Dataset Viewer
Auto-converted to Parquet Duplicate
ulid
stringlengths
18
18
task_hand
stringclasses
2 values
object_phrase
stringclasses
4 values
obj_px_frame0
int64
5.28k
18.6k
grasp_frame
int64
110
288
release_frame
int64
156
434
n_frames_scored
int64
132
210
box_trim_pct
float64
1
1
occlusion_pos
float64
0
1
occlusion_neg
float64
0
0
01JMXF99BYFPWU5NOL
R
bagel
7,451
194
322
149
1
0.287478
0
01JMXGMVZPODEGWHV5
R
croissant
11,093
116
250
134
1
0.425854
0
01JMXGPPFC2ZRTLNWH
R
croissant
9,989
140
268
132
1
0.188407
0
01JMXGQEAGFX55VOQT
L
croissant
15,964
122
258
151
1
0.558256
0
01JMXGR8GTQVLVREEX
L
croissant
8,369
212
390
189
1
0.481659
0
01JMXGS9NVVF5XPUDQ
L
croissant
16,542
148
322
157
1
0.615585
0
01JMXGT691NQAI4UGX
L
croissant
10,002
144
304
154
1
0.436313
0
01JMXGW40SVV2DKJVY
R
croissant
17,943
176
340
153
1
0.392521
0
01JMXGX388EP67QCGO
R
croissant
8,710
162
296
138
1
0.445121
0
01JMXGXY7MKHTF37DJ
R
croissant
10,265
162
284
134
1
0.524208
0
01JMXGYNQNIYVIKPAK
R
croissant
5,284
160
310
145
1
0.108062
0
01JMXGZNHTYMUPD77N
R
croissant
9,824
204
330
151
1
0.138029
0
01JMXH0JZ2UM6QEJFZ
L
croissant
15,905
158
302
150
1
0.745677
0
01JMXH39SZD6SNVGKZ
L
croissant
12,778
132
262
140
1
0.577399
0
01JMXH44Z0CAPNWQGL
L
croissant
16,955
124
262
136
1
0.429372
0
01JMXH50JR4AAJWETN
L
croissant
17,826
140
268
134
1
0.496578
0
01JMXH6MCHZJLL46NL
L
croissant
10,505
196
328
157
1
0.245597
0
01JMXH7M1DAFQ5ZAWF
R
croissant
5,548
184
330
152
1
0.381038
0
01JMXH8F3YCAGKVL6D
R
croissant
16,804
154
316
149
1
0.46239
0
01JMXH9DK2YOAL5T5B
L
croissant
11,822
130
266
134
1
0.925309
0
01JMXHB4XXYG2ZBSR4
L
croissant
9,261
188
344
155
1
0.677573
0
01JMXHC55J2L4AQXR4
L
croissant
16,941
130
280
139
1
0.529957
0
01JMXHD0RABKJ5RDEL
R
croissant
7,568
158
280
132
1
0
0
01JMXHDWD4JHEF3RMU
R
croissant
16,059
114
270
137
1
0.457999
0
01JMXHFJNGEK7ODZIQ
R
croissant
15,343
122
270
143
1
0.463729
0
01JMXHGFJ8KRVXBW6O
L
croissant
12,685
148
276
139
1
0.402759
0
01JMXHJAJFQFQODPYS
L
croissant
8,615
178
342
168
1
0.570517
0
01JMXHK95EE35IFEXL
L
croissant
17,515
154
308
172
1
0.631858
0
01JMXHM95FBDOCPX2L
L
croissant
17,735
138
264
138
1
0.466761
0
01JMXHP1M4VC5AVYM6
L
croissant
10,582
168
296
