rove_slam

Custom lidar-inertial SLAM for the Clubcapra Rove rover. C++17 front-end (KISS-ICP + GTSAM), Python tooling for the rosbag converter and viewer. No ROS dependency in the C++ tree — sensors arrive over UDP at runtime, or from file-backed recordings offline.

Architectural shape ported from a prior DJI M400 + L2 deployment; phase plan and platform-specific deviations are written up in docs/phase-0.md.

Status

Phase 3.0 — GTSAM ISAM2 pose-graph wiring (byte-identical to Phase 1).

Phase	What	Branch	Gate
0	ISensorSource, ReplaySource, LiveSource, wire format	phase-0	—
1	KISS-ICP front-end, TUM + PCD output	phase-1	ROVE_SLAM_USE_KISS_ICP=ON (default)
3.0	ISAM2 + BetweenFactor pose graph, byte-identical to P1	phase-3	ROVE_SLAM_USE_GTSAM=ON (default OFF, ~6 min compile)

Validation:

• docs/phase-1-validation.md — KISS-ICP across the bag fleet.
• docs/phase-3-validation.md — pose-graph byte-identity vs Phase 1.

scripts/smoke.sh runs end-to-end (build, convert, replay, live, SLAM) in <15 s on x86_64.

$ scripts/smoke.sh
── convert
  LIDAR       335 vs rosbag 335
  IMU         1757 vs rosbag 1757
  GPS         1759 vs rosbag 1759
── live (fake VN-300)
  slam_live: vn_pkts=149 imu_emitted=149 gps_emitted=149 …
── slam (KISS-ICP)
  poses       335 vs rosbag 335 lidar frames
  drift       0.039 m (expect <1 m on a clean loop)
[smoke] PASS

KISS-ICP processes ~335 Livox MID-360 scans in ~0.7 s wall on 8 threads; the moving_short_bag2 loop closes at 0.17 % drift end-to-start.

Target platform

• Compute: Jetson-class aarch64 Linux on the rover; x86_64 Linux for dev/replay.
• Lidar: Livox MID-360, read straight off raw UDP — no Livox SDK link. Per-point timestamps preserved when the source provides them (Livox CustomMsg, or PointCloud2 with a per-point timestamp field).
• INS: VectorNav VN-300 (accel + gyro + lat/lon/alt + heading) over UDP via the rover's rove_sensor_api on data port 5000. GPS is bundled in the VN data packet, not a separate driver.
• Bag input: ROS 2 rosbag2 SQLite (Humble), topics /livox/lidar, /livox/imu, /imu/data, /fix. Converter is a one-shot Python tool using the rosbags pip package — no ROS install needed.

Repo layout (planned)

rove_slam/
├── include/rove_slam/        # ISensorSource, sinks, wire format
├── src/                       # ReplaySource, LiveSource, SlamPipeline
├── apps/                      # slam_offline, slam_live CLIs
├── tools/                     # Python: bag2recording.py, rerun viewer
├── tests/                     # fixtures + smoke scripts
└── docs/                      # phase-0.md, frame conventions, wire format

Build

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j8
ctest --test-dir build --output-on-failure

System deps: libeigen3-dev. Python deps for the converter:

python3 -m venv .venv && .venv/bin/pip install rosbags

KISS-ICP is on by default (-DROVE_SLAM_USE_KISS_ICP=OFF to skip — Phase 0 build only). Pulled via FetchContent at v1.2.3 — newer tags require CMake 3.24+. GTSAM lands in Phase 3 behind ROVE_SLAM_USE_GTSAM=ON.

Run SLAM on a converted recording:

slam_offline /tmp/rec --slam --out /tmp/out --map
# writes /tmp/out/trajectory.tum and /tmp/out/map.pcd