Microstrain 3DM-GX3-45 + consumer GPS (NCLT Segway)¶

Platform: Segway RMP campus rover (University of Michigan NCLT dataset) Status: Field validated. 12 sequences, 940 minutes total, scored against the dataset's RTK GPS ground truth. Current results: Benchmark Results.

Sensors¶

Sensor	Model	Notes
IMU	Microstrain 3DM-GX3-45	Accelerometer and gyro fused at 100 Hz. It also has a magnetometer, which the NCLT player does not publish. Its orientation output is published as invalid, because its roll and pitch are the accelerometer's own tilt.
GPS	Consumer-grade receiver (NCLT `gps.csv`)	Reports a 3 m sigma; measured errors are far wider, see the config comment below.
Wheel odometry	Segway RMP encoders	Differentiated from odometry at 100 Hz.

IMU datasheet specs used: - Angular random walk: 0.07°/√hr → noise density ≈ 0.003 rad/s at 100 Hz - Velocity random walk: 0.03 m/s/√hr → noise density ≈ 0.001 m/s² at 100 Hz

Config¶

This is the exact configuration used for the NCLT benchmark. Identical across all 12 sequences, no per-sequence tuning.

fusioncore:
  ros__parameters:
    base_frame: base_link
    odom_frame: odom
    publish_rate: 100.0

    # Microstrain 3DM-GX3-45: 6-axis, no magnetometer.
    # Yaw initializes from GPS motion after ~5m of travel.
    imu.has_magnetometer: false
    imu.gyro_noise: 0.003           # rad/s : ARW 0.07 deg/sqrt(hr)
    imu.accel_noise: 0.1            # m/s^2 : conservative for campus vibration
    imu.remove_gravitational_acceleration: false  # driver publishes raw specific force

    # Segway RMP wheel encoders, differentiated from odometry_mu_100hz.csv.
    # Loosen vel_noise if your surface is rough or wheels are pneumatic.
    encoder.vel_noise: 0.05         # m/s
    encoder.yaw_noise: 0.02         # rad/s

    # Consumer GPS: stated sigma 3m, but NCLT measured p95 errors range
    # from 9.7m (clean day) to 53.1m (bad multipath day).
    # Adaptive noise compensates automatically: gnss.base_noise_xy is the
    # floor, not the ceiling.
    gnss.base_noise_xy: 3.0         # m  : matches the stated sigma
    gnss.base_noise_z: 5.0          # m

    # Which of the two quality gates below actually runs depends on your driver.
    # If the NavSatFix carries a non-zero position_covariance (most drivers do),
    # FusionCore gates on the receiver's own reported sigma in METRES and the
    # max_hdop/max_vdop lines are never consulted. They are the fallback for
    # drivers that publish no covariance at all. Before 0.3.6 the DOP numbers
    # were compared against metres, which silently rejected every fix from a
    # receiver reporting honest sigma; see docs/configuration.md.
    #
    # Leave the sigma gate at its defaults (25 m / 50 m) unless you have logged
    # what your receiver actually reports. It is deliberately permissive: a gate
    # tuned to clean-sky numbers throws away the degraded fixes you most need.
    # gnss.max_sigma_xy: 25.0         # m : reject a fix reporting worse than this
    # gnss.max_sigma_z:  50.0         # m
    gnss.max_hdop: 10.0             # fallback only: lenient, NCLT doesn't publish HDOP
    gnss.min_satellites: 0          # lenient: NCLT doesn't publish sat count
    gnss.min_fix_type: 1
    gnss.heading_topic: ""          # single-antenna: no heading from GPS
    gnss.azimuth_topic: ""

    gnss.lever_arm_x: 0.0
    gnss.lever_arm_y: 0.0
    gnss.lever_arm_z: 0.3           # GPS antenna ~30cm above base_link

    outlier_rejection: true
    outlier_threshold_gnss: 16.27   # chi2(3, 0.999)
    outlier_threshold_enc: 11.34    # chi2(3, 0.999)
    outlier_threshold_imu: 15.09    # chi2(6, 0.999)

    gnss.coast_n: 3                 # enter coast mode after 3 consecutive rejections
    gnss.coast_q_factor: 10.0       # inflate Q_position 10x during coast
    gnss.coast_timeout_s: 30.0      # also coast if GPS silent for 30s

    adaptive.imu: true
    adaptive.encoder: true
    adaptive.gnss: true             # key: adapts to actual GPS noise, not just stated spec
    adaptive.window: 50
    adaptive.alpha: 0.01

    ukf.q_position: 0.01
    ukf.q_orientation: 1.0e-9
    ukf.q_velocity: 0.1
    ukf.q_angular_vel: 0.1
    ukf.q_acceleration: 1.0
    ukf.q_gyro_bias: 1.0e-5
    ukf.q_accel_bias: 1.0e-5
    ukf.q_encoder_wz_bias: 1.0e-7

    input.gnss_crs: "EPSG:4326"
    output.crs: "EPSG:4978"
    output.convert_to_enu_at_reference: true
    reference.use_first_fix: true

Topic remaps¶

The NCLT benchmark uses fusioncore_datasets/nclt_player.py which publishes to default topics. For a real Segway RMP or Microstrain setup, remap as needed:

ros2 launch fusioncore_ros fusioncore.launch.py \
  fusioncore_config:=/path/to/this-config.yaml \
  --ros-args \
  -r /imu/data:=/microstrain/imu/data \
  -r /odom/wheels:=/segway_rmp/odom \
  -r /gnss/fix:=/gps/fix

Validation results¶

Measured 2026-10-26 on all twelve sequences with this config: scored at GPS fix times, it does not beat raw unfiltered GPS (raw GPS lower on 11 of 12, one tie at 9.186 m against 9.187 m). With gnss.reanchor_enabled: true added, it is lower than raw GPS on 2 of 12 held-out sequences (2012-03-31 by 22.0%, 2012-10-28 by 5.6%). Full table, pipelines and scoring: Benchmark Results.

The robot_localization comparison that used to be on this page ("+55% better" and similar) is withdrawn. It was measured through a dataset player with two input errors that fed both filters wrong data, and the corrected comparison is being re-run.

Reproduce one sequence through the ROS node:

bash tools/run_nclt.sh 2012-01-08 <output_dir> 1.0

Adapting this config to similar hardware¶

Different GPS (standard consumer, 3-5m CEP): Keep gnss.base_noise_xy: 3.0 and adaptive.gnss: true. The adaptive window adjusts upward automatically.

Better GPS (F9P RTK float, ~0.4m CEP): Set gnss.base_noise_xy: 0.5, gnss.min_fix_type: 3.

Different IMU with similar specs (VectorNav VN-100, Lord 3DM-GX5-25): Change imu.gyro_noise to match the ARW from your datasheet: σ ≈ ARW_deg_per_sqrthr × (π/180) / 60 × √sample_rate.

Mecanum or holonomic platform: Remove motion_model: "DifferentialDrive" (or set to "CVA") since the robot can slide laterally.