Map maintenance

Overview

A robot that repeats a taught route localises by matching the features in its current camera view against a stored set of features for the nearby part of the route 2603.12460 §III. Each run adds information about the scene, but the stored features also age. Light, weather and season change which features are visible, so a map that is never touched stops matching, while a map that takes in every new view accumulates noise 2603.12460 §I.

Map maintenance is the rule that decides, after each run, which stored features leave and which features from the current view enter. The rules differ in what they use as evidence: nothing, the newest view alone, the match outcome of each feature, or a model of when each feature tends to be visible 2603.12460 §IV-A.

The evidence every rule uses is the same. Each stored feature is labelled as matched correctly, matched incorrectly, or not matched, based on a horizontal shift estimated by histogram voting over the matched pairs 2603.12460 §IV 2603.12460 §III.

Variants

Static map

The map is taught once and never changed. This is the baseline. In the paper's day-to-night sessions, navigation on static maps was stable until the illumination changed significantly, and then it failed 2603.12460 §V. Card: Halodová et al. 2026.

Latest-view map

The whole local map is replaced with features from the current image. The paper expects a copy-of-a-copy drift, because the robot's heading is never estimated exactly, so new features sit slightly off from the originals 2603.12460 §IV-A. In the experiments, drift became visible only after many traversals, and the paper calls this rule's performance "rather low" 2603.12460 §V.

Replace-on-failure rules (aggressive, strict, summary)

Three rules remove features that were matched incorrectly or not matched, and add new ones from the current view 2603.12460 §IV-A:

• Aggressive: replaces every feature that was not matched or matched incorrectly.
• Strict: replaces only the features matched incorrectly, with the same number of new features.
• Summary: removes the features matched incorrectly and adds 10 % of the features from the current image, so the map grows over time.

The text gives no error values for these rules. Their results appear only in fig. 4 of the paper, and the text names only FreMEn and the score-based rule as the best performers 2603.12460 §V.

Multiple maps

When too few features in the current view match correctly, the current features are stored as an extra local map for the same place. At run time the map with the most matches is used 2603.12460 §IV-A. The paper does not state the threshold that triggers a new map, and this rule has no separate result in the text.

Score-based adaptive map

Each stored feature has a score. The score rises when the feature is matched correctly, falls when it is matched incorrectly, and falls by a smaller amount when it is not matched. After a map is used, the features with the worst scores are removed and the same number of features is added from the current view 2603.12460 §IV-A. The paper's experiments set the correct-match change and the incorrect-match penalty to 1, and the not-matched penalty to 0, with 500 features per local map and 5 % of features exchanged per run 2603.12460 §V. It was the second-best rule in the replay comparison, and in the field it guided the robot through sessions that ran from late afternoon into the night 2603.12460 §V. Card: Halodová et al. 2026.

New features are chosen by their descriptor distance to the nearest feature already in the map. Features far from the map are treated as more distinctive, and choosing by this distance beat random choice in the paper's comparison 2603.12460 §IV-B 2603.12460 §V.

Time-aware map (FreMEn)

The score is replaced by a per-feature model of visibility over time, fitted with spectral analysis, so that a feature's score can be predicted for the current time of day 2603.12460 §IV-A 2603.12460 fig. 2. Each match outcome is added to the feature's model at the time it happened. At run time, the features with the highest predicted scores are used. Removal uses each feature's mean score over time 2603.12460 §IV-A. In the replay comparison, it was the best of the rules tested 2603.12460 §V. The paper gives no harmonic count, window length, time units, prediction horizon or number of features used at run time, so the rule cannot be reproduced from the text alone 2603.12460 §IV-A.

When to use which

The paper's comparison covers one robot, one taught path and a forest park, so the ranking is not known to hold elsewhere 2603.12460 §V.

Implementation sketch

Score-based update, the common core of the adaptive rules. FreMEn replaces the score lookup with a model prediction at the current time.

# Score-based map update  [arxiv:2603.12460 §IV-A]
# n features are exchanged per run; the paper uses 5 % of the map  [§V]

def update_map(local_map, outcomes, view, n):
    # 1. Adjust each stored feature's score from its match outcome  [§IV-A]
    for feature in local_map:
        if outcome(feature) == CORRECT:
            feature.score += 1          # correct-match change  [§V]
        elif outcome(feature) == INCORRECT:
            feature.score -= 1          # incorrect-match penalty  [§V]
        # not matched: score unchanged (not-matched penalty 0)  [§V]
    # 2. Remove the n features with the lowest scores  [§IV-A]
    local_map.remove_lowest(n)
    # 3. Add n features from the view, most distinctive first  [§IV-B]
    for feature in most_distinct(view, local_map, n):
        # shift by the current horizontal offset before adding  [§IV-B]
        local_map.add(shift(feature, delta))
    return local_map

The initial score of a new feature is not stated in the paper, and removal depends on it, so record the value you use. The paper also does not state the local-map spacing or the bin width of the voting histogram 2603.12460 §III.

Contradictions

• Sign convention. The paper's text calls the incorrect-match and not-matched values penalties, but the experiment section gives them as positive magnitudes (1 and 0) without a sign 2603.12460 §IV-A 2603.12460 §V. An implementer has to pick the sign.
• Session count. The experiment section says "almost 200" sessions, and also gives 178 traversals and 5696 images from 32 local maps 2603.12460 §V. The two counts describe the same dataset in different ways and do not add up to one number.
• Public code. The paper describes eight update rules and FreMEn. The public repository, as checked for the card, has only the score-based rule and updates it on a distance trigger rather than on each local-map use. See the card's Paper vs code section. A reproduction from the public code therefore tests a different system from the one compared in the paper.

Papers

• Halodová et al. 2026, on predictive and adaptive map updates for long-term visual navigation: compares static, latest-view, aggressive, strict, summary, multiple-map, score-based and FreMEn updates for teach-and-repeat navigation over a three-month outdoor dataset.