To ensure an autonomous truck can operate safely in a dump area, it is crucial to detect a berm accurately in advance. However, there are two challenges. First, the berm is not a static terrain but a movable one because of soil dumping. Second, berms are often irregular in shape-they are neither straight lines nor smooth curves. We considered two types of possible existing methods, but only to find they are not accurate and can’t provide height information. Therefore, this paper proposes a berm detection algorithm, which includes three steps. First, extract berm candidate 3D LiDAR points based on a 2D height difference grid map. Second, use a binary Bayes filter to build and update 3D dynamic probability grid maps. Last, use a fitting rectangle technique to recognize the berm. We call this algorithm a Probability Grid Berm Detection (PGBD) algorithm. Off-line experimental evaluations on PGBD carried on datasets show good performance, compared with two curb detection algorithms, which are Hough Transformation and Haar Wavelet Transformation. And the good performance of the PGBD algorithm is further verified in the real-time experiment.