This report presents the work carried out in the project AI for reconnaissance sensors. The project has conducted research in the area of automatic target recognition, positioning and sensor planning. The aim of this research has been to investigate different AI algorithms that will eventually facilitate the work of the sensor operator and thereby increase the efficiency and quality of reconnaissance by helping operators to have a better situational awareness, higher system confidence, reduce risk of human error and lower mental workload. Within target recognition, the work has focused on the development of methods for generating synthetic training data as modern target recognition methods need vast amounts of training data. To achieve high target recognition performance, a very large set of representative images is generally required, showing objects in different scales (distances) and views, in different terrain types and weather. An interesting option is to use simulated images as these can be generated quickly, although it has been a challenge to get them faithful enough for training target recognition algorithms. Therefore, research has been conducted to make simulated images more realistic with domain translation algorithms. A distance measure between image distributions has been used called the Fréchet distance, and by using it we have seen that domain translation improves simulated data so that it is more like real images. Within automatic positioning, the work has focused on studying how high positioning accuracy is possible if LiDAR and geographical information have been used instead of satellite navigation that can be jammed. Three different methods have been studied for this, which we have called "positioning from individual trees", "positioning against terrain" and "positioning against tree density". Results show that the first two have the best positioning, typically below 1 m and thus on par with the current GNSS. The two methods require robust matching of individual landmarks (such as a specific tree), while the last requires only known areas (such as location of a grove of forest). For the last, we have achieved a positioning of the same order of magnitude is achieved as for older GNSS receivers i.e. 6-20 m. Within sensor planning of sensor systems, work has focused on studying how the modern machine learning method Reinforcement Learning can be used. Several experiments have been carried out using this method and these show that in several cases machine learning can be used for planning of sensor systems. In one of the experiments, a reconnaissance agent has been given the task, based on visual sensor images, of aiming the sensor towards targets in the terrain to enable e.g. classification and identification of detected targets. With the correct parameter adjustment, the reconnaissance agent was often able to find, zoom and steer onto targets down to the size of about 200 pixels. The experiment shows that the agent can support a sensor operator to find targets more quickly while giving the operator better access to classification for further threat assessment. The method has the potential to in the future be applied to more complex scenarios with partially hidden targets and sensor platforms in motion.