6.2. Numerical Inverse Kinematics (Part 2 of 2)

This video applies the Newton-Raphson root-finding method for numerical inverse kinematics when the end-effector configuration is represented as a transformation matrix.

In the previous video, we derived the Newton-Raphson numerical algorithm for inverse kinematics when the end-effector configuration is represented by a minimum set of coordinates x_d. In this final video of Chapter 6, we modify the algorithm so that the desired end-effector configuration is described by the transformation matrix T_sd. We also need to replace the error vector e by something else. The proper way to interpret e is as a velocity which, if the end-effector followed it for unit time, it would move from the current configuration f-of-theta_i to the desired configuration x_d. For our modified algorithm, we need to find the twist that would take the end-effector {b} frame to the desired frame in unit time.

Let's illustrate using these three frames. The {s}-frame is fixed in space. The {d} frame is the desired configuration of the end-effector, and it is represented in the {s}-frame as T_sd. The {b}-frame is the configuration of the end-effector if the joint vector is theta_i, given by the forward kinematics T_sb-of-theta_i. We can calculate the configuration of the frame {d} relative to the frame {b} as T_bd equals T_sb-inverse times T_sd. We are looking for the twist that moves the {b} frame to the {d} frame in unit time, and the matrix representation of this twist is just bracket-V_b equals log of T_bd. If the {b} frame follows the body twist V_b for one unit of time, it will end up at the desired configuration {d}. Thus V_b serves a similar role as the error vector e in the coordinate version of the algorithm.

Now we can write the modified algorithm as shown here. We begin with an initial guess theta_zero, then we calculate the matrix representation of the body twist V_b that moves the {b} frame to the {d} frame. If the angular component omega_b and the linear component v_b of the body twist are both small, then theta_zero is a good solution to the inverse kinematics problem. Otherwise, we update our joint vector guess by adding the pseudoinverse of the body Jacobian times the body twist V_b and repeat.

We apply this algorithm to the inverse kinematics of a 4-joint RRRP arm. The desired end-effector configuration is illustrated by the frame {d}, but our initial joint vector guess theta-zero would put the end-effector frame {b} at the configuration shown. Let's animate the robot to move to the joint angles found at each of 4 iterations of the Newton-Raphson algorithm. This is iteration 1 ... then 2 ... then 3 ... then 4. After 4 iterations the remaining configuration error is imperceptible, and the numerical inverse kinematics algorithm has converged to a good solution.

Of course in practice the robot would not actually move until it has calculated a solution; the animation here just visually demonstrates the iterations. Let's watch the animation one more time.

We have been focusing on finding the joint positions that achieve a desired end-effector configuration, but in some cases we only need the joint velocities that achieve a desired end-effector twist. We call this the inverse velocity kinematics, where the desired twist V_d and the Jacobian J are expressed in the same frame, either the space frame {s} or the end-effector frame {b}. If the robot has more than 6 joints, the use of the pseudoinverse ensures that the sum of the squares of the elements of theta-dot is the smallest among all joint velocity vectors that achieve the desired twist. If the robot has less than 6 joints, meaning it cannot exactly achieve an arbitrary desired twist, the use of the pseudoinverse ensures that the error of the actual twist from the desired twist minimizes the sum of squares of elements of the error. The book describes other types of inverses that yield solutions minimizing other quantities.

This concludes Chapter 6. With the forward kinematics of Chapter 4, the velocity kinematics of Chapter 5, and the inverse kinematics of Chapter 6, you are now prepared to design kinematic controllers for open-chain robots, as discussed in Chapter 11. But before we do that, in Chapter 7 we look at the forward kinematics, inverse kinematics, and velocity kinematics of closed-chain robots, which exhibit features not present in open chains.