This chapter explores the estimation of lower limb joint angles from muscle signals in able-bodied adults across various locomotion modes. The study utilizes an existing open-access dataset, containing motion data of lower limb joints and surface EMG signals from 11 right lower limb muscles of 22 able-bodied adults across different terrains and locomotion conditions. The chapter is organized around two primary objectives. The first objective is to analyze motion and EMG signals by extracting data from the dataset, selecting data from 10 subjects engaged in treadmill walking at 1.3 m/s speed, and preprocessing it for consistency and compatibility for subsequent analysis and intelligent model training. The second objective involves using the preprocessed data to train backpropagation neural network models with three training optimizers, Levenberg-Marquardt (LM), Bayesian Regularization (BR) and Scaled Conjugate Gradient (SCG) to estimate the desired joint angles during treadmill walking. The performance of the trained models is then assessed by comparing their predictions with actual joint angle measurements from the dataset, using evaluation metrics such as mean absolute error or root mean square error. The performance of the trained neural network was evaluated using one test subject for all three methods. The LM method achieved the best performance with an MSE of 83.2474 at epoch 103 and R > 0.91. The BR method had an MSE of 140.3847 at epoch 124 with R > 0.84. SCG method showed the best performance, with an MSE of 171.0699 at epoch 1000 and R > 0.81. This analysis interprets the results to provide insights into the relationship between muscle signals and joint angles during treadmill walking, and to apply these findings to the dynamics and control of lower-limb rehabilitation robots.