spatial-soft/spatial_pastry005_data_quality_check_20260625.md

Spatial Pastry005 Data Quality Check

Dataset:

/vepfs-C区/visuotactile/Tabero/formal_collections/spatial_soft_pastry005_10tasks_50each_gripperjitter_replay_format_20260624

Summary

The dataset is structurally clean and usable for training/evaluation.

Main caveat: force fields are not clean physical force labels. Most force values are proxy force generated from gripper closing; task5 contains a mixture of contact-sensor force and proxy force.

Counts

tasks: 10
hdf5 files: 10
demos: 500
success demos: 500
failed demos: 0
total frames: 49,423
frame range: 81 - 120
mean frames/demo: 98.846

Each task has:

1 replay-format HDF5
50 demos
100 camera videos
100 tactile videos

HDF5 Integrity

Passed:

missing required fields: none
NaN fields: none
Inf fields: none
shape issues: none
source_hdf5 missing: 0
duplicate demo ids: none

All required training fields are present:

actions
obs/actions
obs/eef_pose
obs/arm_joint_pos
obs/gripper_pos
obs/gripper_width
obs/gripper_close_norm
obs/gripper_marker_motion
obs/gripper_net_force
obs/fem_bbox_dims
obs/fem_deformation_max
obs/fem_deformation_rms
soft_extras/*
states/*
initial_state/*

Action-State Alignment

Checks:

actions vs soft_extras/actions_13d max abs diff: 0.0
obs/gripper_net_force vs soft_extras/gripper_net_force max abs diff: 0.0
action xyz vs obs/eef_pose xyz max abs diff: 0.0
action rotation vs obs/eef_pose quaternion converted to axis-angle: max diff 4.6e-7

So the axis-angle branch issue is not present in this dataset.

Note: obs/eef_pose stores quaternion rotation, while /actions stores axis-angle rotation.

Videos

Per task:

camera videos: 100 = 50 demos x 2 views
tactile videos: 100 = 50 demos x left/right marker videos

Formal video tree does not contain active preview_2x2 videos. Old previews were moved under:

_preview_2x2_moved_20260624_165515

Gripper

Recorded for all 500 demos:

obs/gripper_close_norm
obs/gripper_width
obs/gripper_pos
soft_extras/gripper_close_norm
soft_extras/gripper_width

Metadata gripper close range:

0.4601 - 0.4990

Observation gripper close range:

0.4601 - 1.0

The 1.0 values are open-gripper phases.

FEM / Safety

Peak values:

fem_deformation_max peak: 0.5383656025
fem_deformation_rms peak: 0.5142209530

With current default safety thresholds:

fem_max <= 0.60
fem_rms <= 0.55

all 500 demos pass the safety check.

Force Fields

Force source counts:

proxy force, source=1: 46,573 frames
contact sensor, source=0: 2,850 frames

Task-level force source:

task0-4: all proxy
task5: contact sensor + proxy mixed
task6-9: all proxy

Nonzero frames:

raw force nonzero frames: 2,850
processed force nonzero frames: 27,105

Important:

obs/gripper_net_force == soft_extras/gripper_net_force

These are processed force fields. They are not clean physical-force supervision labels.

Issues / Caveats

Hard data corruption:

none found

Training caveat:

force labels are mixed proxy/contact.

Recommendation:

For clean training, prefer 7D no-force tactile baseline, or regenerate/normalize force labels before using a 13D force-supervised baseline.

Per-Task Peaks

Task	Frames Mean	FEM Max Peak	FEM RMS Peak	Force Source
task0	86.98	0.2001	0.1910	proxy only
task1	108.84	0.3226	0.3119	proxy only
task2	81.68	0.2643	0.2503	proxy only
task3	89.84	0.2254	0.2044	proxy only
task4	102.52	0.3230	0.3046	proxy only
task5	95.58	0.3355	0.2936	contact + proxy
task6	100.00	0.2897	0.2713	proxy only
task7	101.68	0.4719	0.4643	proxy only
task8	102.54	0.2374	0.2250	proxy only
task9	118.80	0.5384	0.5142	proxy only