Cyber-physical systems (CPS) integrate computation with physical processes, driving innovations in fields like agriculture, aviation, healthcare, and transportation. However, this integration introduces new risks and vulnerabilities that traditional cybersecurity measures cannot address. To protect CPS, researchers have employed machine learning (ML) and deep learning (DL) techniques to model system behaviors for attack detection and recovery. Existing studies often rely on simulated data and focus on a single DL technique, lacking comprehensive comparisons. This study addresses these gaps by using real data from a robotic vehicle testbed to compare various ML/DL techniques based on prediction accuracy and generalization. Our experiments reveal that simpler models, particularly Small Dense and Small LSTM networks, exhibit superior generalization and adaptability, maintaining low errors across diverse conditions. Dense networks demonstrate the best overall performance, with the Small Dense model achieving the lowest errors in position and linear velocity predictions. LSTM models also show robust performance, balancing accuracy across all metrics. In contrast, GRU and RNN models exhibit mixed results, and TCNs consistently underperform, indicating challenges in adapting to new environments. These findings underscore the importance of model selection and complexity management for accurate and adaptable predictions.