knaxuslearn · build · ship

Roadmaps / Robotics / Phase 3

Phase 3 of 4 · Study weeks 9-12

Autonomous navigation

This phase makes the third checkpoint possible: a robot that drives to goals by itself and a measured account of when it does not.

Topics every week

  1. Week 98 h
    • Nav2 architecture and lifecycle nodes
    • Costmaps: static, obstacle and inflation layers
    • Localisation on a known map (AMCL)
    • Bringing the stack up

    Hand inThe robot localised on your saved map, with the costmap shown in RViz

  2. Week 108 h
    • Global planners
    • Local controllers (regulated pure pursuit, MPPI, DWB)
    • Velocity and acceleration limits
    • The robot footprint

    Hand inA comparison of two controllers on the same 10 goals

  3. Week 118 h
    • Behavior trees
    • Recovery behaviors
    • Waypoints and following several goals
    • The simple commander API

    Hand inA mission script of 5 waypoints with a recovery you triggered on purpose

  4. Week 128 h
    • Evaluating navigation: success rate over many runs
    • Classifying failures
    • Dynamic obstacles
    • The collision monitor

    Hand inCheckpoint 3, with the run table and the failure classes

Practice projects

  • Practice 1, the inflation radius (Week 9): vary the inflation radius and plot clearance against the number of goals reached. Done when: a table shows a value that is too small (contact) and too large (the robot refuses an aisle), and picks one with reasons.
  • Practice 2, the controller comparison (Week 10): run two controllers on 10 goals with 3 seeds each. Done when: a table gives success, time and path smoothness for both, and a sentence says when you would pick each.
  • Practice 3, the failure taxonomy (Week 12): classify every failed run from the 100-run table. Done when: each failure has a cause class and a bag, and the top two causes each have a fix tested on the same seeds.
  • Optional, a narrow door: add a door only 0.7 m wide. Done when: the robot passes at least 9 of 10 times, or you report what stops it.

GitHub repos to study

Status checked 22 Nov 2026 by shallow clone, as above.

RepoWhat to readWeek
ros-navigation/navigation2The bringup launch files, the costmap layers and the behavior trees9
ros-controls/ros2_control_demosA reference for how a robot's hardware interface is built10
ros2/rclcppExecutors and callback groups, to understand controller timing10
gazebosim/ros_gzSimulation bridge settings that affect sensor rates12

Milestone checkpoint

Run Checkpoint 3. Define the 20 goals and the 5 seeds first, and write your predicted success rate. Run all 100 and fill the table, with a bag for every failure. Classify the failures, fix the top two causes and rerun the same seeds. Add the moving obstacle at 1 m and measure the stopping distance. Hand in the run table, the failure classes, the before and after rates, ADR-5 (controller choice and its limits) and the stopping measurements. Report the success rate as a count out of 100, not a percentage alone.

Check yourself

Answer each question in your head first, then open it to check.

  1. What is the difference between AMCL and SLAM?
    AMCL finds the robot's position on a map you already have, using particles. SLAM builds the map and finds the position at the same time.
  2. Why inflate a costmap?
    The robot has size, and inflation turns the body's radius into extra cost around obstacles so the planner keeps clear of them. The radius must be at least the robot's circumscribed radius.
  3. What do the planner and the controller each do?
    The planner computes the whole path to the goal. The controller follows it and produces velocity commands many times a second.
  4. What does a behavior-tree recovery do, and why bound the retries?
    It tries something that can free the robot, such as clearing the costmap, backing up or spinning, then plans again. Unbounded retries hide a real failure and can trap the robot in a loop.
  5. Why report 95 of 100 and not just "95%"?
    The count shows how many runs there were. With 20 runs a 95% result could hide a true rate near 76%; with 100 runs the lower bound is near 89% (Wilson interval).