Autonomous Mobile Robot Navigation Stack

The problem

The fleet needed an operator intervention roughly once every X hours. Most were the same failure: the planner treated a hand-tuned inflation radius as ground truth, so pallet wrap, floor drains, and low-hanging shrink film were either invisible or treated as walls.

What I built

An end-to-end autonomy stack: 2D LiDAR SLAM for localisation, Nav2 for global and local planning, and a learned traversability costmap trained on field data that replaced the inflation heuristic. A behaviour tree handles recovery so the robot degrades gracefully instead of stopping dead.

How it works

• C++17 ROS 2 nodes; the control path runs on a PREEMPT_RT kernel with the executor and node composition tuned to hold the latency budget.
• SLAM Toolbox for mapping and localisation, with a re-localisation routine for the docking approach.
• Traversability model in PyTorch, trained on labelled LiDAR + camera field recordings, exported to ONNX and served on the robot.
• Nav2 with a custom controller plugin and a BehaviorTree.CPP recovery tree.
• Sensor fusion (EKF) over wheel odometry, IMU, and LiDAR scan matching.

How it was tested

A hardware-in-the-loop rig replays recorded field rosbags on every merge and fails the build on a regression in intervention-triggering events. Controller changes are checked against a fixed set of simulated scenarios before they reach a robot.

Results

• Operator interventions cut from 1 per X hours to 1 per Y, across N customer sites.
• Docking accuracy held under Z cm.
• Perception-to-planning latency reduced from X ms to Y ms after restructuring the executor.
• Migrated ROS 1 → ROS 2 with zero fleet downtime.

What I learned

• The learned costmap was only trustworthy once I added a floor on it. A model that is confidently wrong about a drain is worse than a crude inflation radius, so the planner now clamps how much it will trust the network.
• Most of the real work was data, not modelling — the active-learning loop over field failures moved the metric far more than any architecture change.
• I underestimated how much the HIL rig would pay for itself. It should have been the first thing built, not the third.

My role

I owned the perception-to-planning interface and the costmap model, and led the ROS 2 migration. Two other engineers worked on the docking and fleet-management sides.