Intentional or unintentional contacts are bound to occur increasingly more often due to the deployment of autonomous systems in human environments. In this paper, we devise methods to computationally predict imminent collisions between objects, robots and people, and use an upper-body humanoid robot to block them if they are likely to happen. We employ statistical methods for effective collision prediction followed by sensor-based trajectory generation and real-time control to attempt to stop the likely collisions using the most favorable part of the blocking robot. We thoroughly investigate collisions in various types of experimental setups involving objects, robots, and people. Overall, the main contribution of this paper is to devise sensor-based prediction, trajectory generation and control processes for highly articulated robots to prevent collisions against people, and conduct numerous experiments to validate this approach.