mskweon/FWAM-Omni

FWAM Omni

Release status: uploading. This repository is incomplete until metadata/release.json records completion.

This release contains the exact processed episode populations used for the current FWAM joint training and fixed-test evaluation, plus the external RGB-D files needed by the reader, the Bench2Dex force correction, normalization statistics, cached text embeddings, frozen splits, and evaluation populations. It is a data snapshot, not a claim of model convergence or benchmark superiority.

Dataset	Train episodes	Validation episodes	Test episodes	Additional test
GRAIL v3	4,808	593	651	—
ManiSkill dynamics v2c	1,395	139	242	110 dynamics OOD
HRDexDB v1	967	96	69	—
DexTaG v2	191	25	25	—
Bench2Dex v4 ready	840	105	105	—
VPM Real-Robot v2	—	—	100	Evaluation only

The 8,201 training episodes define 1,071,969 candidate windows. The fixed 1 Hz test evaluation uses 13,788 windows; these counts describe different populations. They do not mean that a particular training run visited every window.

Contents

• data/fwam/<dataset>/*.npz: original processed episode bytes, including valid masks, action/state, body and scene trajectories, sensor values, and available RGB-D arrays.
• data/fwam/bench2dex_forcefix_v1/: mandatory represented-link force corrections for affected Bench2Dex episodes. The old force arrays must not be used without this overlay.
• metadata/splits/: portable repository-relative split lists. provenance/frozen_splits/ preserves the original lists and their identity.
• provenance/test_1hz/: original fixed evaluation-window definitions.
• data/fwam/stats_omni_v3_q.json: training normalization.
• data/fwam/text_cache_qwen3vlemb2b_v8.pt: the text embedding cache used by the pipeline.
• External HRDexDB FRONT/SIDE RGB-D packets and VPM depth dependencies are being added with their original resolution and bytes. The processed HR episodes alone do not contain images; the processed VPM episodes alone do not contain depth.

No new image resize or lossy re-encoding is performed for this release. Existing model-sized images inside the processed episodes remain the historical conversion output. Their resolution is distinct from the native external packet resolution.

Evidence and units

GRAIL, ManiSkill, DexTaG, and Bench2Dex use simulator-derived trajectories. A pointwise force field obtained by distributing contact forces onto sampled robot points is a derived label, not a measured dense tactile image. ManiSkill force supervision is the target-object normal resultant on each finger. DexTaG excludes self-contact from its environment-force target. Represented-link support and unknown missing-link loads remain explicit.

HRDexDB scene targets are FRONT TrackEverything pseudo-GT. SIDE is input only. Native tactile scalars are normalized, unitless values and must not be treated as Newton force labels. HRDexDB validation is a held-out object set, not a model-selection set for this project.

VPM depth is estimated by Depth Anything 3 and its scene trajectories are pseudo-GT. All 100 VPM episodes are test-only. Do not create training or validation splits from them.

Failed rollouts remain data. Clean-window filters, including DexTaG's relaxed penetration and hammer-load thresholds, are documented in the producer notes. Cross-dataset CAD overlap exists; this release does not imply globally unseen-object evaluation. Mass, inertia, and friction metadata are not model inputs or prediction targets in the FWAM protocol.

Sources

Producer documentation is preserved under provenance/. This repository assembles derivatives of GRAIL-Phys, ManiSkill, HRDexDB, DexTaG, Bench2Dex, and VPM/VP-WorldBench. Source provenance and recorded upstream terms remain attached; no single blanket relicensing of all upstream components is asserted.

The project owner authorized public redistribution of this prepared collection. See the per-source notes for attribution, acquisition provenance, known limitations, and original source references.