optimal hitting configuration

[anas-abulehia]

Hello ,
I started working on this recently, I considered the given literature as a reference for my work, I have the following question

1. Cartesian Trajectory: I have used similar approach to the presented in the paper. Line -> arc -> hit. how did you define the stopping trajectory , how did you define the stopping point.
2. I have found the best configuration for the hitting movement as presented in 'Efficient and Reactive Planning...' paper , but result is not satisfactory since it has elbow up to down switching when move from start position to targect position. I suspect the optimization pick another configuration ( how can I debug this) maybe it is intial quess issue. see the figure, i think this configuration is correct it gives the desired point in cartesain, but i think it is a bad one
   
3. point 2 is very importnant to find the hitting speed.

[PuzeLiu]

Hi,
Thanks for paying attention to our paper.

1. In that paper, the stopping point is chosen based on heuristics. So we concatenate another line-arc segment to a stopping point that tries to go to the table's center.
2. The optimal hitting configuration is determined without consideration of the initial point. A simple solution would be to add a penalty that keeps the configuration close to the initial one in the objective. We would also be happy to see new solutions!
3. In the baseline example, we have simplified the optimization and omitted Linear Programming Part.

[PuzeLiu]

Can you show me the command you run?

You should set the argument "n_episodes" to be larger than the "n_cores"

[anas-abulehia]

I ran this command
python run.py --example baseline -e 3dof-hit -n 1 -r ====> gave the above error
python run.py --example baseline --n_cores 1 --n_episodes 5 ===> gave a report,
how can i get baseline live visual demo ?

[PuzeLiu]

You can try python
python run.py -r -e 3dof-hit --example baseline --n_cores 1 --n_episodes 5

This will not wrong the whole baseline but only hit the agent example. There will be no difference in this scenario.

[anas-abulehia]

switching to another topic,
I have seen the basline implementation of trajectory generator. It updates it every 5 steps.
I have my trajectory in cartesian from start to end. I think computing the inverse kinematics for the whole trajectory and then send each q, q_dot to the robot will lead to a deviated final endeffector position subsequecntly a very bad hitting situation.
I have tried to understand it from the base line but couldn't figure out the used logic.
It would be great to explain to me this point.

the figure shows the hitting path , computed based on circles and lines sofar I have cartesian trajectory, start and hitting configuration.

[PuzeLiu]

Hi, Have you tried to check the trajectory in the joint space? For example, the joint velocity. If your trajectory is very fast and violates the joint velocity limit, the controller will truncate the joint trajectory output or may have a bad interpolation. This is the major reason that we propose a QP optimization.

[anas-abulehia]

This was the question: I have followed the paper, which outlines the following tasks:

1. Create a path from the starting point to the hitting point (x,y) in Cartesian coordinates.
2. Find the best hitting configuration.

Now, I have a starting and hitting configuration, as well as Cartesian coordinates for the trajectory. I am trying to convert this to joint space by formulating an optimization problem that will solve for q and q_dot every time draw action is executed.

However, I have encountered two problems with this approach:

1. Proper discretization of the Cartesian points to guarantee a solution. I am unsure how the baseline discretizes the Cartesian trajectory.
2. The approach is rigid and does not account for the actual end-effector position. Instead, it relies solely on the Cartesian trajectory. The baseline has addressed this issue, but it is unclear to me how.

alternatively, in the paper you mentioned v_desired in Weighted Null Space Optimization section of paper, how did you get v_desired for the trajectory ?

[PuzeLiu]

Hi, You will need to properly interpolate the cartesian trajectory as well, i.e., compute the via point and interpolate between them. What I did in the past was to build a quartic polynomial to fit the time-phase (t-z) trajectory. Then I'm able to get dx / dt = dx / dz * dz / dt. What you could also try is to set v_desired = K (x_{t+1} - x_{t})