Using PaGMO for industrial arm trajectory optimization?

[rkliman]

Over the past few months, our team has been attempting to use PaGMO to help improve the capabilities of a real-time path planning application for robotic industrial arms, specifically within ROS. While using it, we have found that most of these algorithms are quite slow compared to the one we are using. Specifically, we have been using algorithms in PaGMO that are single-objective, constrained, and non-stochastic due to the requirements of the use-case. Additionally, we have found that there is a lot of tuning needed to get good results. This seems infeasible for the scaling ability of our project, and we were curious if there are any improvements we could make or things that we are missing which could make PaGMO a feasible for our use. The parallelization ability and ability to handle high degree-of-freedom use cases is very attractive to use but we have not been able to fully leverage PaGMO to get these results.

For reference, we are trying to integrate PaGMO into ROS Industrial's trajopt_ros library.

[darioizzo]

If you have good algorithms for constrained non linear optimization, please feel free to suggest them, we would then consider adding them.

In pagmo, for your particular application, I would personally use an Interior point method (ipopt) or a sequential quadratic programming method (snopt) but if you have better ones, feel free to propose them.

Also, keep on mind that the performance of algorithms in inverse dynamics and trajectory optimization station is highly dependent on the transcription used.