Simulation has played an important role in advancing robot learning [1, 2, 3, 4] by providing a controlled yet versatile environment for developing and testing algorithms. Data-driven approaches, in particular, typically deploy robots into simulations to undergo extensive training across a variety of diverse and randomized settings to enable generalizable and adaptable behaviors. Significant advancements in robot learning have been achieved by randomizing object shapes [4, 5], textures [6, 7, 8, 9], and dynamics [10]. However, the layout of objects, despite being another critical factor, remains challenging to reach fully open-ended randomization. Unlike object properties, which can be easily specified within a range without interfering with other objects, object layout must consider the presence of other objects and physical feasibility. For instance, arranging objects in a scene requires ensuring that they do not overlap and are placed in stable positions instead of falling down from the air. Existing efforts often prevent this issue by fixing the object bases [11, 4, 12, 13], but this strategy is not suitable for many objects like bottles or cups. As the number of objects increases within a limited space, generating a randomized yet stable object layout becomes exponentially difficult.