145
1
0.553109
0
01JMXHPXJZG46CDQM5
R
croissant
15,119
126
246
134
1
0.13784
0
01JMXHRMJMMEN2ZKO5
R
croissant
8,501
186
312
145
1
0.199388
0
01JMXHSH5GGRR4ZWBM
R
croissant
17,725
138
276
134
1
0.433343
0
01JMXHT9J86FA7GEJB
L
croissant
13,765
126
284
150
1
0.570795
0
01JMXHV4WQ45DIDTZG
L
croissant
9,365
194
332
168
1
0.233422
0
01JMXHW9V3XC6SBX7R
L
croissant
18,464
134
262
151
1
0.511319
0
01JMXHX7PMRRVTZI73
L
croissant
11,654
186
326
155
1
0.331989
0
01JMXHY52YOGZKZEJB
L
croissant
10,484
190
340
158
1
0.166349
0
01JMXHZ5XDADBGASM5
R
croissant
9,082
200
338
153
1
0
0
01JMXJ04M1ELEJ73K5
R
croissant
16,950
136
274
135
1
0.484897
0
01JMXJ0W2GY7FRZJ5Z
L
croissant
12,517
138
312
156
1
0.489574
0
01JMXJ1QE1JENE3VYV
L
croissant
15,037
132
274
158
1
0.265744
0
01JN0A85HBTIASV5LM
R
croissant
10,053
150
356
195
1
0.451706
0
01JN0A91NCBEM2TXLL
R
croissant
10,484
122
374
200
1
0.394315
0
01JN0A9WPETY4J25V5
R
croissant
10,208
178
434
198
1
0.330525
0
01JN0AAQZZJ4W2RMJS
R
croissant
9,836
172
422
210
1
0.324319
0
01JN0ABKHEIDL5PFHJ
R
croissant
9,997
140
386
192
1
0.151746
0
01JN0ACD62XX5SQMDB
R
croissant
10,332
132
402
187
1
0.335172
0
01JN0B1N7DIF2OCR2Z
R
croissant
12,245
178
326
159
1
0.512781
0
01JN0B2BNZFXLIMPNY
R
croissant
12,081
162
328
165
1
0.479348
0
01JN0B35F9DZ6T4NDH
R
croissant
13,771
140
310
165
1
0.486021
0
01JN0B3VETIMF4BRK7
R
croissant
11,050
136
284
153
1
0.510317
0
01JN0B4HRRYY52AGBE
R
croissant
11,591
142
310
159
1
0.510569
0
01JN0B59XC5PAHKPR2
R
croissant
12,241
156
328
162
1
0.501348
0
01JN0B6144I3JRR245
R
croissant
13,052
154
326
172
1
0.472724
0
01JN0B6SM7ETEVDWLM
R
croissant
10,663
152
320
161
1
0.473038
0
01JN0B7MWN7VS3RR7Q
R
croissant
12,153
134
294
153
1
0.506459
0
01JN0B8G8VT6WWBBEV
R
croissant
12,142
158
334
171
1
0.408005
0
01JN0B981DR4HI32AT
R
croissant
13,091
142
292
141
1
0.488198
0
01JN0B9XZASBEFU6PQ
R
croissant
12,118
144
296
151
1
0.523436
0
01JN0BAMK9C6VGFKV2
R
croissant
12,765
156
298
160
1
0.559812
0
01JN0BBBYQSYO4JUYW
R
croissant
11,866
140
304
160
1
0.502107
0
01JN0BC2Z6ZAXRWIPI
R
croissant
11,620
140
316
160
1
0.486833
0
01JN0BCTN2BA76U4D3
R
croissant
12,227
148
310
160
1
0.546495
0
01JN0BDHGV7EESL4NM
R
croissant
12,196
126
284
