- Contents (4 dataset configs)
- Schema (per-episode
.npz) - Sampling law (per episode, independent, seeded)
- Physics (identical across the whole corpus)
- Validity taxonomy (
validity,valid_training) — maintainer ruling 2026-07-20 - Admission gate (how the expansion object set was chosen) — see
admission_matrix.md - Recommended split —
held_out(metadata inmanifest.json) - Known boundaries
- Verify
CDWM Drop Corpus — v1.0 (spec_version = labeling_spec_v2)
A rigid-body drop / placement dataset for training a world model of where an object comes to rest after being
released above a table. Object released above a table → free-fall → impact → settle. Sibling to the CDWM grasp
world model (BWangCN/cdwm-grasp-dataset); conditioning uses the same object Gaussian-cluster assets.
Contents (4 dataset configs)
| config | episodes | role |
|---|---|---|
train_corpus/ |
160,200 (16,200 corpus-v1 core + 144,000 WM#1-universe expansion) | training; ALL validity classes present with flags intact |
v0_grid_eval/ |
990 | held-out-configuration eval — a structured table grid (fixed heights × tilts × reps), same objects, unseen exact layouts |
tier_b/ |
30 | finger-open gripper-release episodes; validates the zero-velocity kinematic-release model |
spread_eval_v2/ |
180 | multimodality calibration — 20 deployment-jittered reps at 9 near-boundary release nominals |
train_corpus = ~80 meshes / 108 (object×config) configs, 153,996 SETTLED (valid training). The 6 corpus-v1
configs (box, potted-can, tomato-can upright+lying, marker, apple) are the signed-off core; the expansion covers 75
newly-admitted objects from the ~171-object WM#1 universe. Bowl / receiver (containment) episodes are excluded
(frozen for a follow-up paper).
Schema (per-episode .npz)
object · config_id/config_kind · drop_h · tilt_deg · yaw_rad · xy · roll_mu · seed ·
release_quat (wxyz) · obj_pos (N,3) · obj_quat (N,4, wxyz) · t (N,) — the settling trajectory, world
frame, logged at 31.25 Hz · settle_t · status · validity · valid_training · outcome_label /
outcome_subclass · release_basin / rest_basin / basin_transition · resting_pen_mm · net_rot_from_release_deg.
Storage: tar-sharded (<config>/<prefix>_NNN.tar, each holding ~5000 .npz members) — HF rate-limits 161k
per-file uploads. Use load_example.py (iter_episodes(root, config)) to stream episodes.
Per-config schemas: train_corpus carries the full spec-v2 schema above (validity + basins). v0_grid_eval is
the raw structured-grid schema (obj_pos/obj_quat/t/settle_t/status + v1 label/subclass; re-derive v2 validity
from the trajectory as verify_dataset.py does). tier_b uses a finger_open_release schema (object, probe,
release_h, v_release_lin_cmps, v_release_ang_dps, release_frame, obj_pos/obj_quat/t). verify_dataset.py
checks the train_corpus spec-v2 labels.
Sampling law (per episode, independent, seeded)
- drop height
h~ U[5, 80] mm — defined as the clearance between the object's LOWEST point at the released orientation and the table surface (the only size-invariant definition across meshes of very different scale). - tilt ~ U[0, 15°] about a uniformly-random horizontal axis · yaw ~ U[0, 2π) · release XY ~ uniform disk r = 5 cm.
- orientation = candidate stable pose ⊕ tilt ⊕ yaw.
Physics (identical across the whole corpus)
MuJoCo 3.9, condim=6 + rolling friction μ = 1e-4 (single value), harness-default contacts, zero-velocity
kinematic release, NO gripper, 4 s horizon, 31.25 Hz logging. Object↔table contact is governed by the table geom
(priority=1) → identical regardless of object. Collision = per-object CoACD decomposition (obj2mjcf, ≤24 hulls)
from the WM#1 physics asset set (assets573). Masses: measured for the 9 YCB objects; all others a 400 kg/m³ ×
convex-hull-volume proxy (nearly irrelevant to rigid settling; see Known boundaries).
Validity taxonomy (validity, valid_training) — maintainer ruling 2026-07-20
Validity = settling. Any episode that reaches converged rest within the 4 s observation window is valid training data — including tip-overs (basin transitions: "it fell over, but that is another graspable stable pose") and roll-then-settle. Episodes still moving at 4 s are NOT valid (recorded + flagged, never deleted). The at-rest criterion is unchanged (<6 mm/s ∧ <6 °/s sustained 150 ms); 4 s is the observation ceiling, not a fixed-duration truncation. Outcome labels are descriptive strata for balancing/eval — they do not gate trainability.
SETTLED— converged rest within 4 s (valid_training = True). Includes TIPPED / ROLLED outcomes.NO_REST— still moving at 4 s (flagged, never deleted).ARTIFACT— resting penetration > 2 mm or launch/NaN (peak transient penetration is explicitly not a trigger).SETTLED_OFF_SUPPORT— rest with COM off the table (none on this table set).
Outcome strata (descriptive only): outcome_label ∈ {SUCCESS, FAILURE} with outcome_subclass ∈ {TIPPED (>30°),
ROLLED (>25 mm), NEVER_SETTLED, …}. Basins: release_basin/rest_basin = nearest stable-pose (yaw-modded);
basin_transition = they differ. UNASSIGNED for 3 degenerate/near-sphere meshes.
Admission gate (how the expansion object set was chosen) — see admission_matrix.md
Each (object, stable-pose) faced 4 calibration drops {2, 8} cm × {0°, 15°}. ADMIT iff all 4: (i) SETTLED on table
(roll-off = fail; tip into a different basin = fine); (ii) [2 cm] ≤2 bounces & rebound < 0.3 h; (iii) no NaN/launch;
(iv) resting-pen < 2 mm. 102 ADMIT / 181 REJECT across 283 candidate configs (75 of 170 NEW objects admitted).
All rejects failed (i) — the mechanized roll-prone / ball-class exclusion. Per-object verdicts in admission_matrix.md.
Recommended split — held_out (metadata in manifest.json)
12 held-out meshes (cross-object generalization eval), stratified by size, incl. 013_apple +
010_potted_meat_can. Hard rule: all configs of one mesh on the same side. Held-out objects are still present in
the files; use the manifest.json meshes[*].held_out flag to split. (dt_train, H) is left to the loader — raw
31.25 Hz is stored; candidates (192 ms, 16) vs (96 ms, 32).
Known boundaries
- Mass proxy — non-YCB masses are a 400 kg/m³ proxy; this drives an elevated resting-penetration tail (expansion median 0.87 mm vs the 6 hand-picked corpus-v1 objects' 0.068 mm; 1.8 % of expansion episodes > 2 mm → ARTIFACT).
- Near-boundary configs — 3 configs have NO_REST > 20 % (roll-prone bottles/mouse); 8 configs settle-time p95 > 3.2 s (near the 4 s ceiling). Their NO_REST episodes are data, not failure.
- Ball-class excluded by the gate (ideal-sphere / roll-dominant objects never settle).
- tier_b: median finger-open release velocity 0.15 cm/s ≪ 6 cm/s gate (validates zero-velocity release); 1/30 episodes reached 7.79 cm/s (tail).
- Reproducibility: bit-exact re-simulation from the stored seed/params (sub-mm; occasional boundary-settle sensitivity).
Verify
python verify_dataset.py <root> --n 500 — re-derives the spec-v2 labels from the stored trajectories on a random
sample and asserts equality with the stored labels + schema.
- Downloads last month
- 33