Regional Path Planner

NVIDIA Docs Hub NVIDIA Isaac AMR
Isaac SDK v2.0 Regional Path Planner

Overview

The regional path planner (RPP) connects the global route planner and the local trajectory planner. It converts the waypoint sequence from the global route planner into a smooth and safe path that avoids obstacles within the sensing range. The local trajectory planner then uses this path to guide the robot.

Inputs:

PoseTree: This library provides the pose frame transform queries.
Route: This message from route planner provides the global guidance for path searching.
Evidence Grid Map: This message from onboard perception shows the distance to obstacles in real time.
Occupancy Map: This map represents the pre-mapped environment occupancy.
Semantic Map: This map contains the human annotated environment context.
Planner Operation Mode: This message from the behavior executor controls the path searching mode.

Outputs:

Regional Path: The smooth and safe path that the trajectory planner uses to guide the robot.
Path Health: This message indicates the status of the regional path solving process and whether a rerouting request is needed.

Components

The regional path planner consists of four major components that are easy to customize and configure:

Problem Formulator: This component defines the path problem based on the robot model, search space representation, cost formulation and waypoint management.
Path Solver: This component solves the path problem using a selected algorithm and searches the optimal regional path for trajectory optimization.
Path Smoother: This component optionally smooths the solved path to make it more friendly for trajectory optimization.
Failure Handler: This component handles any failures that may occur during the path search process. It can provide a fallback path or trigger a rerouting request.

The following diagram shows how these components work together:

Problem Formulator

Path problem formulator helps to define the path problem in a generic way that can be easily configured and used by different path solvers. It consists of four subcomponents:

Robot Model

This subcomponent defines the robot’s properties and motion capabilities. It uses the following information:

Shape: The robot’s shape as a polygon and a list of disks for efficient collision detection.
Motion: The robot’s kinematics as either a point model or a motion primitive model.
Limits: The robot’s constraints and limits such as maximum curvature, etc..

With this information, the robot model can validate state transitions and generate feasible motions for path searching.

Space Discretizer

This subcomponent divides the search space into discrete configurations. It can use two different methods depending on the environment:

Cartesian coordinate: For unstructured space with uniform coverage.
Frenet coordinate: For structured space with a reference path.

Each discrete grid has a unique id for cost caching. The robot model can check the motions within the grids.

Waypoint Manager

This subcomponent manages the waypoints within the search range and identifies the target waypoint for path searching. It performs the following steps:

It chooses a target waypoint from the route by calculating where the route and search boundary cross.
It verifies that the target waypoint is clear of obstacles using the obstacle maps.
If valid, it commits the target waypoint. If not, it either reports a failure or searches for another waypoint to substitute.

Cost Manager

With the robot model and the space discretizer, this subcomponent is to compute the cost based on the navigation objectives.

Table 1 *Cost Components and Objectives*
Cost	Objective
Travel Distance	Produce shorter path
Motion	Penalize the unwanted motion segment, e.g., max curvature, lateral acceleration, etc.
Waypoint Guidance	Encourage the path to follow the guidance of waypoint sequence.
Path Flickering	Stabilize the paths between consecutive cycles.
EGM Clearance	Penalize the proximity to the real-time perceived obstacles.
Semantic Distance	Penalize the proximity to the annotated high-risk regions (e.g., stairs).
Occupancy Map	Penalize the proximity to the pre-mapped occupancy grid.

Path Solver

This subcomponent finds the best regional path by using a chosen search algorithm (e.g., A*). It uses the robot model’s motions to explore the search space without a predefined graph. It also stores the cost of each node it visits until it reaches the target waypoint.

Path Smoother

This subcomponent can smooth the solved path as an extra refinement to make the path less jagged and the rotations less sharp.

We can select different smoothers based on the criteria. Some are model-based (e.g., iLQR), which respect the robot constraints but require more computation. Others are for signal denoising (e.g., Savitzky-Golay filter), which are more lightweight but do not ensure feasibility.

Failure Handler

This subcomponent helps the regional path planner deal with blocked routes. It tries these strategies:

Fallback Search: This strategy expands the search space before asking for a rerouting.
Rerouting: If fallback search fails too, this strategy requests a global route search that can find a different way.

The path health and the rerouting request are sent to the behavior executor to work better with other navigation modules.