"Tabular experimental registry with synthetic illustrative images"

"This dataset provides a preregistered experimental design for evaluating causal depth gating in visual reinforcement learning for robotic manipulation. It defines 7,500 unique method\u2013task\u2013seed\u2013condition runs, corresponding to 750,000 planned evaluation episodes across three RL-ViGen Robosuite tasks, ten learning methods, ten confirmatory seeds, and twenty-five evaluation conditions. The condition matrix separates in-distribution observations from appearance, illumination, camera-view, depth-sensor, cross-embodiment, and combined shifts. Depth-specific extensions cover heteroscedastic noise, structured missingness, systematic bias, and RGB\u2013depth misregistration. The dataset also specifies eleven ablations, outcome definitions, explanation-fidelity measures, exclusion rules, and a robust statistical analysis plan. All empirical outcome fields are intentionally empty because experiments have not yet been executed. Six PNG files provide clearly labeled synthetic illustrations of representative RGB, depth, corruption, viewpoint, and causal-gate conditions; they are not measurements from a trained policy, physical robot, RL-ViGen, or Robosuite execution. The resource is intended to support preregistration, reproducible experiment orchestration, audit-ready result logging, and methodological comparison without fabricating performance evidence.\r\n "