A collision-avoidance system for use with an autonomous-capable vehicle can continuously receive image frames captured of the roadway to determine drivable space in a forward direction of the vehicle. The system can determine, for each image frame, whether individual regions of the image frame depict drivable space. The system can do so using machine-learned image recognition algorithms such as convolutional neural networks generated using extensive training data. Using such techniques, the system can label regions of the image frames as corresponding to drivable space or non-drivable space. By analyzing the labeled image frames, the system can determine whether the vehicle is likely to impact a region of non-drivable space. And, in response to such a determination, the system can generate control signals that override other control systems or human operator input to control the brakes, the steering, or other sub-systems of the vehicle to avoid the collision.