In shape stacking the agent must stack shapes to make as high of a tower as possible. A variety of polygons are available for the agent to place.

Below are the types of polygons available for the agent to place. The size of each shape is modified to make 810 shapes in total.

The purpose of the shape observations is to convey the properties of a particular shape. The agent must then infer how the shape should be used from these action observations. To consider 3D geometry, we simply view the shape from different angles and distances. This gives a complete view of the shape. While in the experiments there are 1,024 such images per shape, here we display 6 sample observations for 3 shapes.

We also demonstrate different strategies for placing the shapes. Our primary results are reported for when the agent selects the "x" position of where to place the shape on the horizontal axis. Here we show choosing "(x,y)" where to place on the horizontal and depth axis, and not choosing where to place at all.