Can DWB be tuned to penetrate deeper into the XY goal region before final rotation, without changing xy_goal_tolerance?

[maartenkoks]

Hi all,

I am doing a sim-to-real analysis of MPPI vs DWB on a skid-steer robot in a narrow warehouse task, and I am trying to understand whether DWB can be made to drive deeper into the XY goal region before final rotation, without tightening DWB’s own internal near-goal tolerance.

Context

• Platform: skid-steer / differential-drive robot
• Stack: Nav2 Jazzy
• Controllers compared: DWB vs MPPI
• Same robot, same task, same localization pipeline, same goals, same goal checker
• Goal checker: SimpleGoalChecker(xy_goal_tolerance=0.10, yaw_goal_tolerance=0.05, stateful=false)
• I am explicitly comparing native controller behaviour, so I am not using Rotation Shim or controller switching

Observed behaviour in the controller-visible map frame

• DWB sim: 29/30 success
• DWB real: 1/15 success
• MPPI sim: 29/30 success
• MPPI real: 15/15 success

Near-goal metrics:

• DWB median min_xy_err: 0.094 m in sim, 0.105 m in real
• MPPI median min_xy_err: 0.023 m in sim, 0.030 m in real

What I consistently see is:

• DWB tends to operate right at the XY tolerance boundary
• Near goal, DWB often becomes rotate-dominant while still having very little XY margin
• MPPI tends to drive much deeper into the goal region before final orientation correction

This makes DWB much more sensitive to small real-world effects such as skid-steer motion behaviour, wheel separation / ICR mismatch, low-speed rotation effects, and localization imperfections.

What I can already do
I was able to make DWB converge much deeper by reducing DWB’s internal xy_goal_tolerance from 0.10 to 0.05, while keeping the external goal checker at 0.10.

In a runtime single-goal probe, that gave:

• final_xy_error: 0.057 m
• min_xy_error_est: 0.053 m
• success: True

So tightening the controller’s internal near-goal threshold clearly changes the behaviour in the direction I want.

My actual question
I do NOT mainly want to know whether DWB can be made deeper by changing its internal xy_goal_tolerance, because I have already confirmed that.

What I want to understand is:

Can DWB be made to converge deeper before final rotation while keeping its internal xy_goal_tolerance unchanged?

In other words:

• is deeper XY penetration fundamentally tied to the internal threshold at which DWB switches into terminal behaviour
• or are there supported ways to get similar behaviour from the algorithm itself, without moving that threshold?

More concretely:

1. Is DWB fundamentally expected to stop around its internal xy_goal_tolerance and then rotate?
2. If I keep that internal threshold unchanged, are there tuning paths that can still make DWB translate deeper before final rotation?
3. If yes, which parameters matter most?
   • RotateToGoal
   • slowing_factor
   • GoalDist, GoalAlign, PathAlign
   • sampling / horizon settings
   • any other near-goal mechanism
4. Or is this effectively a structural limitation of DWB, where deeper terminal convergence mostly requires changing the internal threshold itself?

What I already tried

• Changing critic weights
• Adjusting RotateToGoal related settings
• Trying to bias the controller toward stronger translational convergence near goal

So far, the only thing that clearly changes the behaviour in the desired way is reducing DWB’s internal xy_goal_tolerance.

That is why my main question is whether I am missing an intended algorithmic tuning path, or whether this is simply not how DWB is designed to behave.

If useful, I can also share:

• a minimal config
• the runtime single-goal probe output
• short near-goal trajectory plots comparing DWB and MPPI

[SteveMacenski]

Check out #6052 and give it a review / test!

It will not just impact DWB but also MPPI and all other algorithms.