A-OctoMap: Adaptive 3D Volumetric Mapping

• A-OctoMap is a volumetric mapping framework that fuses multi-modal sensor data using Bayesian log-odds updates to reduce occlusions and improve map accuracy.
• It utilizes an adaptive hierarchical octree structure and geometric downsampling to balance high-fidelity 3D representation with computational efficiency.
• Integrated with efficient online path planning algorithms, A-OctoMap enhances obstacle recovery and dynamic navigation in complex robotic environments.

A-OctoMap is an advanced volumetric mapping framework that extends the standard OctoMap octree representation in robotics to enable adaptive, high-fidelity 3D environment modeling and efficient online motion planning. It comprises two distinct but related lines of research: one focuses on probabilistic fusion of multi-modal sensor data to reduce occlusions (Strong et al., 2021), while the other centers on hierarchical adaptive resolution and efficient path planning through downsampled but geometrically faithful occupancy grids (Mao et al., 2024). Both leverage octree structures but introduce novel algorithms for sensor fusion, adaptive downsampling, and multi-resolution search, enabling improved map completeness, precision, and computational efficiency in complex robotic scenarios.

1. Probabilistic Sensor Fusion for Occluded Environment Mapping

A-OctoMap incorporates an octree of cubic voxels with a fixed minimum resolution ($r = 0.04$ m) and extends the standard OctoMap framework with a unified Bayes-filter update operating in log-odds form. Each voxel’s state $L(n)$ maintains the log-odds of occupancy: $L(n) = \ln\frac{P(n)}{1-P(n)}$ where $P(n)$ is the posterior voxel occupancy probability. When a new sensor measurement $z_t$ arrives, the log-odds update for a voxel $n$ is: $L(n|z_{1:t}) = L(n|z_{1:t-1}) + L(n|z_t)$ and $P(n|z_{1:t})$ is recovered via

$$P(n|z_{1:t}) = 1 - \frac{1}{1 + \exp(L(n|z_{1:t}))}$$

A-OctoMap fuses input from both an external depth camera (DC, e.g., Kinect-V2) and one or more onboard proximity sensors (PS, e.g., VL53L1X), each modeled probabilistically. Both sensor types use range-dependent hit-updates and constant miss-updates, applied via ray-casting along each sensor beam. The overall recursive fusion equation is: $$L(n|z_{1:t}) = L(n|z_{1:t-1}) + L(DC_t) + \sum_{k=1}^{M} L(PS_{k,t})$$ with $L(n)$0 the number of proximity sensors. All beams are assumed independent given the map. Proximity sensors are treated as perfectly collimated, and potential sensor correlation is ignored.

2. Sensor Model Formulations and Parameterizations

Each sensor modality is encoded with its own hit/miss update functions. For the VL53L1X proximity sensors, the hit log-odds update for a voxel at distance $L(n)$1 is: $L(n)$2 The external depth camera's model is: $L(n)$3 Occupancy and free-space state thresholds are set via

$L(n)$4

Voxel size, log-odds slopes, and thresholds are directly tied to sensor calibration and typical operating ranges, supporting map accuracy and tunable trade-offs for application needs (Strong et al., 2021).

3. Adaptive Hierarchical Octree Structure

A-OctoMap generalizes the standard OctoMap to a binary ($L(n)$5-ary) tree structure augmented with hierarchical boundary fields—specifically, the “SplitBoundary” and “NodeBoundary” records per node, which respectively store the coordinates of parent split axes and the tight bounding box of included points. This enables precise adaptive spatial subdivision:

• At construction, the maximum tree depth is set in terms of the robot’s minimum controllable region (MCR),

$L(n)$6

• The required octree depth $L(n)$7 is determined as

$L(n)$8

with $L(n)$9 as map side length, $L(n) = \ln\frac{P(n)}{1-P(n)}$0, and $L(n) = \ln\frac{P(n)}{1-P(n)}$1 (default).

