Motion latency remains a challenge for industrial exoskeleton robots. At least 300 ms of feedforward control is required to ensure human–robot synchronization. To secure the feedforward control, we utilize EMG sensor data, which is known to detect movement intention faster than actual motion. We collected data on 10 subjects using seven EMG sensors attached to the lower body. The subjects performed 10 sets of squats, a common lifting exercise, and each set was divided into segments based on the knee angle, with 15 repetitions in each set. We categorized motion intention into four categories labelled Set, Descent, Pause, and Ascent. CNN-LSTM was used to train a model that predicts the subject’s intended action 300 ms in advance. The model showed 80.44% accuracy in predicting motion intentions. The results of this study are expected to address the motion latency issue of industrial exoskeleton robots and enhance the acceptance of using exoskeletons in actual industrial environments.