148
1
0.404969
0
01JN0BE8GCDGAXDSCI
R
croissant
11,733
146
290
154
1
0.418648
0
01JN0BEYVX3GS4FX5B
R
croissant
11,064
146
274
140
1
0.499367
0
01JN0BFMVDSCEZRCOC
R
croissant
12,609
116
264
137
1
0.35086
0
01JN0BG9G1WQIU236S
R
croissant
12,216
110
270
150
1
0.26981
0
01JN0KMMEED4B6XL2R
L
donut
7,864
182
306
150
1
0.999746
0
01JN0KNC6ENBIIANCP
R
bagel
8,764
142
270
149
1
0.306481
0
01JN0XAGCRDDJ33ML4
R
croissant
9,820
208
366
161
1
0.21334
0
01JN0XB99NMRGSN34B
L
croissant
10,165
224
390
171
1
0.451648
0
01JN0XC8EEQBMRX73C
R
croissant
9,764
202
332
156
1
0.225113
0
01JN0XDNHPW3MSHRCZ
R
bread
13,179
150
242
133
1
0.272934
0
01JN0XEAG7A6NDMS7R
R
croissant
12,010
144
252
138
1
0.318235
0
01JN0XF11CKS6YENQJ
L
croissant
13,196
148
156
153
1
0.271143
0
01JN0XFRTCSU3HUYBK
R
croissant
9,781
166
336
160
1
0.265924
0
01JN0XGJ9NIQQDTF3O
R
croissant
11,207
236
388
176
1
0.236727
0
01JN0XHCY4E6E6HE5Z
L
croissant
13,919
188
354
168
1
0.312307
0
01JN0XJ7Q8CH7OG5XN
L
bread
12,130
182
350
168
1
0.178318
0
01JN0XK2P5SFFTTIJU
R
croissant
11,282
164
312
165
1
0.380429
0
01JN0XKY46ORUF2SQ7
R
croissant
11,300
118
272
149
1
0.435398
0
01JN0XMMZRPUNRCL4Q
R
croissant
10,993
204
362
168
1
0.288911
0
01JN0XNG2226HCR27P
R
bread
12,082
170
344
164
1
0.306406
0
01JN0XP8ABNHWX4ZRP
L
croissant
15,058
202
368
179
1
0.234028
0
01JN0XQ2MFKJ3F6WCW
R
croissant
11,371
168
354
171
1
0.373934
0
01JN0XQYZ75FPQ4GIQ
L
croissant
17,496
288
428
187
1
0.64249
0
01JN0XSNAK5A6GJF76
L
croissant
18,606
204
344
154
1
0.499194
0
01JN0XTD94FJISJQBQ
R
croissant
15,572
146
306
152
1
0.125225
0
01JN0XV4GT5TNULAUJ
R
croissant
14,303
178
334
174
1
0.198769
0
01JN0XVX6376EU4HK6
L
croissant
16,274
188
194
201
1
0.468047
0
01JN0XWS9V2Z56BCFO
L
croissant
16,609
168
298
159
1
0.282136
0
01JN0XXHP8WDWQIHB3
L
croissant
10,045
180
352
159
1
0.540866
0
01JN0XYBN8F6CWOU7K
R
croissant
11,010
172
316
158
1
0.27257
0
01JN0XZ4VWVGAZWEOQ
R
croissant
11,242
202
378
173
1
0.371909
0
01JN0XZYE15R6DF3VW
L
croissant
13,653
158
338
161
1
0.467004
0
01JN0Y0R46ZBRGEFQI
L
croissant
9,904
204
404
170
1
0.446587
0
01JN0Y1K3SR7N5LYF5
R
croissant
11,010
190
328
163
1
0.234605
0
01JN0Y2AWRMLCMJWDP
R
croissant
11,556
154
314
152
1
0.244116
0

