Skip to content

FenglongSong/Quadraped-kinematic-controller

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Kinematic walking controller

This is one of the assignments of 263-5806-00 Computational Models of Motion at ETH Zurich, Spring Semester 2021. The original code framework were implemented by Dongho Kang.


How to run the code

I strongly recommend to build in Release mode with following commands. Or you can directly build with Release mode in VS code. $ cd build $ cmake -DCMAKE_BUILD_TYPE=Release .. $ make


Exercise description

Following are the original assignment description by Dongho Kang.

In this assignment, we implement a kinematic walking controller for a legged robot!

Let's see the figure below.

figure: overview

Figure 1: The control pipeline of kinematic walking controller: 1) trajectory planning 2) computing desired joint angles by IK.

We start from five trajectories: 1 for base, and other 4 for feet. We plan the trajectories given target velocity of the robot base (body) and a timeline of the foot contacts (i.e. when and how long the foot is contacted with the ground.) The details of how we plan the timeline of foot contacts, and how we generate the target trajectories, are out of scope of this assignment. But in Ex.3, we will have a sneak peek of trajectory planning procedure for the robot's base.

Our ultimate goal is tracking all of the target trajectories at the same time. We will simplify this problem by assuming the robot's base (somehow...) always perfectly tracks the target base trajectory. Then, we want to find desired joint angles for each leg that allow the foot (i.e. the end effector of individual leg) to reach the target position. We can effectively formulate this as an IK problem. Since the robot has four legs, by solving four IK problems, we can obtain desired configuration of the robot.

figure: overview2

Figure 2: Don't freak out!, this is a skeletal visualization of our Dogbot. Assuming the base is at the target position, we want to find desired joint angles for each leg that allow the foot (i.e. the end effector of individual leg) to reach the target position. Note. You can render skeletal view by Main Menu > Draw options > Draw Skeleton.

Hand-in

Once you complete this assignment you should hand in

  • a short video demonstrating your baseline (Ex.1 ~ Ex.3) implementation. (filename: demo1.mp4)
  • a short video demonstrating Ex.5 implementation. (filename: demo2.mp4)
  • unit tests all passed (automatically checked by GitHub.)
  • code pushed to your github repository.

The grading scheme is as follows

  • baseline (80 %): based on your code and the demo video. Each exercise is counted as 20%.
  • advanced (20 %): based on unit test, code and demo videos. Each exercise is counted as 10%.

Please leave your questions on GitHub, so your colleagues also can join our discussions.

Exercises

Okay now let's do this step-by-step :)

Ex.1 Forward Kinematics (baseline)

In order to formulate an IK problem, firstly, we have to express the positions of a feet as functions of joint angles (and the base position). Formally speaking, given a generalized coordinates vector

equation: generalized coordinate

which is a concatenated vector of position of robot's base, orientation of robot's base and nj joint angles, we need to find a map between q vector and end effector position pEE expressed in world coordinate frame.

equation: forward kinematics

In the previous lecture, we learned how to find this map by forward kinematics.

Code:

  • Files:
    • src/libs/simAndControl/robot/GeneralizedCoordinatesRobotRepresenetation.cpp
  • Functions:
    • P3D getWorldCoordinates(const P3D &p, RB *rb)

Task:

  • Implement forward kinematics for point p.

Details:

  • GeneralizedCoordinatesRobotRepresenetation represents the generalized coordinate vector q.
  • P3D getWorldCoordinates(const P3D &p, RB *rb) returns position of a point in world coordinate frame. The arguments p is a position of the point in rigidbody rb's coordinate frame.
  • You may want to implement getCoordsInParentQIdxFrameAfterRotation(int qIndex, const P3D &pLocal) first: this function returns position of a point in the coordinate frame of the parent of qIdx after the DOF rotation has been applied.

Once you implement getWorldCoordinates function correctly, you will see the green transparent sphere around the feet of the robots.

figure: forward kinematics

Figure 3: Check if your implementation is correct. You should see the green spheres around the robot's feet.

Ex.2-1 Inverse Kinematics - Jacobian by Finite Difference (baseline)

Okay, now we can express the position of the feet as a function of joint angles. It's time to formulate an IK problem: we want to find a generalized coordinate vector qdesired given end effector target position pEEtarget.

equation: inverse kinematics

In the last assignment, we learn how to formulate the inverse kinematics problem as an optimization problem.

equation: inverse kinematics as optimization

We can solve this problem using gradient-descent method, Newton's method, or Gauss-Newton method. Whatever optimization method you choose, we need a Jacobian matrix of the feet point. Remember, Jacobian is a matrix of a vector-valued functions's first-order partial derivatives.

equation: jacobian

For now, we will use finite-difference (FD) for computing Jacobian. The idea of finite difference is simple. You give a small perturbation h around jth component of q, and compute the (i,j) component as follows.

equation: jacobian by finite difference

Code:

  • Files:
    • src/libs/simAndControl/robot/GeneralizedCoordinatesRobotRepresenetation.cpp
  • Functions:
    • void estimate_linear_jacobian(const P3D &p, RB *rb, Matrix &dpdq)

Task:

  • Complete estimate_linear_jacobian functions that computes a Jacobian matrix of position/vector by FD.

Details:

Ex.2-2 Inverse Kinematics - IK Solver (baseline)

Now, it's time to implement a IK solver. Choose one of optimization methods we learned in the previous lecture: gradient-descent, Neton's, or Gauss-Newton.

