M6-T7 — Choosing a Local Planner

Learning objectives

After this lesson, you should be able to:

• State what DWA, TEB, MPPI and regulated pure pursuit each optimise and how.
• Compare their computational cost with a simple count.
• Match a planner to the robot’s kinematics, environment and available computing.
• Describe the characteristic failure of each and the first thing to check.

The intuition

Picture four drivers on the same road. One looks at a handful of steering-and-pedal combinations for the next second and takes the best (DWA). One draws the whole route for the next few seconds as a stretched elastic band, with times along it, and relaxes the band until the route is short, quick, smooth and clear (TEB). One imagines thousands of slightly random futures, scores them, and steers by the average of the good ones (MPPI). One simply follows the road ahead and slows for corners and for things nearby (regulated pure pursuit, RPP).

All four work. Which one is right depends on the road. A narrow, winding corridor, a robot that cannot turn on the spot, an embedded computer with little to spare, and a hall full of moving people each favour a different driver.

The mathematics

What each one optimises

How they work, in one sentence each

• DWA (Fox, Burgard and Thrun, 1997): sample velocities reachable now, drop those that cannot stop (lesson M6-T3), score the rest.
• TEB (Rösmann and colleagues): minimise a weighted sum of penalties over a band of timed poses using sparse graph optimisation; the band is re-optimised at each cycle starting from the last solution.
• MPPI (Williams and colleagues): add noise to a nominal control sequence, roll out K samples, and update the sequence by an exponentially weighted average of the noise by the costs.
• RPP (Macenski and colleagues, 2023): pure pursuit’s curvature 2·y/L² towards a look-ahead point, with speed reduced in tight curves and near obstacles.

A cost count

Compare candidates by this product before comparing quality. It also tells you which knob to turn first when the CPU is saturated.

Typical failures

A worked example

Counting the work. A DWA configuration of 20 speeds × 40 turn rates, 30 prediction steps, at 20 Hz:

MPPI’s rate is 7 times DWA’s here. The numbers are illustrative: a rollout step of MPPI can be vectorised or run on a GPU, which is what makes it practical, whereas DWA’s step count is small enough for a single core. Run the actual count on your hardware, with your costmap lookups, before deciding.

Choosing for Scout. Scout is a differential-drive robot with an 8-core embedded computer. It drives 60 m corridors, passes through 0.9 m doors, and shares the space with staff.

Scout’s default is RPP on a global path that already respects the footprint and the lanes, with the collision monitor as a safety net. Where people fill the corridor, switch to a sampling planner for that zone. The decision is by zone, not for the whole building.

The first check. When a robot drives badly, the planner is the last place to look. Compare footprints and costmaps (lab M1-L1), confirm the localisation is stable (lab M7), and check the rates and timestamps. Only then retune.

Engineering implications and limitations

Choose by constraint first

Kinematics rule out planners (a car-like robot cannot use unconstrained arcs). Computing rules out others. Quality decides what is left.

Keep the safety net separate

Run an independent collision monitor that stops or slows the robot from raw sensor data, whatever the planner commands. A planner’s model of the world can be wrong; the monitor’s rule is simple (lesson M7-T1).

Tune with scenarios

Fix a set of scenes: a door, a bend, a person crossing, a dead end. Measure the metrics of lesson M1-T6 on each after every change. Never tune on one corridor.

Plan for switching

A navigation stack lets you choose a controller per zone or per task. Define zones in the map (lesson M2-T5) and switch with the behaviour tree, so the corridor controller and the crowd controller can differ.

Mind the plugins

Nav2 ships DWB, RPP and MPPI controllers. TEB comes from ROS 1’s teb_local_planner and is available as a community port for ROS 2. Check that the version you rely on is maintained.

Common misconceptions

Engineering takeaways

1. DWA samples velocities, TEB optimises a timed band of poses, MPPI averages sampled rollouts, RPP follows a path with regulated speed.
2. Compare cost with rate × samples × steps × cost per step.
3. Kinematics and computing narrow the choice; the environment decides among the rest.
4. Each planner has a typical failure and a first thing to check.
5. Keep an independent safety layer, choose per zone, and tune on scenarios rather than one corridor.

References and further reading

The papers behind each planner, and the Navigation 2 documentation.

• D. Fox, W. Burgard and S. Thrun, “The Dynamic Window Approach to Collision Avoidance”, IEEE Robotics & Automation Magazine 4(1), 1997.
• C. Rösmann, W. Feiten, T. Wösch, F. Hoffmann and T. Bertram, “Trajectory Modification Considering Dynamic Constraints of Autonomous Robots”, German Conference on Robotics (ROBOTIK), 2012.
• G. Williams, P. Drews, B. Goldfain, J. M. Rehg and E. A. Theodorou, “Aggressive Driving with Model Predictive Path Integral Control”, IEEE ICRA, 2016.
• S. Macenski, S. Singh, F. Martín and J. Ginés, “Regulated Pure Pursuit for Robot Path Tracking”, Autonomous Robots 47, 2023.
• Nav2 documentation, “Controller Server” and the plugin pages for DWB, MPPI and Regulated Pure Pursuit, 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 module: staying safe when things go wrong, stopping distances, uncertainty and recovery.

Related content

Nothing below is required to finish this lesson.

• Before
  M6-T6: Social Navigation

  Describe personal space, passing side and the freezing-robot problem, and how costs can encode them.