GINIO v1.0-corl2026 — code and checkpoints (CoRL 2026)

GINIO v1.0-corl2026 — code and checkpoint release for CoRL 2026

Paper: GINIO: A Geometric SO(3)-Equivariant Interface for Neural Inertial Odometry (CoRL 2026, poster) — OpenReview · arXiv 2609.25338.

Assets

asset	contents	size
ginio_checkpoints_tlio.tar.gz	4 TLIO checkpoint folders (tlio_*: GINIO-W, EqNIO, ResNet+SCEKF+SO(3) aug., ResNet-S+InEKF)	241.2 MB
ginio_checkpoints_ridi.tar.gz	3 RIDI checkpoint folders (ridi_*: GINIO-W, SO(3)-augmented EqNIO, ResNet+SCEKF+SO(3) aug.)	155.8 MB
ginio_checkpoints_nanobench.tar.gz	9 NanoBench checkpoint folders (nanobench_*: all nine Table 2 rows)	233.3 MB
paper_results.tar.gz	709 stored result files behind the "reproduced" values (per-sequence outputs, aggregates, recomputed filter metrics), for all 32 documented models	18.4 MB
SHA256SUMS	digests of the four archives	—

bash scripts/download_checkpoints.sh [tlio|ridi|nanobench|results] downloads and verifies them into checkpoints/.
Each checkpoint archive also carries INDEX.json (per-file SHA-256, size, parameter counts, epoch, stored training arguments), the SHA256SUMS of the released folders' files and RENAMES.txt (archive directory -> release folder).

What each checkpoint reproduces

See checkpoints/MANIFEST.md: one table per benchmark with the paper row, registry key / loader note, parameter count, paper value vs reproduced value, status (PINNED / CKPT_ONLY / RETRAIN / NOT_PAPER_MODEL / UNSUPPORTED) and 18 traps to read before loading anything. Paper-vs-evidence discrepancies are listed in docs/KNOWN_ISSUES.md; the cell-by-cell map is docs/PAPER_ARTIFACT_MAP.md.

Documented in checkpoints/MANIFEST.md but not in this release (status NOT_IN_V1.0; available on request): the Fig. 5(b) frozen-mean covariance arms, the Table 8 covariance-ablation models other than the Table 2 GINIO-RS, the Table 9 SO(3)-augmented NanoBench baselines, the 2026 ResNet+SCEKF retrain and eight non-paper controls.

Not released because nothing survives: the Fetch remounting models (Tables 10–12), AquaticVision (Table 13), GINIO-S on TLIO/RIDI, the RIDI ResNet+SCEKF model, and the generator of the Table 2 ID/SO(3) column for the non-equivariant rows.

Code

Three instantiations in one repository: src/ (Lie-Neuron family: GINIO-W/S/RD/RS/R, SCEKF and InEKF trackers), airio/ (Vector-Neuron AirIO family: GINIO-A), eqnio/ (canonicalised family: GINIO-E). scripts/paper/ regenerates the paper's NanoBench tables and figures. Datasets are not redistributed; the exact split lists are in splits/.