Our approach accounts for object-object interaction, which is crucial in some scenarios.

Methods assuming sub-goals (corresponding to different objects) are independently reachable, aim to decompose and simplify the RL task by sequentially solving single-object tasks. This approach overlooks potential object-object interaction, as made evident by the failure cases demonstrated below.

Adjacent Goals

Goals are sampled randomly on the table such that all cubes are adjacent.

The single-object approach often pushes cubes away from their goals while manipulating others.

Small Table

Table is substantially smaller.

The single-object approach often pushes cubes off of the table.

Ordered Push

A narrow corridor is set on top of the table such that its width can only fit a single cube, which requires pushing in a specific order.

The single-object approach is not aware of the ordered relationship between sub-goals or the object-object physical interaction, 

resulting in failure to reach either sub-goal.