Evaluators and Sensor Models

How a candidate’s gain is estimated: the sensor model simulates the view from a node, and the evaluator turns it into gain anchors and a frontier flag, refreshed as the map grows. The classes are under Evaluators and Sensor Model, the keys under Evaluator, SimulatedSensorEvaluator and Sensor Models.

Gain

• The SensorModel ray-casts a view from the node’s sensor pose, and evaluator::SimulatedSensorEvaluator keeps the visible voxels inside the target bounding volume (bounding_volume_args). The evaluator turns them into the node’s gain anchors; Node::getSingleNodeGain() sums them.

• A path’s gain sums its nodes after the start, each node once. With the path planner’s is_gain_correlated, which task:=points sets, an anchor shared by several nodes of the path also counts once.

• By default a node the robot has reached is spent (Node::markObserved()) and is never evaluated again.

• A node is a frontier node (is_at_frontier) when more than frontier_min_unknown_ratio of the voxels it sees are unknown and its view exceeds frontier_min_visible_ratio times the largest view seen so far this episode. Both default to 0, which turns detection off; the shipped task evaluator sets 0.8 and 0.6. Frontier nodes feed expected_gain, table_gain, SPT and TSP (Path Planners).

• getRealAccumulatedGain() measures what the episode has actually gained, logged as <layer>.RealAccumulatedGain and as real_accumulated_gain in episode_summary.yaml.

The Evaluator of Each Task

Every shipped planner reads its evaluator from /task_evaluator, set by task:=:

Task (task:=)	Evaluator	A node’s gain	Real accumulated gain
Volumetric exploration (explore, default)	NaiveEvaluator	Unknown voxels in view	Explored volume [m³]
Surface reconstruction (recon)	SurfaceFrontierEvaluator	Unknown voxels in view, weighted by frontier_voxel_weight next to the mapped surface, else new_voxel_weight	Mapped surface area [m²]
Point collection (points)	PointsEvaluator (open_apero_ros)	One anchor per observed (voxel mapped free), uncollected target point within its collection radius, shared by every node in range; collected once an executed waypoint puts the sensor within the radius	Summed gain of the collected points

SimpleYawPlanningEvaluator wraps a SimulatedSensorEvaluator (following_evaluator_args) and keeps each tree node’s best orientation; the rrt and rrtstar configs use it. VoxelWeightEvaluator also credits occupied voxels by how much a new view would add to their TSDF weight; no shipped setup uses it. PointsEvaluator’s CSV input is in the open_apero_ros README.

Gain Suppression

With the path planner’s gain_suppression_radius above its default 0, each search first suppresses every node cluster (the nodes at one sensor position) closer than the radius to a higher-scoring kept cluster (Node::is_gain_suppressed). Suppressed nodes stay in the graph as waypoints worth 0; the start node is never suppressed. The beam searches, SPT and TreeBestValue honor it; TSP does not.

Re-Evaluation on Map Updates

Each cycle PlanningLayer::update() hands every node cluster to Evaluator::updateEvaluation(). The simulated-sensor evaluators re-evaluate a traversable node whose gain is above update_gain or that is a frontier node, and, with only_update_for_updated_map (default true), only if its view box overlaps the map region updated since the last cycle (Map::getUpdatedMapRegion()).

Sensor Models

Rays run from the sensor pose to ray_length or the first occupied voxel, which counts as visible; a camera with observation_boundary: planar, shipped to match the depth camera and Voxblox, ends them at camera depth ray_length instead. run_experiment_sim.launch sets ray_length (shipped 3.0 m) and observation_boundary from its arguments, as for Voxblox; a robot integration sets them to what its map integrates with. downsampling_factor (shipped 3.0) sets the ray spacing at the observation boundary in voxels, trading evaluation time against accuracy. Every sensor model requires an OccupancyMap.

Type	Model	Rays
SimpleRayCaster	Pinhole camera (CameraModel)	One fixed grid
IterativeRayCaster	Pinhole camera (CameraModel); shipped 480 × 640 px, focal_length 625.58 (42° horizontal)	Coarse first, doubling the density in each farther distance section
IterativeRayCasterLidar	Spinning LiDAR (LidarModel), field_of_view_x × field_of_view_y degrees	As IterativeRayCaster