One small cube = one task: pick an object, place it in a receptacle, then press a coloured button. 4 objects × 4 receptacles × 4 buttons = 64 tasks in total.
Pick a training set from the menu above. Blue cubes are tasks it was trained on, orange ones it never saw; darker means a higher success rate. Drag to rotate, scroll to zoom, hover a cube for its numbers.
A rollout is one attempt at a task in the simulator, scored success or failure. Each percentage pools every rollout across all tasks in that group and all random seeds, so the counts below say how much evidence is behind it.
A held-out task still shares pairs of ingredients with training — object+receptacle, object+button, receptacle+button. Grouped by how many of those three pairs appeared in some training task.
Success rate over all tasks using each object, receptacle, or button — shows whether one ingredient is harder than the rest.