A vast number of applications for legged robots involve the performance of tasks in complex, dynamic environments. One consequence of this expectation is that these environments put legged robots at high risk for limb damage. The ideal response to limb damage is for the robot to adapt its locomotion to avoid further damage and move to a secure area for retrieval. A substantial prior robotics literature addresses the problem of adaptation to limb damage, with a preponderance of attention to locked-joint failures recovered through quasi-static gaits explored mainly through simulation studies. In contrast, this dissertation focuses on the development of highly energetic recovery gaits for a physical robot with a lost (non-load-bearing) limb whose diagnosis is to be automated by a failure identification algorithm. This thesis has two main contributions. The first contribution is the development of a tripedal dynamic gait for a power autonomous unconstrained direct drive quadrupedal robot with no external sensing that allows it to execute fore-aft locomotion regardless of the location of the "missing" limb. This controller takes inspiration from the comprehensive, longstanding literature establishing that running animals {anchor} the Spring Loaded Inverted Pendulum (SLIP) {template}. Empirical results confirm the resulting dynamic gait is more energetically costly than its quadrupedal sibling but less costly than its quasi-static counterpart. The second contribution is the implementation of two deep learning techniques to accurately identify and locate a single limb failure. With these learning algorithms, it is possible to input either a labeled or unlabeled dataset of both non-damaged and damaged gait data and predict the faulty limb. Empirical data establishing the stability of fore-aft tripedal dynamic gait are used to train deep learning models and test their validity.