MicroDuck SLAM & Odometry: Data, Replay and Accuracy Limits

What is new, and which version does it describe?

Pollen Robotics published odometry and SLAM engineering notes on October 23, 2026. Their author is Antoine Pirrone. The replay dataset was produced on October 14; it is newly connected here, not newly recorded today. Official stable daemon remains 0.17.8. The dataset names the separate maploc-v2 replay baseline 34b106c; do not treat the blog as an upgrade command.

Relative movement and a room map answer different questions

Contact odometry combines leg joint geometry with IMU orientation. It estimates motion relative to boot, so a carried robot can lose its room position. The mapper adds short-range ToF observations and matches at stops; its camera is not the mapping input. The external camera and floor tags belong to the evaluation rig. This distinction matters when selecting a recording or diagnosing a drifting trail.

The odometry report measures 20 walking stretches, 47.5 m total: endpoint error median 5.7 cm, worst 11.8 cm. Its median per-stretch percentage is 2.9%, while total error divided by total distance is 2.4%. Neither number is an absolute-room accuracy guarantee. These odometry numbers come from the blog’s separate evaluation, not a recalculation of the stored replay metrics below.

Read accuracy together with coverage

The fixed dataset card and its original s2 metrics / s3 metrics describe an alpha robot, policy velstand, driven with a pad in one living room.

The last row is conditional on tracking time. It does not mean 99% or 94% of the entire recording. The SLAM report also describes failure in an unknown room and longer lost periods without stops; autonomous multi-room exploration remains future work.

Start a replay review without a robot

1. Open the pinned dataset tree. Keep rig.toml, room/, the chosen session, its film/track.csv and ground_truth/clock.json together. State/ToF, camera and external truth use different timestamp fields; the supplied clock fit connects them.
2. Read sessions/s2/baselines/maploc/report/report.md before comparing a custom replay. Coordinates are metres, world z is up, and quaternions are [w,x,y,z].
3. Match replay source 34b106c and the included code/maploc_truth snapshot. The dataset documents these two commands from their respective project directories; adjust only the dataset path. They are reference commands, not a run performed by this site.

# From the fixed microduck checkout; replace this absolute path
SLAM_DATA="/absolute/path/SLAM_JJ_Salon_with_GT"
cargo run -p maploc --release --example replay -- "$SLAM_DATA/sessions/s3/session/20261014-172242-graphite-bang" --export "$SLAM_DATA/run.json" --grids
# Then enter the bundled toolkit and complete its documented setup
cd "$SLAM_DATA/code/maploc_truth"
uv run mtruth --rig "$SLAM_DATA/rig.toml" evaluate "$SLAM_DATA/sessions/s3/session/20261014-172242-graphite-bang" --track "$SLAM_DATA/sessions/s3/film/track.csv" --room "$SLAM_DATA/room" --replay "$SLAM_DATA/run.json" --out "$SLAM_DATA/report"

Keep loss/recovery events beside position and map scores. This site read the fixed card, code and small metrics files; it did not download the full recordings, compile the replay, reproduce scores or test hardware. The dataset is marked Apache-2.0. A personal HD1910 build needs its own sensor transforms and runtime contract: see replica source notes and DIY compatibility steps.

Where to go next.