KISS-IMU: Self-supervised Inertial Odometry
with Motion-balanced Learning and Uncertainty-aware Inference

🏆 IEEE ICRA 2026 Award Finalist

Jiwon Choi, Hogyun Kim, Geonmo Yang, Juhui Lee, Younggun Cho^†

🤖 Spatial AI and Robotics Lab (SPARO)

<Keep IMU Stable and Strong>

---

📰 News

• 🏆 May 6, 2026. Selected as an IEEE ICRA 2026 Award Finalist.
• 🎉 Jan 31, 2026. Accepted to IEEE ICRA 2026.

💡 Overview

KISS-IMU denoises raw IMU streams against a self-generated LiDAR-odometry pseudo-label. A GMM-based motion-balanced sampler with a frequency gate keeps under-represented motion regimes from being drowned out during training.

🚀 Getting Started

🧪 Just want to see it run? See RUNNING.md for a self-contained synthetic-data demo that needs no external dataset and no GPU, plus environment/CPU notes and the src/data/ data contract.

🐳 Docker (recommended)

The image sparolab/kiss-imu:v1.0 ships with all runtime dependencies (CUDA, PyTorch, pypose, kiss-icp, small_gicp, pygicp, scikit-learn).

git clone https://github.com/sparolab/KISS-IMU.git
cd KISS-IMU

Edit docker/docker-compose.yml and replace {dataset_folder} under volumes: with the absolute path to your datasets (e.g. /mnt/hdd/datasets:/storage1). Then launch:

docker compose -f docker/docker-compose.yml up -d
docker exec -it kiss-imu-ws bash

Inside the container, pwd is already /home/test_ws/src:

bash scripts/train.sh

🛠️ Native

pip install -r requirements.txt
bash scripts/train.sh

📂 Dataset Layout

One top-level data_root with one sub-directory per sequence:

📁 <data_root>/
└── 📂 <SEQUENCE_NAME>/
    ├── 📄 imu.csv
    ├── 📄 gt_pose.csv
    └── 📂 points/
        ├── 📂 data/
        │   ├── 🟦 000000.bin
        │   └── ...
        └── 📄 timestamps.txt

The loader handles dataset-specific column indices and coordinate transforms once --data-type is set (kitti, mulran, diter_os, …). See examples/dataset_layout.md for custom formats.

🏋️ Training

bash scripts/train.sh

Hyperparameters live as env vars at the top of scripts/train.sh:

Variable	Notes
DATA_DIR	Root directory of your dataset
DATA_TYPE	diter_os | kitti | mulran | ...
TRAIN_SEQS / VALID_SEQS	Sequence names under DATA_DIR
LO_MODEL	kiss_icp | fast_gicp | small_gicp
GMM_COMP_NUM	0 = auto-pick K via BIC, otherwise fixed K
USE_GT	GT pose supervision (ablation only, no ICP/PGO)
USE_SUBMAP	Aggregate scans into a sub-map for small_gicp
TRAIN_RATIO	Fraction of training windows used

Ablations

• GT supervision. Lets you isolate the GMM reweighting contribution from the LO pseudo-label.

  USE_GT=true bash scripts/train.sh

• Raw-IMU + PVGO. Reproduces the Baseline* reported in the paper.

  bash scripts/raw_pvgo.sh

📊 Evaluation

Evaluate the best checkpoint:

CKPT=results/.../best_model.ckpt \
  EVAL_SEQS="Forest_new Lawn_lower_night Park_in_day" \
  bash scripts/evaluate.sh

Reports per-window endpoint RPE (translation, rotation) and end-point APE, averaged per sequence. Run bash scripts/evaluate.sh --help for all options.

🛰️ Inference

Save full ICP / PGO trajectories as .npz + top-down .png for plotting or downstream stages:

CKPT=results/.../best_model.ckpt \
  SEQS="Forest_new Park_in_day" \
  bash scripts/inference.sh

Variable	Notes
CKPT	Path to best_model.ckpt (required)
SEQS	Sequence names to run inference on (required)
LO_MODEL	kiss_icp | fast_gicp | small_gicp
USE_ADAPTIVE_WEIGHT	Weight ICP factors by overlap, IMU by integrated cov
USE_SUBMAP	Aggregate scans into a sub-map for small_gicp

Outputs land in <ckpt-dir>/inference/<seq>/{inference.npz, trajectory.png}.

🔬 Encoder t-SNE

Project encoder features to 2D and color by component id to verify motion-regime separation:

CKPT=results/.../best_model.ckpt \
  TRAIN_SEQS="Forest_new" EVAL_SEQS="Park_in_day" \
  bash scripts/tsne_encoder.sh

Variable	Notes
CKPT	Path to best_model.ckpt (required)
GMM	Override the fitted GMM (defaults to <ckpt-dir>/gmm.joblib)
TRAIN_SEQS / EVAL_SEQS	One t-SNE plot per split
PAIR_MODE	all plots every component. farthest keeps only the two components whose GMM means are most separated (focused separability check).
PER_COMP	Points per component (0 = auto, capped by MAX_WINDOWS)
MAX_WINDOWS	Total cap on plotted points
MIN_PER_COMP	Drop components with fewer windows than this
PERPLEXITY	Auto-clamped down if needed

Outputs: <ckpt-dir>/tsne/{tsne_train.png, tsne_eval.png}. The fitted GMM is saved alongside best_model.ckpt during training, so the default path usually just works.

🔗 Resources

• 📄 arXiv
• 🌐 Project page
• 🎬 Video

📝 Citation

@inproceedings{choi2026kissimu,
  title     = {KISS-IMU: Self-supervised Inertial Odometry with
               Motion-balanced Learning and Uncertainty-aware Inference},
  author    = {Choi, Jiwon and Kim, Hogyun and Yang, Geonmo and Lee, Juhui and Cho, Younggun},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026}
}

📬 Contact

Jiwon Choi: jiwon2@inha.edu

📜 License

BSD 3.0 for academic use. For commercial use, please contact the authors.

✨ Contributors