ActionNet subset100 with ground-truth depth

This is a 100-episode subset redistribution of a third-party dataset, plus our derived artifacts. Read the attribution before using it.

Attribution and license

Upstream dataset FourierIntelligence/ActionNet (Fourier Intelligence)
What is redistributed 100 episodes out of 30,121, byte-identical to the upstream tars: rgb.mp4 (1280x800 fisheye), depth.mkv (lossless 16-bit), timestamps.json, <ULID>.hdf5
Upstream terms Whatever the upstream repository states. The upstream terms govern the episode payload; if they disallow redistribution, tell us and this repo comes down.
Ours (CC-BY-4.0) Everything under derived/, the mapping tables, build_subset100.py, and this document
Robot Fourier GR1-T1, 6-DoF Fourier DexHands, single head-mounted camera (top)

Why this subset exists

The widely used LeRobot conversion (lerobot/action_net) drops depth and downscales RGB to 192x288, and its observation.state is indexed at the robot's 60 Hz while its rows are labelled at the camera's 30 Hz โ€” so image and state drift apart by up to half an episode. Both problems are documented, with measurements, below. This subset keeps the original 1280x800 RGB, the lossless ground-truth depth, and a corrected per-video-frame state built from the original hdf5.

What is corrected here (2026-08-20)

Measured
observation.robot_joints[k] vs hdf5 /state/robot[k] within 1e-5 on 97/100 episodes (median max abs diff 5.95e-08); the same tar-independent check passes 21/21 on a second tar
the documented [2k] rule 0/100, median max abs diff 0.740 rad; and in 18/21 episodes of the second tar 2k runs past the end of the hdf5
FK of our corrected states vs the dataset's own /state/pose left EEF median 0.29 cm, right 0.27 cm, head 0.59 cm, over 3 cm on 0/100
the same check with states indexed by video frame the moving arm sits median 7.59 cm off (range 2.96-18.66), 99/100 over 3 cm; the idle arm stays at 0.33 cm, which is why this hid for so long

Pre-Contact Level Filtering โ€” the inputs, the verdicts, and how to apply it

Added 2026-08-24. Everything above is about the capture: what was redistributed, what was corrected, and what the depth is. This part is about what we then did with it -- the filter that decides whether a video is consistent with the action stream that goes with it -- and it ships the inputs and verdicts so the filter can be re-run rather than taken on trust.

The top-level directories are new and are the authoritative ones. derived/ is the earlier partial staging and is kept for the artifacts that did not move: derived/points/, derived/hand_points2/, derived/hand_tracks3/, and derived/states/ (the un-tau-corrected states). Two of its folders were removed rather than left to be picked by accident -- derived/masks/ held only 43 of the 100, and derived/states_tau/ was byte-identical to the new states/ (all 100 checked by md5); that folder's evidence note moved to states/README.md.

F0. What the filter decides

Given a video and the action stream that goes with it, is the video consistent with those actions? The question exists because generated video is used as training data: a clip whose pixels show an arm going somewhere its actions never went teaches a policy the wrong thing.

The filter answers in two levels, and both must pass:

level question space what it can see
Pre-Contact did the acting fingertip ever reach the target object's boundary? 1280x800 pinhole one state, the action stream, the frame-0 object. Never the pixels.
Occlusion at that instant, does the arm the video draws actually cover the object? the video's own resolution the pixels

The split is deliberate. Pre-Contact is a geometric screen computed from actions alone, so it cannot be fooled by a rendering. Occlusion is the pixel check, and it is only asked at the instant Pre-Contact names. Neither is a contact test; together they are a screen.

This set is the reference, not the target. Every episode here is a real teleoperated demonstration, so the arm does reach the object. Running the filter here answers does it destroy good data? -- the answer is 98/100 -- and calibrates the constants. The set it exists to judge is generated video: see glory-hyeok/robocurate-synth100.


F1. What is in here

manifest.csv                    one row per episode: ulid, task_hand, object_phrase,
                                obj_px_frame0, grasp/release frame, occlusion_pos, occlusion_neg
camera.json                     K_pinhole, K_fisheye, D, and the space definitions

gtdepth/depth_<U>.npz           GROUND-TRUTH depth, already undistorted + downsampled:
                                  depth        (T, 400, 640) float16, metres
                                  frame_indices (T,)     which video frames these are
                                  K_ds         (3,3)     the intrinsics for THIS array
                                  ds           ()        the downsample factor (2)
                                  K_full       (3,3)     the 1280x800 pinhole K
                                  zero_frac, sat_frac    per-frame sensor quality
masks/object_masks_<U>.npz      target-object masks, (T, 800, 1280) bool + frame_indices,
                                  n_px, object_phrase, method, n_obj
states/states_<U>.npz           states (T,44) + actions (T,44) + frame_indices, tau-corrected
handmasks/<U>_{L,R}.npz         tracked arm masks (24 episodes -- the sampled subset)
world3d/world3d_<U>.json        the Pre-Contact verdict, per episode + per frame
occlusion/occlusion_full.json   the Occlusion verdict, all 100 in one table
occlusion/occlusion_series.json the per-frame score inside the window

