The only
real failsafes I have are overcurrent detection on the blade motor (when you hit something bad, current will spike) which triggers immediate shut-off, as well as a big red emergency stop button that will be on top of a pedestal at between knee and waist height.
Otherwise I intend to implement naive obstruction avoidance to start where I detect proximity to an obstruction and simply steer away from it when I'm too close. In the case where the distance to the obstruction decreases faster than warranted by vehicle velocity, I'll simply shut off the blade and stop moving until either the obstruction goes away, or some timeout expires.
Later on I'll add some smarts to this w/ a bit of mapping so that the static object "too close threshold" varies depending on whether or not the object is known in the current map - that way I can implement wall-following for close edging without that behaviour conflicting with the lower-level "don't turn the neighbour's cat into sausage" behaviour.
Edit 1:
My yard is quite private, and lawn mowers (even electric ones) make scary noises, so I don't really expect that much will want to interfere with it while it is running.
Edit 2:
It'll also have a camera and I'll be experimenting with various monocular SLAM and odometry algorithms. If nothing else, I'll use the camera to capture video for remote FPV control and for offline structure-from-motion mapping purposes. My hope is that if I can get decent enough pose estimation (likely from fusion of VSLAM + wheel odometry + AHRS/IMU data + GPS via an EKF), that I can compare the live camera image with one taken from the map at a similar pose to find and avoid dynamic objects in the scene. Of course, to do that well I'll actually need some sort of statistical map of the environment, as a single textured 3D mesh won't convey the expected variance of the scene (things blowing in the wind, etc).