Mobile robots are more and more efficient but limited by the problem of variation of motion properties, as is the case for applications in natural environments. Sharp transitions can be estimated reactively, but are difficult to predict, and lack of anticipation can lead to inappropriate or even hazardous behaviors. This project aims to overcome this problem by proposing adaptive mechanisms for robotic behavior by anticipating these variations from scene perception. The project proposes to develop machine learning approaches to predict and map the interaction conditions. It will also develop stable supervision processes to select and modify on-line several control modes. Tested on realistic scenarios using the robotic platforms available from the project members, such developments will strengthen the autonomy of robots to offer efficient and safe solutions to societal issues, particularly for agriculture.