Simulation

A Gazebo simulation of Mote runs the same controllers, the same scan pipeline and — crucially — the same launch files as the real robot, so nothing can drift between sim and hardware. It needs no hardware, and its dependencies live in their own pixi environment (sim) so the robot install stays lean.

pixi run sim            # headless gz server + robot + controllers
pixi run sim-mapping    # the real mapping mission on top of it
pixi run sim-nav        # the real nav mission against the world's saved map
pixi run sim-test       # ~20 s headless smoke test (needs a GPU)
pixi run -e sim -- gz sim -g   # optional: attach the Gazebo GUI

The sim, sim-* and benchmark tasks are defined only in the sim environment, so pixi selects it for you. An ad-hoc command has to name it:

pixi run -e sim -- ros2 launch mote_bringup slam_launch.py use_sim_time:=true

How it delegates

sim_launch.py supplies a base — a Gazebo robot with gz_ros2_control in place of MoteHardware, a simulated lidar, and the /clock and /scan bridges — and then includes the real mission launch files with base:=false:

• mode:=mapping includes mapping_launch.py (that is pixi run sim-mapping)
• mode:=nav includes robot_launch.py (pixi run sim-nav)
• mode:=none, the default, is the base alone

So the missions are defined once. The sim also pulls controllers.yaml, laser_filters.yaml and localization_launch.py straight out of mote_bringup's share, and the URDF is the same xacro with use_sim:=true — without that flag its output is byte-identical to the robot's. The dependency runs mote_simulation → mote_bringup, never the reverse.

The world ladder

Worlds live in mote_simulation/worlds/ and get harder in order. Pick one with world:=:

pixi run sim world:=hospital_world.sdf

World	Tier	What it stresses
mote_world.sdf	easy	A simple walled room. The default, and the smoke test's world.
office_world.sdf	medium	A corridor of near-identical rooms — localisation.
hospital_world.sdf	hard	~58 × 38 m, a looping corridor grid, ~50 rooms, clutter.

The hospital world is generated by worlds/gen_hospital.py, which is committed beside its output — edit the script's layout and regenerate rather than hand-editing the SDF.

Every world has a sibling <world>.zones.yaml carrying the same waypoint zones (pickup, dropoff, home) plus a few room zones with footprints, so the fetch and goto missions run anywhere on the ladder with matching coordinates. The generator asserts that every zone clears the walls and the furniture, and lies inside its own footprint.

Sim maps are real site bundles

mote_simulation/sim_home/ is a committed, in-repo MOTE_HOME: one real site bundle per world (site name = world stem, floor ground) holding that world's SLAM map and zones. The sim environment points MOTE_HOME at it, so sim-nav loads a world's own map and never touches the robot's real ~/.mote.

Those maps are built the same way the robot builds one:

pixi run sim-map-world <world.sdf> [budget_s]

That launches the real mapping mission headless, runs the same autonomous coverage tool the robot runs (explore, with --sim-time), and saves into the world's site. Sim maps are ground-truth clean, so the save keeps the raw map_saver output — the FFT declutter pass is tuned for real sensor noise and would strip thin true walls. Re-running adds a revision.

Measuring, not just running

Three harnesses sit on top of the sim:

• Benchmark harness — scores the nav mission against Gazebo ground truth into metrics JSON plus a report.
• Parameter sweep — runs the benchmark once per parameter set and ranks them against the baseline, so a Nav2 parameter is changed on evidence.
• Bag-replay scoring — replays a real robot's mapping bag through SLAM under N parameter sets and scores it truth-free. The reality check that complements the benchmark; it needs no sim at all.

pixi run segment-eval scores room segmentation against worlds/<world>.rooms.yaml, the walkable rectangle of every enclosed room.

Isolation

Every entry point that starts a gz server — the benchmark, the smoke test, map_world.sh — claims a free ROS_DOMAIN_ID and GZ_PARTITION through one picker (mote_simulation/tools/sim_domain.py), so two runs on one machine stay separate. The sim environment also pins ROS_AUTOMATIC_DISCOVERY_RANGE=LOCALHOST, so a run neither advertises itself to nor discovers a robot on the LAN.

That makes the sim invisible to a default-range tool, which is why there are two RViz tasks: pixi run rviz for the robot, and pixi run rviz-sim joined to the sim's host-local graph.

Teardown is scoped the same way. Each launch is setsid-ed, so its session id is the exact reaping scope: it can clean up stragglers under any node name and reach nothing else. Bare name matches (pkill -f mote_world) are what that replaces — see clearing stray ROS processes.