The measurement of biological signals for assistive rehabilitation devices such as wearable exoskeletons, particularly angle measurement, has garnered increasing research interest. Since surface electromyography (sEMG) precedes movement onset, it enables early angle prediction with lower latency than direct measurement. However, existing angle prediction models are often computationally intensive and inadequately tailored to the unique characteristics of sEMG signals. To address this, we propose a novel Fourier basis mapping and adaptive normalization (FBMAN) framework for lumbar sEMG-based angle prediction. Specifically, sEMG and motion signals were acquired using Delsys equipment to establish a dataset, followed by signal processing via dual-tree complex wavelet transform (DTCWT) combined with the Nelder-Mead algorithm. An entropy-weighted evaluation scheme was developed to construct a comprehensive denoising assessment system. Within the prediction module, slice-level adaptive normalization (SAN) and Fourier basis mapping (FBM) were innovatively introduced to accommodate sEMG’s nonstationarity and time–frequency properties. Remarkably, when integrated with lightweight architectures such as multilayer perceptrons (MLPs), the FBMAN-MLP framework demonstrated superior performance in both cross-subject and cross-day experiments. Experimental analysis of sliding window sizes revealed that a 200 ms window optimally balanced real-time capability and predictive accuracy, achieving a root mean square error (RMSE) of 0.054° and a prediction latency of $14.078{\,}\pm {\,}0.421$ ms. By leveraging sEMG’s time–frequency and nonstationary features without additional sensors, our method outperforms existing approaches, validating FBMAN’s potential for assistive rehabilitation applications.

Off-policy reinforcement learning is usually used to train the grasping task model of the manipulator. However, in the training process, it is difficult to collect enough successful experience data and rewards for learning and training; that is, there is a problem of sparse rewards. Hindsight experience replay (HER) allows the agent to relabel the completed states. However, not all failed experiences have the same effect on learning and training. Facing the many transitions generated by the environment during operation, adopting a random uniform sampling method from the experience replay buffer will result in low data utilization and slow convergence. This paper proposes using a prioritized sampling method to sample the relabelled transitions, and then combines various off-policy reinforcement learning algorithms with it for training in simulated environments. This paper uses the prioritized sampling method, which allows the agent to access more important transitions earlier and accelerate the convergence of training. The results demonstrate that hindsight experience replay with prioritization (PHER) exhibits significantly faster convergence compared to other methods.