Industrial robotic manipulators increasingly rely on software-defined, network-connected motion pipelines, expanding the attack surface into motion-command semantics. False data injection attacks (FDIAs) can alter joint-command trajectories while preserving syntactic validity, potentially evading network-only IDS. We propose and experimentally evaluate a semantic-aware pre-execution validation gate as Stage-1 defense. This layer combines unsupervised anomaly detection using an Isolation Forest trained on benign joint-command dynamics and lightweight semantic constraints bounding stepwise displacement. In simulation on a 5-DOF robotic arm, we consider two attack models: an abrupt single-step injection and a low-rate multi-step drift. The step-wise semantic check reliably blocks abrupt manipulations, but slow drift can evade it. To address this, we introduce a trajectory-consistency rule that compares each command against a trusted reference trajectory within tolerance. Under the chosen thresholds, the extended gate achieves a false-positive rate of 0.002 and an F1-score of 0.941 on the drift attack, meeting the 50 ms control-cycle deadline.