Also in the repo, from an earlier partial staging: derived/points/ (VLM object points), derived/hand_points2/ and derived/hand_tracks3/ (10 episodes each), derived/states/ (the un-tau-corrected states), and subset100_gtdepth.tar. The top-level directories above are the authoritative ones -- derived/masks/ held only 43 of the 100 and derived/states_tau/ was a byte-identical copy of states/, so both were removed rather than left to be picked by accident.

gtdepth/ is multi-frame, but the filter only uses frame 0 -- the object is not moved before it is touched, so its boundary is built once and held (static_box: true, box_frame: 0 in every verdict). The rest of the frames ship because they are what a depth model gets scored against.


F2. The scene, and why it matters

Every episode has n_obj = 1: a single pastry on a plate, on a striped tablecloth. The masks are ours, from a SAM 3 text prompt (croissant x94, bread x3, bagel x2, donut x1), keeping the largest connected component inside an area band.

That single-object scene is doing quiet work, and it is the main reason numbers measured here are looser than what a cluttered scene needs. When a mask's border bleeds a pixel past the object, here it lands on the tablecloth immediately beside it -- at nearly the pastry's own depth. In a multi-object scene the same one-pixel error lands on a neighbouring object tens of centimetres away, and the lifted point cloud gains a second cluster that stretches the boundary. See ยงF6.


F3. How to apply Pre-Contact filtering

F3.1 What you need beyond this repo

robot model GR1T1_fourier_hand_6dof.urdf, from FFTAI/teleoperation @main:assets/urdf/, md5 78ec53140a125ef50c3a8701249fb036 (not redistributed here)
forward kinematics any URDF FK that returns the world pose of every link

The 44-channel state layout is ActionNet's own. Channels 0-6 are the left arm, 22-28 the right; the 6 hand slots per side are declared in configs/hand/fourier.yaml order (pinky, ring, middle, index, thumb_pitch, thumb_yaw) and are DexHand raw units in [0, 10.3], not radians, on this set.

F3.2 The trajectory

Roll the state forward from the episode's first state using the recorded actions, subsampled to 15 Hz (actions[::2]), with a first-order tracking model:

q[t+1] = q[t] + beta * (q[t] - q[t-1]) + alpha * (a[t] - q[t])
         beta = 0.0    alpha = 0.45 (body)   alpha = 0.60 (hand)

Then add two offset layers, in this order:

  1. rollout_offsets -- both arms and the waist/head chain, largest term 4.29 deg. These correct the controller's standing bias, which the rollout inherits.
  2. left_arm_offset -- the left arm only: shoulder pitch 0.40, roll -0.27, yaw 4.04, elbow pitch 7.21 deg. This is a left-arm encoder zero bias, not a symmetric fudge: under it, 9 of 10 fit episodes reach 0.26-1.12 px reprojection error, which is the level the right arm already reaches with no offset at all. Do not mirror it to the right arm. It moves the left fingertip by 6.11 cm, so getting this wrong on a left-handed episode is larger than the object.

Both files are in the repo root alongside this card.

F3.3 The object boundary

At frame 0 only:

z  = np.load(f"gtdepth/depth_{u}.npz")
D  = z["depth"][list(z["frame_indices"]).index(0)].astype(float)   # (400, 640) metres
Kd, ds = z["K_ds"], int(z["ds"])

m  = np.load(f"masks/object_masks_{u}.npz")
M  = m["masks"][list(m["frame_indices"]).index(0)][::ds, ::ds]     # match depth's grid

ok = M & (D > 0.05) & (D < 2.6)                 # sensor range gate
v, u_ = np.nonzero(ok); zc = D[v, u_]
lo, hi = np.percentile(zc, [5, 95])             # drop depth outliers
k = (zc >= lo) & (zc <= hi); v, u_, zc = v[k], u_[k], zc[k]

