In the first phase, we individually train both 2D and 3D classifiers with the classification loss of each domain. Here, in the second phase, please note that we first train our framework without the 3D-to-2D loss for the first few epochs. Here, weempirically observethat using abigger patch size isineffectiveinterms ofclassification, since having awider relative receptive field is crucial for understanding the scene from the image. On the other hand, when we use asmaller patch size, fine details ofthe image could not be preserved. Also, we filter the points ofthe occluded object which should not exist onthe image.

Our code for the StarCraft II micromanagement tasks (Figure 1 of the main paper) is available at https://github.com/mzho7212/LICA. Apart from focus-firing where LICA agents simultaneously focus on individual enemy units, they also learned to build strategic formations (e.g. Fig. A3) where high health Marines are moved to the front toattract theattacks oftheheuristics-based enemy Marines thatoften prioritize closer targets. Wealsoobserve thatLICAagents learned topullMarauders forwarddespite theirlonger shooting rangecompared to Marines; this is because Marauders are more durable and can afford to divert possible enemy attention andtakemoreincoming damage. Furthermore, we observethat the ally Medivaclearned to switch between multiple ally units and prioritize units with low health.