I have recently implemented a model predictive controller (MPC) to calculate the necessary reaction forces for a legged robot. The work presented here is based on "Highly Dynamic Quadruped Locomotion via
Whole-Body Impulse Control and Model Predictive Control" by Kim et al. If you don't know what model predictive control is, I recommend this excellent explanation by Steve Brunton. I also found this guide to model predictive control with CasADI to be extremely helpful. CasADi is an open source nonlinear optimization tool which I'm using for MPC.
However, I'm not entirely sure yet whether this code works the way it's intended--I'm not getting any error messages, but the force values I'm getting seem too low. If you spot anything wrong with my implementation, please let me know.
Here's the code.
Part 1. The Dynamics Formulation
From the paper:
First off, you may be wondering what this notation means.
This refers to a skew-symmetric matrix. The 3x1 vectors r1 and r2 (footstep positions) are converted into a 3x3 matrix in order to perform a cross product operation with the inverse of the global inertia tensor.
A few more notes:
The end result is a very long equation which will be used as a constraint for the QP. The states and controls are defined as symbolics, but other values such as the inertia tensor, rotation matrix, and foot positions are fed in as numerical values from sensor feedback.
Part 2. The Objective Function
The objective function is shown above.
Let's start with this notation. What does it mean? Well, double bars represent a matrix norm. However, the R in subscript clues us into the fact that this is a special case of the matrix norm with an inner product. Q and R are weighing matrices and should be diagonal.
However, I'm not sure what the purpose of the square root would be here. It seems to me that minimizing the square root of f(x) is basically equivalent to minimizing f(x) by itself, but more computationally expensive. In fact, running the code with the square root operation actually causes a "QP is infeasible!" error. As such, I'm leaving the square root operation out.
So, the objective function minimizes x and f (the state and control vectors) based on the difference between actual and reference state as well as the value of f, scaled to Q and R respectively, over k time steps where k is between 0 and m. Here, m refers to the prediction horizon, which is chosen by the user.
Both my Q and R are identity matrices right now, as I'm not sure how best to weight them.
Part 3. The Constraints Vector
In the same for loop, we can now add the constraints vector. For each iteration of the loop, the discrete dynamics equation from Part 1 is subtracted from the next state vector, x(k+1).
There are more constraints, however. Namely, friction cone constraints, governed by friction of the feet with the ground. As stated in the paper:
Here's my implementation below. Both the Fx and Fy inequalities must be repeated due to the absolute value function. Fz is not required here because it's extremely simple and can be implemented as an input constraint.
Part 4. Problem Setup
The solver used here is qpOASES. The optimization variables (in this case, the state and controls), objective function, constraint functions, and reference state are declared as shown. It is important to specify a bound tolerance and termination tolerance.
Part 5. Upper and Lower Bounds
The upper and lower bounds for the constraint vector still needs to be set. The dynamics equation is set up as an equality constraint, so the upper and lower bounds are both zero. The initial condition constraint is also an equality constraint, since the first state vector must be equal to the input state vector. The rest of the constraints have a lower bound close to infinity.
Constraints for the optimization variables are set separately. This is where the friction cone restraint for Fz comes into play. The reaction force perpendicular to the ground (assumed flat) must be upward, because the robot can only push itself away from the ground rather than pull itself toward the ground. Therefore, the lower bound on all Fz is 0.
Additionally, the MPC must check if each leg is actually in contact with the ground when calculating forces for it. If that leg is not in contact, all bounds are set equal to zero.
Part 6. Solving
Firstly, initial values must be fed into the solver.
I've set the initial values for my control inputs as an array of zeros, although I'm wondering if that should really be the last set of control inputs applied.
The initial values for the state are pulled in and calculated from the simulator. This 12x1 array is repeated N+1 times, where N is the prediction horizon, in order to fill the entire array including future states.
Finally, the initial states, bounds, and parameters are plugged into the solver. The controls vector is pulled from the solution, and the first row of optimized control values is returned. The rest of the values (representing future time steps) are ignored.
I'm currently working on the Python code to control a simulated version of my latest bipedal robot design in PyBullet. My focus over the past few weeks was getting the operational space control to work (many thanks to Travis DeWolf's incredibly helpful blog). After finally getting it to work properly, I have decided to share my math in the hopes of providing a useful example for anyone else having trouble with this. There really aren't enough resources on the internet that explain this in a succinct manner.
Solving for the transformation matrices and centers of motion correctly is the trickiest part of the process, and deceptively so. For more information on how to set up the transformation matrices, I recommend this Youtube tutorial and this blog post by Travis DeWolf.
I find it easiest to solve for the transformation matrices by "stringing" the robot out, as shown below, such that all of the joints' coordinate systems are oriented the same way. This saves you from having to perform additional linear algebra operations, which would raise your chances of making a mistake.
Shown below are the transformation matrices. The first transformation represents an x-axis rotation. See the first joint in the kinematic diagram above for reference. Transformations (2) through (4) are z-axis rotations. Despite the z-axis pointing up in the world coordinate system, the default angle (90 degrees) for the first x-axis rotation points the z-axis of the second through fourth joints parallel to the y-axis of the WCS.
Next, the locations for the centers of mass at each link are found, where l₀, l₁, l₂, and l₃ represent the distance of the center of mass from the base joint of each link, along a line connecting the base joint to the next joint. Keep in mind that the centers of mass may not be located directly along a line from one joint to the next in any given robot link, and that in this case I am able to make this approximation for the purpose of mathematical simplicity (but that may not always be the case).
The end-effector offset (Equation 9) should also be kept in mind, but in my case I don't need an offset from the end-effector. As such, I'm expressing it as zero.
Next, I solved for the full transformation matrices from the origin to each joint. This isn't necessary for the first joint, because it's already in the base coordinate system.
Now, the Jacobian for each COM and the end-effector can be found as a matrix of its partial derivatives with respect to each joint qᵢ. Both linear and angular velocities must be accounted for, so I have x, y, z, ωx, ωy, and ωz to deal with.
So far I haven't explained angular velocity. This I have solved separately, and it's quite simple as long as your robot is serial and doesn't have spherical joints. The partial derivative for the angular velocity w.r.t. each joint is expressed as 1 if that joint rotates along that axis, and 0 if it doesn't. For example, in J₀ (Equation 14), joint q₀ rotates around the x axis, so the partial derivative of ωx w.r.t. q₀ is 1. In J₁ (Equation 15), joint q₁ rotates about the z axis, so the partial derivative of ωz w.r.t. q1 is 1.
I can only show the first two Jacobians here, as the rest are too long to fit on the screen. You'll notice that the Jacobians for the centers of motion of each link (J₀, J₁, J₂, and J₃) are calculated as a function of the transformation matrix from the link's base joint to the origin multiplied by that link's center of motion. This confused me the first time around, and I don't want you, the reader, to make the same mistake.
For example, in the case of J₁ (the Jacobian for the second link), the transformation matrix from joint 0 to the world coordinate frame (origin) is multiplied by COM₁, the center of motion of the link between joints 0 and 1 (the first and second joints, so the second link).
For J₀ (the Jacobian for the first link), no transformation matrix is required because the first joint (q₀) is already in the base coordinate system.
I'm probably just confusing you more because my joint indexing starts at zero, but that's for programming purposes and is standard...
So there you have it. After several headaches, I was able to verify that this setup works in simulation. The chief cause of my problems was that I had solved for the transformation matrices incorrectly! Everything else is easy--though it looks intimidating, all of the symbolic math can be computationally solved by Matlab/Octave. Just remember, garbage in = garbage out.