School Bus 1.0 Garbage Truck 0.99 Punching Bag 1.0 Snowplow 0.92 Motor Scooter 0.99 Parachute 1.0 Bobsled 1.0 Parachute 0.54 Fire Truck 0.99 (a) School Bus 0.98 (b) Fireboat 0.98 (c) Bobsled 0.79 (d) FIGURE 1. A deep learning algorithm prediction (probabilities following the algorithm's label) for typical road vehicle poses in a 3D simulator (a) and for unusual poses (b)-(d) [1]. This previous section focused on the concept of meaningful human control and dynamic targeting, but there are additional lessons to be learned by examining autonomous weapons and static targeting, dis- cussed in the following section. lined that because of serious flaws in computer vision algorithms and high uncertainty, human control is necessary due to the inability of an autonomous system to perform pre- dictably and reliably. However, such operations are also prone to human error due to their dynamic nature. A different picture emerges when looking at missions that prosecute static targets, like the Chinese Embas- sy mission discussed earlier. These missions are typically planned days, Low Uncertainty Engagements Are Very Much in the Realm of Autonomous Weapons Today For dynamic targeting in offensive missions, the previous section out- DECEMBER 2019 ∕ IEEE TECHNOLOGY AND SOCIETY MAGAZINE if not weeks in advance with teams of analysts and weapons special- ists evaluating multiple courses of action. Unlike the pilot in a dynamic targeting scenario, the pilot asking for the target is not the person who picks the target (this is typically a team of people), nor is the person who authorizes the target (which is a senior official who consults with a team of lawyers). The pilot of the plane that bombs a predesignated t arget simply confirms that the 23