Robot Models, Cost Computers and Backtrackers

The parts that describe the robot and its motion: the robot model, the bounding volumes that limit where it plans, the cost computer that prices every edge, and the backtrackers that produce recovery motions. The classes are on the Plugin Interfaces and Built-in Plugins API pages, the keys under Robot Models, Bounding Volumes, Cost Computers and Backtrackers in the parameter reference.

Robot Models

The robot model is a shared service: generators sample and steer through it, and every layer checks traversability with it. A state is kTraversable, kPotentiallyTraversable (not yet observed well enough) or kUntraversable. A potentially traversable state counts as traversable only within safe_radius of the initial sensor position (default 0). Nodes must also keep the robot’s preferred obstacle margin.

SimpleRobot is a body of spheres checked against a TsdfMap’s distance field:

• Collision spheres (collision_sphere_{x,y,z,r}, base frame) must not collide, and nodes keep the extra collision_sphere_safe_margins. Contact spheres (contact_sphere_*) must touch an obstacle such as the ground. A state is kTraversable once min_observed_spheres spheres (default all) were observed passing.

• Sampling draws from the bounding volume, which must be active; sample_z: false keeps the initial height.

• Steering is a straight line: rotate to face the travel direction, translate at max_linear_velocity, rotate to the goal orientation. Start and goal must be level (zero pitch); a pitched goal throws.

• Body bounds (getBodyBounds()) are the box over every sphere grown by its safe margin; the generators re-check a node or edge only where the updated map region meets them (Generators).

SimpleFallbackRobot, the shipped robot, falls back to an OMPL planner (shipped RRTstar, 2 ms per query) when the straight line is blocked but both ends are traversable; random_seed ≥ 0 seeds OMPL (Random Seeding).

Bounding Volumes

SimpleRobot and every Evaluator build a BoundingVolume from the namespace their bounding_volume_args names: the robot’s (/map_bounding_volume) bounds where it plans, an evaluator’s (/target_bounding_volume) where gain counts. A volume is a stage mesh’s bounding box (mesh_folder, scene_id) grown by margin, as in the HSSD scenes, or the box x_min … z_max in the map frame, either rotated by rotation degrees about z. Unless max > min on every axis it is inactive and contains every point.

Cost Computers

The agent’s one cost computer prices every edge of every layer, recoveries included, in the budget’s unit. No cost computer has keys.

Type	Edge cost
IdentityCost	1 per edge
TrajectoryLength	Length of the sensor path [m]; a rotation in place costs 0, so a path of rotations alone has an infinite ratio
TrajectoryTime	Duration of the trajectory [s]; shipped

getCostLowerBound() is a cost no planned edge between two states undercuts, which lets a search skip planning an edge that cannot win: the straight-line distance for TrajectoryLength, the larger of that distance over max_linear_velocity and the yaw difference over max_yaw_rate for TrajectoryTime, and 0 otherwise.

Backtrackers

The BacktrackingLayer’s strategy for a recovery motion. Each recovery trajectory becomes one edge flagged is_backtrack, which backtrackers never undo.

Type	Recovery
RotateInPlace	Turns on the spot by rotation_step; at episode start it first sweeps initial_rotate_angle
Reverse	Replays the most recent executed edge backwards; remembers up to queue_size edges
RotateReverse	Rotates during the initial sweep and when no edge is remembered, else reverses the most recent edge. Shipped, with rotation_step: 0.3 and initial_rotate_angle: 6.28