Equivalence and Divergence of Bayesian Log-Odds and Dempster's Combination Rule for 2D Occupancy Grids

1. Problem Statement

Occupancy grid mapping is a cornerstone of mobile robotics, typically implemented using Bayesian log-odds (the standard in systems like GMapping and Cartographer) or Dempster-Shafer (DS) belief functions. While prior literature has suggested that belief function approaches offer advantages in dynamic environments and uncertainty representation, these comparisons suffer from a critical methodological gap: they fail to control for sensor model equivalence.

In previous studies, sensor parameters (e.g., log-odds increments for Bayesian vs. mass assignments for DS) were chosen independently for each framework. This lack of alignment meant that observed performance differences were often driven by the aggressiveness of the single-observation update rather than the fusion rule itself. The authors demonstrate that this confound can completely reverse conclusions; in their own preliminary experiments, unmatched parameterization favored DS by +12% in boundary sharpness, whereas matched parameterization reversed this to a -22% deficit for DS.

2. Methodology: Pignistic-Transform-Based Matching

To isolate the effect of the fusion rule, the authors developed a rigorous methodology to ensure per-observation decision equivalence between the two frameworks.

• Theoretical Foundation: The paper establishes a formal bijection ($\phi$) between Bayesian log-odds ($L$) and consonant Basic Belief Assignments (BBAs) on the binary frame $\{O, F\}$.
  • Equivalence Theorem: For a single observation, the Bayesian posterior probability $p = \sigma(L)$ is exactly equal to the Pignistic Probability ($BetP$) derived from the matched BBA.
  • Matching Procedure: Given a Bayesian log-odds increment $l$, the corresponding mass assignments are derived such that $BetP(O) = \sigma(l)$.
    • If $L \ge 0$: $m_O = 2\sigma(L) - 1$, $m_F = 0$, $m_{OF} = 2(1 - \sigma(L))$.
    • If $L < 0$: $m_O = 0$, $m_F = 1 - 2\sigma(L)$, $m_{OF} = 2\sigma(L)$.
• Experimental Design:
  • Simulation: Single-agent and multi-robot (3 robots) scenarios in procedurally generated 2D environments with dynamic obstacles.
  • Real Data: Two standard indoor LiDAR datasets (Intel Research Lab and Freiburg Building 079). Real-data multi-robot conditions were simulated via scan-splitting (dividing a single scan sequence among virtual robots) to ensure ideal alignment and eliminate pose estimation noise.
  • Fusion Rules Compared:
    1. Bayesian Log-Odds (with clamping).
    2. Dempster's Rule (DS).
    3. Yager's Rule (conflict assigned to ignorance).
  • Ablation Studies: Testing $L_{max}$ limits (including $\infty$) and DS regularization ($m_{OF, min}$) to identify the source of performance differences.
  • Metrics: Cell Accuracy, Boundary Sharpness, Brier Score, Map Entropy, and downstream A* path-planning clearance.

3. Key Contributions

1. Fair Comparison Framework: Introduction of a pignistic-transform-based methodology that matches sensor models at the per-observation level. This allows for the first controlled comparison where performance differences are strictly attributable to the accumulation/fusion rule.
2. Identification of the Confound: Demonstration that unmatched sensor parameterization is a massive confound capable of reversing performance rankings. The authors show that previous claims of DS superiority were likely artifacts of this parameterization mismatch.
3. Mechanism Analysis: Through ablation studies, the paper identifies that the performance gap arises not from regularization (clamping) but from Dempster's conflict normalization ($1/(1-K)$). This normalization causes slower geometric convergence to certainty at boundary cells where observations conflict, compared to the additive nature of Bayesian log-odds.
4. Comprehensive Empirical Validation: Validation across simulation, multiple sensor parameterizations, and two real-world datasets, consistently showing Bayesian superiority under fair matching.

4. Key Results

Under the fair comparison (Pignistic matching), Bayesian log-odds fusion consistently outperforms or matches Dempster's rule across all metrics and conditions:

• Directional Consistency: In 15/15 simulation runs and across all real-data configurations, Bayesian fusion was favored.
• Cell Accuracy: Bayesian achieved higher accuracy (e.g., +0.0013 to +0.022 difference on a [0, 1] scale).
• Boundary Sharpness: Bayesian produced sharper boundaries in all simulation conditions and the Freiburg dataset. On the Intel Lab dataset, differences were not statistically significant, attributed to clamping artifacts rather than intrinsic rule differences.
• Brier Score: Bayesian consistently yielded lower (better) scores.
• Ablation Findings:
  • $L_{max}$ Ablation: Removing the Bayesian clamp ($L_{max} = \infty$) did not eliminate the Bayesian advantage; in fact, the gap slightly increased. This proves the advantage is intrinsic to Dempster's conflict normalization, not the clamping mechanism.
  • Yager's Rule: Performed worse than DS on boundary sharpness, confirming that simply removing conflict normalization without changing the multiplicative structure is insufficient.
  • Sensitivity to Matching Criterion: When using Normalized Plausibility ($PPl$) instead of Pignistic ($BetP$) for matching, the boundary sharpness result reversed (favoring DS), but cell accuracy and Brier score still favored Bayesian. This highlights that the comparison is criterion-specific, but the overall trend remains robust.
• Downstream Impact: An A* path-planning evaluation showed functionally identical navigation outcomes (99.8% shared reachability, 76% path equivalence), indicating that the small probabilistic differences do not translate to practical navigation failures.

5. Significance and Implications

• Re-evaluation of Belief Functions: The paper challenges the prevailing view that Dempster-Shafer theory offers inherent accuracy advantages for standard 2D occupancy grid mapping. It suggests that previous reports of superiority were likely due to uncontrolled sensor model parameters.
• Practical Recommendation: For standard 2D binary occupancy grids with LiDAR, Bayesian log-odds is recommended over Dempster's rule due to:
  1. Equal or better point-probability accuracy.
  2. Lower computational cost (addition vs. multiplication + normalization).
  3. Lower communication bandwidth requirements (2 values per cell vs. 3 for DS).
• Scope of Belief Functions: The authors acknowledge that belief functions retain value for:
  • Interval-valued representations ($[Bel, Pl]$) for safety-critical planning (not evaluated here).
  • Multi-hypothesis frames (semantic mapping).
  • Open-world reasoning where new hypotheses can emerge.
• Methodological Legacy: The pignistic-matching methodology provides a reusable standard for future comparisons between Bayesian and non-Bayesian fusion frameworks, ensuring that future research isolates the fusion rule from sensor model artifacts.

In conclusion, the paper establishes that when sensor models are equated at the decision level, Bayesian log-odds is the superior or equivalent choice for 2D occupancy grid mapping, primarily because Dempster's conflict normalization introduces a convergence lag at boundary cells that degrades point-probability accuracy.