Abstract
Abstract
Existing intelligent robots in vocational training lack the ability to perceive and adaptively adjust the trainees’ states in real time. Furthermore, existing lightweight models struggle to balance efficiency and accuracy, limiting their application in real-time interactive scenarios. Therefore, this study proposes an adaptive vocational training robot system based on a lightweight model and emotion perception. The system first constructs a classroom emotion dataset, then uses a residual network and a 16-layer visual geometry network to build a multimodal perception model. Feature fusion is enhanced through attention mechanisms and a feature pyramid network. Subsequently, depthwise separable convolution and pruning are used to achieve model lightweighting, ultimately integrating these elements to construct the adaptive vocational training robot system. Experimental findings demonstrate that when the data batch size increases to 32, the number of floating-point operations reaches 82.1 M. With a pruning rate of 70 %, the compression rate increases to 73.4 %. With a data volume of 8 k, the end-to-end latency only increases to 82 ms. At an illumination intensity of 500 lux, the highest attention assessment accuracy is 89.7 %. With 30 participants, the highest performance improvement rate reaches 34.8 %. The results show that the system has excellent response speed, robustness and teaching effectiveness, providing a feasible technical path and practical reference for adaptive technology in the field of vocational education.