Monte Carlo localisation

Particle-filter implementations usually document their tuning with a recorded bag and a screenshot of RViz, which says nothing about whether the numbers hold on a different seed or a different floorplan. I wanted an MCL implementation where every claim about accuracy was reproducible.

• Implemented the odometry motion model, a likelihood-field range-finder model and adaptive low-variance resampling over a known 2D occupancy grid, following Thrun, Burgard and Fox.
• Kept the core free of any ROS dependency: the node translates messages into plain structs at the boundary, which is what makes the models unit-testable.
• Computed the likelihood field with an O(n) exact distance transform rather than a per-beam ray cast.
• Shipped an offline evaluation harness that drives a simulated robot through the map against known ground truth, so accuracy is reported per seed instead of per screenshot.
• Ran the same harness over six floorplans to find where the tuning stops transferring.

• 0.055 m mean final error over a 20.25 m route, 8/8 seeds converged within 0.50 m, against odometry drifting 0.31–0.95 m.
• 8/8 seeds on every one of six floorplans, 0.055–0.132 m mean final error.
• 200 particles for tracking; 5,000 for global localisation on the shipped map.
• 83 GoogleTest cases, building and running without ROS.

Personal project, solo, in the same repository as the DWA local planner. This page is built from my own report; the derivations and every measurement behind the tuning are in the repository docs.

• Global localisation is tuned to the shipped map: unreliable at 2,000 particles and below, and on five other floorplans 5,000 particles converge only 3–6 times out of 8.
• KLD-adaptive sampling and random particle injection are not implemented, so the filter cannot recover if it locks onto a wrong mode.
• The reported pose is the mean over all particles, not the dominant cluster, so it is meaningless while the cloud is multi-modal. The published covariance is what makes that detectable.
• No scan matching or motion-model adaptation, so fast rotation or wheel slip beyond the noise model will lose the lock.
• Measurement updates run on the latest scan at a fixed rate without interpolating odometry to the scan timestamp: fine at walking speed, a source of bias at speed.
• Single-threaded, O(particles × beams) per update.