This adaptivity allows finest resolution near regions critical for the robot’s dynamic constraints, and coarser mapping elsewhere, balancing accuracy and computational resources (Mao et al., 2024).

4. Geometric Downsampling and Error Bounds

A-OctoMap achieves high-fidelity surface representation during map downsampling through a convex hull–preserving criterion:

• For each leaf node $L(n) = \ln\frac{P(n)}{1-P(n)}$2 with points $L(n) = \ln\frac{P(n)}{1-P(n)}$3, the tightest axis-aligned bounding box is computed and its convex hull $L(n) = \ln\frac{P(n)}{1-P(n)}$4 formed. Points interior to $L(n) = \ln\frac{P(n)}{1-P(n)}$5 are discarded as redundant.
• Remaining points are partitioned by split boundaries into up to $L(n) = \ln\frac{P(n)}{1-P(n)}$6 sub-regions. In each sub-region, the convex hull $L(n) = \ln\frac{P(n)}{1-P(n)}$7 is computed.
• The final downsampled set $L(n) = \ln\frac{P(n)}{1-P(n)}$8 comprises the union of convex hull vertices across all child subregions:

$L(n) = \ln\frac{P(n)}{1-P(n)}$9

• The Hausdorff error between the original surface and this downsampling is bounded by the cell diameter ($P(n)$0 MCR), ensuring no loss of critical geometric information within each region (Mao et al., 2024).

5. Online Path Planning with Adaptive Grids

The adaptive octree induces an occupancy grid $P(n)$1 of non-uniform cell size matched to local MCR, supporting discrete search algorithms. Jump Point Search (JPS), a widely used grid path planning algorithm, is performed on $P(n)$2:

• Each leaf node’s occupancy is defined by intersecting its split and node boundaries and testing for obstacles.
• Standard JPS pruning rules are retained, but connectivity is determined by the non-uniform adaptive grid.
• The result is a significant increase in successful pathfinding in challenging environments, especially narrow passages, as demonstrated in simulation: adaptive grids improved feasible path recovery by ≈28% and reduced average path length by >6% relative to uniform grid baselines (Mao et al., 2024).

6. Computational and Quantitative Evaluation

A-OctoMap’s primary computational characteristics are:

• Tree build:
  • $P(n)$3 with $P(n)$4 input points and $P(n)$5 leaves
  • Example: 2-D, 35K points at depth 7 in 5.04 ms (i9-13900H CPU)
• Downsampling:
  • $P(n)$6 average per leaf
  • Largest observed retention: 27–48% post-sampling, at 101–212 ms for 300K–900K input points at depth 7
• Path planning (JPS over $P(n)$7 m, ≈1.2M points)
  • Adaptive grid: 441/1000 feasible paths (fixed grid: 343), $P(n)$8 ms end-to-end

For sensor fusion, simulation with 34 proximity sensors and one depth camera (tree resolution 0.04 m) for occluded object mapping demonstrated that A-OctoMap achieves higher accuracy in occupied/free voxel recovery compared to depth or proximity sensing alone—recovering 1430 occupied voxels versus 970 with depth only, and reducing missed occupied cells by >1000 at a small increase in false positives (Strong et al., 2021).

7. Limitations and Future Directions

Known limitations include:

• No current modeling of beam divergence or sensor field-of-view (e.g., ~27° for VL53L1X); this could be improved via angular uncertainty modeling.
• Downsampling is strictly geometric, with no semantic or learned partitioning.
• Pathological distributions can yield $P(n)$9 convex hull construction in rare leaf nodes.
• Large-scale scenes (>10M points) may induce high per-leaf memory usage.
• Both hit-log-odds decay and downsampling criteria are simple, linear, and could benefit from data-driven or probabilistically optimal enhancements.

Recommended extensions are incorporation of learned or adaptive sensor models, richer partitioning strategies (e.g., semantics-aware), and explicit modeling of real-world sensor uncertainties for even higher map fidelity (Strong et al., 2021, Mao et al., 2024).