Many manufacturing processes require demanding cycle times that subject industrial robots to sudden motions. For instance, in acoustic perforation of airplane engine housings, thousands of holes are drilled to minimize acoustic emissions. Due to the large number of holes required, drilling is performed at over two holes per second, which requires rapid repositioning of the tool center point. However, due to the structural flexibilities present in industrial robots, rapid movements induce residual vibrations. To avoid violating hole tolerances (<0.2mm), industrial robots must sacrifice cycle time and wait for the vibrations to settle before proceeding to the drilling motion. This thesis presents a novel learning-based framework combined with input shaping techniques to avoid vibrations in sudden motions of an industrial robot. Input shaping is an efficient method to filter harmful frequencies and avoid exciting structural modes in industrial robots. The performance of input shaping relies heavily on the accuracy of the predicted natural frequency. In industrial robots, the dominant natural frequency varies significantly with pose and payload – making prediction within the workspace a challenge. In this thesis, a theoretical flexible-joint dynamic model is developed for the first three joints of an industrial robot to predict the dominant natural frequency for any combination of pose and payload. Model parameters, such as joint stiffness, are experimentally identified to improve the accuracy of the theoretical model. The flexible-joint dynamic model is paired with an artificial neural network for efficient real-time prediction of natural frequency for any pose. Collection and labelling of training data is autonomously performed by simulating the flexible-joint dynamic model. Model adaptation, or transfer learning, is used to quickly map the artificial neural network trained for no payload to a new payload with a 90% reduction in required training data. Finally, the learning-based prediction framework is used to automatically design time optimal yet robust zero-vibration-derivative input shapers for vibration avoidance. The developed methodologies are experimentally validated on a Staubli RX90CR industrial robot paired with an open-architecture controller developed fully in-house. Experimental results show an 85% reduction in residual vibration during aggressive motions.