10.7.1. Iterative Methods

1. Newton-Raphson Method:
   • Linearizes the non-linear kinematic equations.
   • Update rule: $q_{i+1} = q_{i} + J^{-1}(q_{i})(X_{d} - f(q_{i}))$
   • Converges quickly near the solution but requires good initial guess.

The Newton-Raphson method is an iterative technique used to approximate solutions to equations that are too complex to solve directly. In the context of kinematics, it helps find joint parameters (the values that define the robot's joint angles) needed to achieve a specific position and orientation (the desired end-effector pose). The process starts with an initial guess of the joint parameters. The formula provided, which involves the Jacobian matrix, helps adjust this guess. It linearizes the problem around the current guess, making it easier to find the next guess. The method is known for its speed and efficiency when the solution is reasonably close to the initial guess; however, if the guess is far off, it can diverge and fail to provide a solution.

Imagine you are trying to find a path through a dense forest using a map. You start at a point where you think you are. If you make a small adjustment moving north, and it feels right, you keep going that way until you find the correct path. If you had started too far south, you might end up going deeper into the forest, losing your way instead of finding the path. The Newton-Raphson method works similarly by refining guesses until it finds the optimal path (solution).

2. Gradient Descent Method:
   • Minimizes the cost function:
   $$$$

orm{f(q) - X_d}^2$$

• Slower than Newton-Raphson but more stable in some cases.

The Gradient Descent method is another iterative method used to find optimal solutions, particularly when the system is complex or when the Jacobian is difficult to calculate. This method focuses on minimizing a 'cost function', which measures how far off the current joint configuration is from the desired configuration. The function evaluates the error between the current and desired states of the end-effector. The approach involves taking small steps in the direction that most reduces the error, gradually approaching the optimal configuration. While it may take longer to converge to a solution compared to Newton-Raphson, it's more stable especially when the landscape of the cost function is complex.

Consider a person trying to find the lowest point in a hilly landscape blindfolded. They can feel the slope of the ground, which helps them determine which direction to step towards to descend. Initially, they might only be taking small steps in the direction that feels downhill. Though they may take a longer time to find the lowest point compared to someone with a map (like in Newton-Raphson), they’ll often find their way effectively and avoid getting stuck or lost.

3. Damped Least Squares (Levenberg–Marquardt Algorithm):
   • Handles singularities by adding damping:
   $$\Delta q = (J^T J + \lambda I)^{-1} J^T (X_d - f(q))$$
   • Offers a balance between speed and stability.

The Damped Least Squares method, also known as the Levenberg-Marquardt algorithm, is a powerful technique that combines features from both the Newton-Raphson method and the Gradient Descent method. Its primary benefit is in managing situations where the Jacobian becomes difficult to invert or when the solution space has singularities—points where the system cannot operate effectively. By introducing a damping factor, the algorithm adjusts the steps it takes, ensuring that updates are stable and converge smoothly, striking a balance between fast convergence and stability. This makes it particularly useful for manipulator control in complex configurations where harsh changes could lead to system failure.

Imagine a tightrope walker who must traverse a very narrow beam. If they move too quickly or make abrupt movements, they risk falling. However, by taking careful, calculated steps and adjusting their movements based on how stable they feel, they maintain balance. The Damped Least Squares method functions similarly; it ensures the robot’s movements remain stable (like the tightrope walker) while still progressing toward the desired end-effector position.

--