SSIO / README.md

Add teaser video and figures to card

48b27e8 verified 7 days ago

8.02 kB

license: other
library_name: pytorch
pipeline_tag: robotics
tags:
  - robotics
  - imu
  - inertial-odometry
  - self-supervised-learning
  - legged-robots
  - humanoid
  - quadruped
  - drone
datasets:
  - Open-IO/humanoid-g1-odometry
  - Open-IO/mipo-go1-odometry

SSIO: Self-Supervised Inertial Odometry — Checkpoints

Paper | Project Page | Code | Collection

Every model in the tables of SSIO, which trains inertial odometry from a robot's own unlabeled IMU stream, with no velocity or position labels. SSIO trains in two steps: Integration Self-Supervision, where the network learns body-frame velocity from an IMU integration teacher on short windows ("Ours"), and optional Pose Graph Distillation, which fine-tunes it on sequence-consistent velocities from a pose graph ("Ours (w/ PGD)"). The supervised AirIO baselines are released alongside.

Checkpoints

One directory per table cell, <dataset>/<recipe>/, named after the recipe in Config/experiment/<dataset>/<recipe>.yaml that retrains it (airio_full is the airio recipe with network=airio). Scores are the released checkpoint's AVE (m/s) / ATE (m) on the test split. AirIO and AirIO† are supervised; the other columns are self-supervised. Proprio adds joint states (and, on MIPO, the four calf-mounted IMUs) to the network input.

Dataset	Test split	AirIO	AirIO†	Ours	Ours (w/ PGD)	AirIO† + Proprio	Ours + Proprio
TLIO	`TLIO_Test`	`tlio/airio_full` 0.081 / 1.89	`tlio/airio` 0.084 / 2.08	`tlio/ssio` 0.104 / 3.04	`tlio/ssio_pgd` 0.101 / 2.99	–	–
EuRoC	`EuRoC_Test`	`euroc/airio_full` 0.335 / 3.72	`euroc/airio` 0.344 / 3.72	`euroc/ssio` 0.423 / 9.03	`euroc/ssio_pgd` 0.403 / 9.93	–	–
Blackbird	`BlackBird_Test`	`blackbird/airio_full` 0.318 / 6.10	`blackbird/airio` 0.362 / 6.39	`blackbird/ssio` 0.475 / 18.39	`blackbird/ssio_pgd` 0.429 / 17.73	–	–
MIPO (Go1)	`MIPO_Test`	`mipo/airio_full` 0.120 / 0.31	`mipo/airio` 0.121 / 0.31	`mipo/ssio` 0.182 / 1.63	`mipo/ssio_pgd` 0.169 / 1.58	`mipo/airio_proprio` 0.116 / 0.19	`mipo/ssio_proprio` 0.146 / 0.82
GrandTour (ANYmal)	`ANYmal_Test`	`grandtour/airio_full` 0.068 / 3.19	`grandtour/airio` 0.070 / 3.38	`grandtour/ssio` 0.096 / 4.37	`grandtour/ssio_pgd` 0.096 / 4.37	`grandtour/airio_proprio` 0.073 / 3.62	`grandtour/ssio_proprio` 0.091 / 3.66
Humanoid (G1)	`HumanoidOdom_Test` (Proprio: `HumanoidOdom_Joints_Test`)	`humanoid/airio_full` 0.087 / 1.26	`humanoid/airio` 0.096 / 1.40	`humanoid/ssio` 0.167 / 4.36	`humanoid/ssio_pgd` 0.155 / 4.38	`humanoid/airio_proprio` 0.095 / 2.35	`humanoid/ssio_proprio` 0.161 / 3.20

The released "Ours (w/ PGD)" checkpoints do not match the paper's PGD row exactly; Doc/recipes.md lists the paper and checkpoint scores side by side, with the recipe that retrains each run.

Repository Layout

<dataset>/                 tlio, euroc, blackbird, mipo, grandtour, humanoid
├── airio_full/            AirIO              (train_ssio.py, supervised)
├── airio/                 AirIO†             (train_ssio.py, supervised)
├── ssio/                  Ours               (train_ssio.py)
├── ssio_pgd/              Ours (w/ PGD)      (train_pgd.py, warm-started from ssio/)
├── airio_proprio/         AirIO† + Proprio   (mipo, grandtour, humanoid)
└── ssio_proprio/          Ours + Proprio     (mipo, grandtour, humanoid)
calibration/grand_tour/anymal_imu_gt_offsets_per_imu.json

Each directory holds best_model.pt, a bare PyTorch state_dict selected on the lowest validation loss (no labels), and the training run's Hydra config in .hydra/. The GrandTour calibration file holds the per-IMU IMU-to-ground-truth time offsets the GrandTour runs were trained with; it is not part of the GrandTour release.

Quick Usage

With the SSIO code installed and the datasets under $DATA_ROOT:

hf download Open-IO/SSIO --include "euroc/ssio/*" --local-dir ckpts
python evaluate.py ckpts/euroc/ssio --checkpoint best_model.pt --data Config/dataset/EuRoC_Test.yaml

Always pass --checkpoint best_model.pt and --data: the released runs ship no checkpoint_*.pt and store their splits under val_dataset / test_dataset. To fine-tune a released "Ours" run with Pose Graph Distillation:

python train_pgd.py +experiment=euroc/ssio_pgd train.resume_from=file:ckpts/euroc/ssio/best_model.pt

The evaluator reports AVE (average velocity error, m/s, the primary metric), ATE (RMSE of the position error, m), Δv error, and RVTE / nRVTE.

Citation

@misc{qiu2026ssioselfsupervisedinertialodometry,
      title={SSIO: Self-Supervised Inertial Odometry},
      author={Yuheng Qiu and Yutian Chen and Carlton Cort and Weison Huang and Shibo Zhao and Pieter Abbeel and Rocky Duan and Koushil Sreenath and C. Karen Liu and Carmelo Sferrazza and Junyi Geng and Sebastian Scherer and Guanya Shi},
      year={2026},
      eprint={2610.25988},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2610.25988},
}