Tuning DWA Planner using trajectory costs

[Sarika Ramroop]

Hi, 

I was having a lot of trouble tuning the DWA planner in ROS Kinetic for a pretty large robot and part of my problem was that I had no idea how the costs were affecting the trajectories that were being chosen. Having gone through the nav stack source code, I realise now that some things didn't quite work how I thought they did. 

To make tuning easier for myself, I modified the DWA planner code to publish the trajectory costs for each set of trajectories. I:

- created a new message type to store all the costs for a single trajectory (overall costs and the sub-costs such as obstacle costs, goal costs etc.)

- created another message type to store an array of all the trajectories generated at a particular instance

- indicated which trajectory was the best in a the set

Taking this data, I was able to publish the costs of the best trajectory, as shown in the below example:

Best trajectory has index: 102  # index of this trajectory, in the trajectory array

Overall cost: 455.06

Oscillation costs: 0

Obstacle costs: 5.06

Goal front costs: 150

Alignment costs: 0

Path costs: 0

Goal costs: 300

Twirling costs: 8.23756e-320

--------------------------------------- 

--------------------------------------- 

I was able to use this to determine which set of costs were having the largest influence on the selected trajectory being chosen, and which probably needed a higher weighting. Is this a feature that might be useful to PR into the nav stack? I'm mainly just trying to get some feedback into whether other people would use this since I've seen a number of questions on ROS answers with tuning-related issues, such as the below:

https://answers.ros.org/question/209287/base_local_planner-does-not-follow-the-global-plan-accurately/

https://answers.ros.org/question/9795/tuning-navigation-robot-keeps-rotating-when-close-to-its-goal/

Thanks in advance for any feedback!

[David Lu!!]

Hi Sarika,

I share a lot of your frustrations with DWA and the inability to look at the values of all the trajectories evaluated. I ended up doing something quite similar to what you did, and it lead to me writing DWB Local Planner. 

https://roscon.ros.org/2017/presentations/ROSCon%202017%20Fundamentals%20of%20Local%20Planning.pdf

https://vimeo.com/236487972

Best,

David!!

--
You received this message because you are subscribed to the Google Groups "ros-sig-navigation" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ros-sig-navigat...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.