Lin, Zhang, and Zhao developed a new deep learning framework called the Parallel Efficient Transformer (PET) for accurately estimating continuous hand movements from surface electromyography (sEMG) signals. sEMG captures the electrical activity of muscles through sensors on the skin, offering a noninvasive way to predict motion intentions, critical for prosthetic control, rehabilitation robotics, and human-machine interfaces. Current sEMG-based deep learning methods, especially those using Transformers, are often too computationally heavy, with high latency and energy use that make them difficult to deploy on wearable or edge devices. To address these issues, the researchers designed PET, which integrates a lightweight external attention mechanism and a parallel transformer structure to reduce computational complexity from quadratic (O(n²)) to linear (O(n)) while keeping prediction accuracy high. PET was tested on four public datasets, Ninapro DB2, Ninapro DB7, FMHD, and SEEDS, involving over 60 subjects. Across these benchmarks, PET achieved superior performance compared to state-of-the-art models like TCN, LSTM, GRU, BERT, and Conformer. For example, in the Ninapro dataset, PET reached a correlation coefficient (CC) of 0.85 ± 0.01, RMSE of 7.26 ± 0.32, and NRMSE of 0.11 ± 0.01, outperforming other methods while maintaining ultra low latency and power usage. Deployed on a Raspberry Pi, PET’s inference time was just 13.65 ms, and its power efficiency (AME) was the best among all compared models. In online experiments with high density EMG sensors and a data glove, PET predicted hand joint angles in real time with a mean CC of 0.86, showing smooth and responsive movement estimation.