Neural Network-Based Forward Dynamics and Model Predictive Control for a n-Link Robotic Arm

This project involves the implementation of advanced control systems for a 3-link robotic arm, combining neural network techniques with Model Predictive Control (MPC). The project is divided into several key phases: learning forward dynamics using neural networks, developing an MPC controller, and evaluating MPC controller using the learned dynamics model.

Phase 1: Neural Network for Forward Dynamics Learning
- Generated a dataset capturing the dynamics of a reference arm under various torque conditions. The data was used to train a neural network model to predict the future state of the arm based on applied torques.
- Developed and trained a neural network model to learn the forward dynamics of a 3-link robotic arm. The model was trained using the collected dataset to accurately predict arm movements and behaviors.
- Tested the neural network's ability to predict the arm's behavior, ensuring it closely mirrored the reference arm's movements.

Phase 2: Implementation of Model Predictive Control (MPC)
- Developed the MPC class to compute optimal actions based on the current state and target goal. Implemented adaptive step sizes and dynamic gain adjustments to enhance control accuracy.
- The MPC controller was designed to minimize the distance and velocity errors to the target position, ensuring precise movement control of the robotic arm.

Phase 3: Evaluation with Learned Dynamics Model
- Trained a neural network to predict the next state of the robotic arm given the current state and actions, using the previously collected data.
- Implemented a class to use the trained model for state predictions. The MPC controller was tested with the learned dynamics model to evaluate performance.
- Evaluated the system's performance based on criteria such as distance to the goal and end-effector velocity, ensuring the robustness and accuracy of the control mechanism.