In a similar way to bacteria that have to navigate in their environment, soaring birds try to minimize their effort by finding and exploiting ascending currents. However the environment is highly turbulent and unpredictable with thermals constantly forming, disintegrating and transported within minutes. How the birds navigate these environments remains unknown. It is a notoriously difficult/impossible task to assess what cues are used by the soaring birds and what strategies they developed to explore and exploit such turbulent environments. For this investigation, we chose to emulate how an agent could learn to see what strategies would emerge. I then set up an experimental reinforcement learning framework that I implemented in a two-meter wingspan glider. Here, the soaring agent measured its state (height and a set of cues), took actions (change the left/right direction it is heading to) and recorded what resulted from these actions given the previous state. After an initial learning period, the glider chose the actions according to their state that would maximize the gain in altitude. In short, the glider learned to soar through its past experience. The learned strategy was based on accurate estimates of local wind accelerations and roll-wise torque. In the field, the glider could typically gain hundreds of meters in height in a few minutes when facing environments with thermals. Our results not only highlight the vertical wind accelerations and roll-wise torques as effective mechanosensory cues for soaring birds but also provide navigational strategy that is directly applicable to the development of autonomous soaring vehicles to increase their time aloft with minimal energy cost.