We solve four independent IK problems (one for each leg.) Let's say qdesired,i is a solution for ith feet.

equation: ik for ith foot

We can just solve each IK problem one by one and sum up the solutions to get a full desired generalized coordinates qdesired.

equation: full desired generalized coordinate

Code:

  • Files:
    • src/libs/kinematics/IK_solver.h
  • Functions:
    • void solve(int nSteps = 10)

Task:

  • Implement an IK solver based on gradient-descent, Neton's, or Gauss-Newton.

Details:

  • Use Jacobian matrix computed by gcrr.estimate_linear_jacobian(p, rb, dpdq) we implemented for Ex. 2-1.
  • It doesn't matter which optimization method you use for IK among aformentioned methods, but I strongly recommend to use Gauss-Newton method. It is easy to tune and has a better convergence. Remember, if your IK solver is not robust enough, you won't get full point from this exercise.
  • I left some useful hints on the source code. Please read it carefully.

Let's see how the robot moves. Run locomotion app and press Play button (or just tap SPACE key of your keyboard). Do you see the robot trotting in place? Then you are on the right track!

figure: trotting in place

Ex.3 Trajectory Planning (baseline)

Now, let's give some velocity command. Press ARROW keys of your keyboard. You can increase/decrease target forward speed with up/down key, and increase/decrease target turning speed with left/right key. You can also change the target speed in the main menu.

figure: main menu

Oops! The base of the robot is not moving at all! Well, a robot trotting in place is already adorable enough, but this is not what we really want. We want to make the robot follow our input command.

Let's see what is happening here. Although I'm giving 0.3 m/s forward speed command, the target trajectories (red for base, white for feet) are not updated accordingly. With a correct implementation, the trajectory should look like this:

figure: base trajectory

Code:

  • Files:
    • src/libs/simAndControl/locomotion/LocomotionPlannerHelpers.h
  • Class:
    • bFrameReferenceMotionPlan
  • Function:
    • void generate(const bFrameState& startingbFrameState)

Task:

  • Your task is completing generate function so that it updates future position and orientation according to our forward, sideways, turning speed commands.
  • Please record a video ~ 10 secs demonstrating your kinematic controller. The robot in the video should follow your keyboard command, and your video should be visible enough to see how well the robot tracks target trajectories. Upload it to this repository with a name "demo1.mp4". Do a double check if your video can be played without any additional encoding/decoding. It's also fine to upload your video to YouTube and add its link to on the top of this README.md. If I cannot play or access your video until the deadline, you won't get full points.

Details:

  • We plan the target base trajectory by numerical integration: we integrate forward, sideways and turning speed to get future position and orientation.

Once you finish this step, you can now control the robot with your keyboard.

By the way, planning the feet trajectories is a bit more tricky. I already implemented a feet trajectory planning strategy in our code base so that once you complete generate function, the feet trajectory is also updated by user commands. I will not explain more details today, but if you are interested, please read the paper, Marc H. Raibert et al., Experiments in Balance with a 3D One-Legged Hopping Machine, 1984. Although this is a very simple and long-standing strategy, almost every state-of-the-art legged robot still uses this simple heuristic. (Sidenote. Marc Raibert, who was the group leader of Leg Laboratory, MIT, later founded Boston Dynamics Company in 1992.)


From now on, we will improve our kinematic walking controller.

Ex.4 Analytic Jacobian (advanced)

Okay, we compute Jacobian matrix with FD. But in general, FD is not only inaccurate but also very slow in terms of computation speed. Can we compute Jacobian matrix analytically? The answer is yes. With a small extra effort, we can derive analytic Jacobian matrix, and implement this into our code.

Code:

  • Files:
    • src/libs/simAndControl/robot/GeneralizedCoordinatesRobotRepresenetation.cpp
  • Functions:
    • void compute_dpdq(const P3D &p, RB *rb, Matrix &dpdq)

Test: Compile and run src/test-a2/test.cpp. Test 4 should pass.

Ex.5 Uneven Terrain (advanced)

Our robot can go anywhere in the flat-earth world. But, you know, our world is not flat at all. Now, we will make our robot walk on an bumpy terrain. Imagine you have a height map which gives you a height of the ground of given (x, z) coordinates (note that we use y-up axis convention i.e. y is a height of the ground.) To make the robot walk on this terrain, the easiest way is adding offset to y coordinates of each target positions.

figure: terrain

You have a full freedom to choose the terrain map you want to use: you can just create a bumpy terrain by adding some spheres in the scene as I've done here. Or you can download a landscape mesh file in .obj format. Please figure out the best strategy to implement this by your own.

Task:

  • Make the robot walk on an uneven terrain (5%)
  • Visualize your terrain in locomotion app (5%).
  • Please record a video ~ 20 secs demonstrating your implementation. The robot in the video should walk on an uneven terrain, and your video should be visible enough to see how well your controller performs the task. Upload it to this repository with a name "demo2.mp4". Do a double check if your video can be played without any additional encoding/decoding. It's also fine to upload your video to YouTube and add its link to on the top of this README.md. If I cannot play or access your video until the deadline, you won't get full points.

Hint:

  • We want to give y-offset to IK targets. But remember, offset value for each target could be different.
  • Do not change double groundHeight = 0; in SimpleLocomotionTrajectoryPlanner. This will merely give the same offset to every target trajectory. We want to give different offset for each individual foot and the base.

Final Note

Congratulations! You can now control the legged robot! Hooray!

Can we use our kinematic walking controller to a real robot? Well... unfortunately it's not very easy. In fact, working with a real robot is a completely different story because we need to take into account Dynamics of the robot. But, you know what? We have implemented fundamental building blocks of legged locomotion control. We can extend this idea to control a real robot someday!

By the way, can you make a guess why using kinematic controller for a real legged robot doesn't really work well in practice? If you are interested in, please leave your ideas on the github issue.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published