Real-time applications

[Adoplot]

I am interested how fast is your algorithm. Can I use it to solve inverse kinematics of 6DoF Kuka robot in real-time (<16ms for every frame) ? I managed to test it using urdf file. I receive end-effector position data every 4ms from udp socket, solve IK and send joints' parameters back. I achieved IK solver speed of 200-300ms. Am I doing something wrong, or it is normal for such type of algorithm?

Thank you for your work! I appreciate it!

[Phylliade]

Hello @Adoplot,

This is a though question: Every IK computations involve maths optimization that are pretty heavy, thus inducing minimal computation time per Chain.inverse_kinematics call.

There are numerous ways to solve this issue, but this depends on your context:

• Decrease the computation time to compute an IK, using the max_iter parameter. Setting it to a lower value (like 1000 instead of the default 15000) will decrease computation time. However this can reduce precision. A good use-case is when you have a trajectory and just use IK to "leapfrog" between different positions (ie do small jumps between different IK targets, instead of going to completely new targets). In this case, do not forget to also use the initial_position parameter.
• Instead of doing an IK per frame, just do less IKs. This is possible in the case you have a trajectory broken down into multiple IK targets. You can decrease the number of targets that compose your trajectory, having a little less granularity in the sampling, but this will still work.
• In a similar manner, you can also cache the IK computations: Instead of recomputing an IK already computed, you can just reuse it.
• Use more powerful hardware: Having a raspberry pi or an i9 would induce different computation times for the IK

Can you give more information about your issue, so we can find a solution that suits your need?

[Adoplot]

Hello @Phylliade,

Thanks for your fast reply.

As you said, I need a robot to do small jumps between close IK targets. I achieved 80-110 ms computation speed on i5-9300H processor. I tried decreasing max_iter parameter, but the result was not much different (70-100ms).

I tried to use initial_position parameter but I can't understand how to use it properly. As I understand, initial_position is a parameter of Chain.inverse_kinematics_frame, not the Chain.inverse_kinematics

For testing I have an array of 4x4 homogeneous transformation matrices (for inverse_kinematics_frame). I solve IK for 100 matrices and divide overall computation time by 100 to get average value. It works ok. But how to use initial_position parameter? I understand that I need to write there joint parameters of the robot from previous frame, but I don't know what syntax to use. Could you show an example, please?

Thank you!

[alxy]

You can pass initial_position to the inverse_kinematics function as well, it will be forwarded to the lower level functions.

You need to pass it the joint positions of a previous run, so just somehow cache it between two calls:

ikResults1 = armChain.inverse_kinematics(translation1)
ikResults2 = armChain.inverse_kinematics(translation2, initial_position=ikResults1)

[Adoplot]

@alxy, Thank you! I managed to decrease average computation time to 40-50ms. However, it is not enough.

Thanks for help!