M2-T4 — Layered Costmaps

Learning objectives

After this lesson, you should be able to:

• Describe the static, obstacle, inflation and social layers and what each contributes.
• Explain marking and clearing, and how raytracing removes obstacles that have left.
• Explain why ghost obstacles appear and list the usual fixes.
• Combine layers into a master costmap and say why order and update rates matter.

The intuition

Think of transparent sheets laid over a floor plan. The bottom sheet is the building: it does not change. The next sheet is drawn by the robot with a marker as its sensors see things, and wiped with a cloth where its sensors see through. Another sheet adds a halo around everything on the sheets below. A last sheet adds “do not go here” stamps, for example around a ward round in progress. The planner looks down through all the sheets at once.

Each layer is simple because it ignores the others. That is also why a mistake in one layer is easy to find: switch it off, and the problem vanishes.

The mathematics

The layers

Combining layers

Each cell of the master takes the largest cost any layer gives it, so no layer can hide another’s obstacle. Unknown has its own value, which the policy of lesson M2-T1 resolves.

Marking and clearing

For each sensor reading, the obstacle layer does two things:

1. Marking: the cell where the beam ended, if it ended on something within the marking range, becomes occupied.
2. Clearing: the cells the beam passed through, from the sensor to its end (a raytrace), become free.

A beam that returns nothing clears all the way to its maximum range and marks nothing. The ranges are separate parameters, so a robot may clear far but mark near. Common defaults are 2.5 m for marking and 3.0 m for clearing.

Why ghosts appear

• No beam passes through: the obstacle left but nothing looks through its cells (a blind spot, or the robot moved away).
• Clearing range too short: the robot marked the obstacle from 2 m but only clears to 1 m.
• Sensor field of view: a camera sees ahead only, so obstacles beside the robot are never cleared.
• Wrong hit flag: maximum-range readings reported as hits create phantom walls (lab M2-L3).
• Dynamic obstacles: a walking person leaves a trail of marks behind them unless the beams behind them clear it.

A worked example

A laser that clears. A 360° lidar with 0.5° steps puts 720 beams in a scan. At 4 m, neighbouring beams are 4 × tan(0.5°) = 0.035 m apart, closer than a grid cell of 0.10 m, so every cell between the robot and a wall at 4 m is crossed by at least one beam, and a moved cart disappears within one scan.

Where it breaks. A camera with a 70° field of view cannot clear the cells beside the robot. A cart parked there stays on the map until the robot turns towards it. The fix is to give the layer an expiry (observation persistence) so that old marks without fresh evidence fade, or to keep a lidar for clearing.

The hospital hallway. A cart that left at 13:00 remains because the only sensor that saw the hallway looks forward. The routes in the lab (M2-L3) show the repair: trace each beam, clear what it crosses, mark only a real return, and treat a beam without a return as free. Check the layer after a change by looking at a picture of it, not only at the planner’s behaviour.

Engineering implications and limitations

Debug one layer at a time

If a path goes the wrong way, publish and look at each layer’s own costmap before the master. The faulty layer is usually obvious.

Expiry is a safety net

Time-based expiry for obstacle marks catches the ghosts that clearing misses. Choose a time long enough that the robot does not forget a wall behind it, and short enough that a parked cart that has gone does not linger.

Rates

The static layer updates rarely, the obstacle layer at sensor rate, the master at a fixed rate. A slow update rate makes a moving robot plan on a stale map, which looks like clipping a corner it had “already seen”.

Layers for people and rules

A social layer raises costs around people and in front of them. A keep-out layer sets cells to lethal or to a speed limit. Both belong in layers, not in the planner, so they can be toggled by the operator and tested alone (lesson M2-T5).

Common misconceptions

Engineering takeaways

1. Layers keep sources separate: static, obstacle, inflation, and social or keep-out rules.
2. The master costmap takes the maximum per cell, so no layer hides another’s obstacle.
3. Marking sets the cell a beam ended in; clearing sets the cells it crossed.
4. Ghosts come from missing clearing, short ranges, narrow fields of view, wrong hit flags and moving obstacles.
5. Expiry and visual debugging, one layer at a time, are the practical cures.

References and further reading

The layered costmap paper and the Nav2 documentation.

• D. Lu, D. Hershberger and W. Smart, “Layered Costmaps for Context-Sensitive Navigation”, IEEE/RSJ IROS, 2014.
• S. Macenski, F. Martín, R. White and J. Ginés Clavero, “The Marathon 2: A Navigation System”, IEEE/RSJ IROS, 2020.
• Nav2 documentation, “Obstacle Layer Parameters” and “Voxel Layer Parameters”, docs.nav2.org.

Knowledge check

Three conceptual questions. Write an answer, then reveal the explanation. Your answers stay in your browser.

What’s next

Next: when a grid is the wrong map: roads, lanes and meaning.

Related content

Nothing below is required to finish this lesson.

• Before
  M2-T3: Distance Transforms and Euclidean Distance Fields

  Compute a distance field with a two-pass sweep and use it to find clearance and a gradient away from obstacles.

• Lab
  M2-L3: Find the Ghost Obstacles

  A cart left hours ago still blocks the hallway on the map: fix the marking and clearing logic.