Skip to main content

Navigation (Nav2): Moving with Purpose

Nav2 is the heir to the ROS Navigation Stack. It answers two questions:

  1. "Where am I?" (Localization)
  2. "How do I get there?" (Path Planning)

The Challenges for Humanoids​

Nav2 was designed for wheeled robots (TurtleBot). Humanoids are distinct:

  • They don't roll; they step.
  • They can step over small obstacles.
  • They can climb stairs.

Standard Nav2 treats everything as a 2D Costmap. For humanoids, we often need 3D Voxel Maps.

The Navigation Pipeline​

graph LR
    Map[Map Server] --> AMCL
    Lider[LiDAR Scan] --> AMCL
    AMCL{Localization} -->|Pose| Planner[Global Planner]
    Planner -->|Path| Controller[Local Planner]
    Controller -->|cmd_vel| Motors

1. Map Server​

Loads a map (usually a 2D occupancy grid .pgm file) into memory.

2. AMCL (Adaptive Monte Carlo Localization)​

This is a particle filter.

  • The robot starts with 1000 hypotheses (particles) of where it might be.
  • It reads the LiDAR scan. "I see a wall 2m away."
  • It checks the map. "Where on the map is there a wall 2m away?"
  • It keeps the particles that match and kills the ones that don't.
  • Eventually, the particles converge on the true location.

3. Global Planner​

"I am in the kitchen. I want to go to the bedroom."

  • Algorithm: A (A-Star)* or Dijkstra.
  • Output: A high-level path (a line of waypoints).

4. Local Planner (Controller)​

"I need to follow this line, but there is a cat in the way."

  • Algorithm: DWB (Dynamic Window Approach) or MPPI (Model Predictive Path Integral).
  • Output: Velocity commands (cmd_vel).
    • For wheeled robots: linear.x and angular.z.
    • For humanoids: Walking velocity vector.

5. Recovery Behaviors​

"I am stuck."

  • Spin: Look around.
  • Back up: Try to reverse.
  • Wait: Maybe the obstacle (person) will move.

Configuring Nav2 for Humanoids​

We need to tune the Costmap.

  • Inflation Radius: Don't get too close to walls.
  • Footprint: The shape of the robot. A simple circle is often enough, but a polygon is better.

3D Navigation​

For traversing stairs, we replace the 2D Costmap with an Octomap or Voxel Grid. We use planners that understand height (Z-axis).

Isaac ROS Nvblox builds a 3D costmap on the GPU, allowing us to spot hanging obstacles (like a table edge at head height) that a 2D LiDAR at knee-height would miss.

Practical Example​

To launch Nav2:

sudo apt install ros-humble-navigation2 ros-humble-nav2-bringup
ros2 launch nav2_bringup bringup_launch.py map:=/path/to/my_map.yaml

You can then give a "2D Nav Goal" in Rviz by clicking on the map. The robot (simulation) should plan a path and follow it.