MRPT CMonteCarloLocalization2D instability ("stuttering") when using CSimplePointsMap vs. OccupancyGrid

[hirenpatel1207]

I am currently working on 2D localization for an Automated Guided Vehicle (AGV). I am attempting to implement Monte Carlo Localization (MCL) using the MRPT library. I followerd the pf-localization (and partially modified pf_localization ros2 node to accept CSimplePoints map) example.
The Setup

Library: MRPT (specifically mrpt::slam::CMonteCarloLocalization2D and mrpt::bayes::CParticleFilter).

Map Type: The map is a 2D Point Cloud, which I am loading into the filter as a mrpt::maps::CSimplePointsMap. Note: I am not using an Occupancy Grid Map.

Inputs: Standard wheel odometry (Action) and 2D Laser Scans (Observation).

The Problem

I have successfully engaged the filter, but I am experiencing significant accuracy issues. Specifically, the estimated pose exhibits a "stuttering" or jittery effect, even when the robot is moving smoothly.

For comparison, I implemented a standard ICP (Iterative Closest Point) localization on the same data.
The ICP result is smooth

and stable, whereas the Particle Filter output is noisy and erratic

.

Configuration

Here is a snapshot of the map I am using:

So the question is why does Particle Filter based localization have this jitter, it is just wrong parameters? Or Particle Filters wont work with PointMaps since they dont provide a (smooth)likelihood over the entire map area?
Problem with ICP based methods is they are too sensitive to noise and if they loose track then they dont recover.