kiss-imu-hilti-encoder

KISS-IMU Encoder — Hilti SLAM Challenge 2023 (Alphasense IMU)

A learned IMU correction/uncertainty encoder (KISS-IMU IMUNet: 6-ch CNN + 2×GRU with accel/gyro correction heads and per-axis uncertainty heads), trained self-supervised against Hilti LiDAR odometry via the --use-gt path.

From raw 400 Hz 6-axis IMU it produces, per ~0.1 s window:

• a 128-d motion feature (encoder embedding),
• bias/noise-corrected accel & gyro (residual corrections added to raw),
• per-axis covariance (calibrated aleatoric uncertainty),
• (via pypose) a short-horizon relative pose preintegration.

Intended use: an uncertainty-aware IMU front-end for a fusion/SLAM backend (IMU preintegration factors weighted by the learned covariance), or motion- feature extraction. See imu_encoder.py.

Training data & supervision

Hilti SLAM Challenge 2023 handheld sequences (Alphasense IMU 400 Hz + Hesai PandarXT-32 10 Hz). Supervision = the provided frame-to-frame LiDAR odometry (relative_poses.npz) expressed in a gravity-aligned, z-up IMU body world frame (LiDAR→IMU extrinsic estimated from data by gyro/LiDAR hand-eye; consistent euler ≈ (-178°,-1°,-89.5°) across all sequences).

• Train: site1_handheld_1,2,3,4, site2_handheld_4 (~9.3k windows)
• Validation (held out): site1_handheld_5

Metrics & honest assessment

Baseline = raw IMU strapdown integration (the Alphasense IMU is high quality).

The encoder improves per-window relative motion and adds calibrated uncertainty — the signal a fusion backend actually consumes. It does not beat raw IMU on long-horizon pure dead-reckoning, because the --use-gt label (raw LiDAR odometry) carries a slow vertical (z) drift that a point-correction network partially fits. The principled fix is the full self-supervised pipeline (ICP + pose-graph optimization fuses IMU+LiDAR to produce gravity-consistent, drift-reduced labels) — supported in this repo (--use-gt off) once a live LiDAR-odometry backend is wired in.

Usage

from imu_encoder import IMUEncoder        # uses modeling_imunet.py
import numpy as np
enc = IMUEncoder("best_model.ckpt", device="cuda:0")
acc = np.zeros((40,3), np.float32); acc[:,2] = 9.81   # (T,3) m/s^2, IMU body frame (gravity incl.)
gyr = np.zeros((40,3), np.float32)                    # (T,3) rad/s
feat = enc.encode(acc, gyr)            # (D,128) motion features
out  = enc.correct(acc, gyr)           # corrected acc/gyr + per-axis covariance
mot  = enc.preintegrate(acc, gyr, dt=np.full(40,1/400.,np.float32))  # rel pose

Architecture

Conv1d(6→32, k=10, s=5) → GELU → Dropout → GRU(32→64) → GRU(64→128) → 4 MLP heads {accel-correction, gyro-correction, accel-logvar, gyro-logvar}.

License

BSD-3-Clause (academic). Based on KISS-IMU (Choi et al., ICRA 2026).

<!-- ml-intern-provenance -->

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern