amcl-tuning

Start all alpha values at 0.2 and adjust based on symptoms. Increase alpha1 and alpha4 if particles diverge during turns, and increase alpha2 and alpha3 if divergence happens during straight-line driving. Alpha5 only applies to omnidirectional robots.

The likelihood_field model is fast and recommended for most cases but ignores max-range and short readings. The beam model is physically accurate and models all phenomena but is slow because it ray-casts at every update.

Corridors lack unique scan features to disambiguate particle poses, so the filter cannot determine position along the corridor axis. Add distinctive environmental features, increase max_particles, or set an accurate initial pose to mitigate this.

Reduce max_beams from 60 to 30 and lower max_particles from 2000. Also check update_min_d and update_min_a, since overly small values trigger constant filter updates even for minor robot jitter.

Enable beam skip in dense, cluttered environments with many dynamic obstacles where single bad scan rays corrupt the measurement update. Set do_beamskip to true with beam_skip_distance of 0.5 and beam_skip_threshold of 0.3.