert Baseline

World models aim to learn predictive latent dynamics for planning from high-dimensional observations, yet most closed-loop control demonstrations remain in simplified domains that do not stress realistic actuation limits or touchdown contact dynamics. We introduce RocketLanding, a physics-based powered-descent-and-landing benchmark built on PyFlyt/PyBullet, designed to stress long-horizon compounding error under constraints that preclude hovering and require precisely timed braking. The task supports pixel observations and goal images for vision-based objectives, while providing a standardized evaluation protocol based on randomized initial-condition perturbations. We contribute a strong, interpretable state-based analytical suicide-burn expert with PD attitude control and gain-scheduled lateral guidance (90\% success on medium perturbations and 20\% on hard), and outline a JEPA-style pipeline (frozen visual encoder, action-conditioned latent dynamics, goal-conditioned MPC) intended to train on offline trajectories generated by this benchmark.