Consider a system of autonomous mobile robots initially randomly deployed on the nodes of an anonymous finite grid. A gathering algorithm is a sequence of moves to be executed independently by each robot so that all robots meet at a single node after finite time. The robots operate in Look-Compute-Move cycles. In each cycle, a robot takes a snapshot of the current configuration of the grid in terms of occupied nodes (\emph{Look}), then based on the perceived configuration, decides whether to stay put or to move to an adjacent node (\emph{Compute}), and in the later case makes an instantaneous move accordingly (\emph{Move}). The robots have \emph{weak multiplicity detection} capability, which enables them to detect if a node is empty or occupied by a single robot or by multiple robots. The robots are \emph{asynchronous}, \emph{oblivious}, \emph{anonymous}, can not communicate with each other and execute the same distributed algorithm. In a faulty system, however, any robot can \emph{crash}, which means that it becomes completely inactive and does not take part in the process any further. In that case a fault-tolerant gathering algorithm is an algorithm that gathers all the non-faulty robots at a single node. This paper considers a faulty system that can have at most one crash fault. With these assumptions deterministic fault-tolerant gathering algorithms are presented that gather all initial configurations that are gatherable in a non-faulty system, except for one specific configuration called the \emph{2S2 configuration}.