P_cam = np.stack([(u_ - Kd[0,2]) / Kd[0,0] * zc,
                  (v  - Kd[1,2]) / Kd[1,1] * zc, zc], 1)
P_base = (base_T_cam[:3,:3] @ P_cam.T).T + base_T_cam[:3,3]

lo3 = np.percentile(P_base, 1, axis=0)          # 1-99 per axis, NOT min/max
hi3 = np.percentile(P_base, 99, axis=0)
centre = (lo3 + hi3) / 2
radius = np.linalg.norm(hi3 - lo3) / 2          # circumscribed sphere of the AABB

base_T_cam = FK(q0)["head_pitch_link"] @ head_pitch_T_cam, with the extrinsic from align_camera_hullfit_20260820.npz in the repo root.

The boundary is the circumscribed sphere of the trimmed box, not the box. The sphere is 4.2x the volume, which is the point: Pre-Contact is a screen, and a screen should be generous. On this set the median radius is 8.15 cm, from an AABB of 13.1 x 7.0 x 7.1 cm -- genuinely a croissant.

F3.4 The distance, and the event

Per frame, over the five *_tip_link fingertips of the task hand:

d = max(0.0, min(norm(FK(q_t)[tip][:3,3] - centre) for tip in tips) - radius)

d == 0 means inside. Pre-Contact is an EVENT, not a state: the entry frame is the first frame at d == 0 having been outside on the frame before.

if entry is None and prev is not None and d <= 0 < prev:
    entry = t

Frame 0 therefore can never qualify -- there is no frame before it to have been outside on. That sounds like it should cost something, because 24 of 100 episodes already start inside the sphere: the demonstration often begins with the hand near the pastry. Measured, it costs nothing. All 24 also produce a crossing later -- the hand leaves and comes back, which is what picking a thing up and putting it down looks like -- so 100 of 100 episodes here have an entry event.

The same holds on the generated set: 21 of 300 clips start inside (7 per method, and exactly 7 because all three share the same 100 conditioning frames, so frame 0 does not depend on the generator), and all 21 get an entry event anyway. Nothing is lost to the frame-0 rule in either set. Clips with no entry are clips whose fingertip never reached the boundary at all, which is the signal the filter exists to find.

F3.5 Occlusion, at that instant only

Window: +-1 s around the entry frame, i.e. +-30 frames at this set's 30 fps.

obj = object_mask_at_frame_0                       # static: nothing moves it before contact
occ = max(|arm_mask[f] & obj| / |obj|  for f in window)
keep = occ >= tol

The denominator is the object, not the union. Measured on the 24 episodes with tracked arm masks, the arm mask is 3.2x the object at the median (range 0.5-10.1x), so on a clean grasp that fully covers the object, IoU reads about 0.31 where coverage reads 1.0 -- it understates by roughly the size ratio, and understates most in exactly the frames where the arm is closest. Coverage has no such bias, and it is also the quantity the question asks for: what fraction of the object is covered.

F3.6 The threshold, and why this set cannot set it

tol task hand passes idle hand falsely passes
0.1 % 98 / 100 0 / 100
1 % 98 / 100 0 / 100
5 % 98 / 100 0 / 100

The task hand's occlusion is 43.1 % at the median (10th percentile 19.8 %); the idle hand -- the negative control, the arm that never moves -- reaches 0.00 % at its maximum. Separation is total, so the threshold is unidentifiable here: three values two orders of magnitude apart give the same answer. It has to be set on generated video, where the scores are lower and the control is much harder (an arm that genuinely moves and grasps, just not the instructed object). Treat any tol in this range as provisional.


F4. What came out on this set

episodes 100
Pre-Contact entry found 76 (24 start inside the sphere at frame 0 -- no observable event)
Occlusion, task hand median 43.1 %, p10 19.8 %
Occlusion, idle hand (control) median 0.00 %, max 0.00 %
both levels pass 98 of the 100 that were scored
boundary sphere median 8.15 cm

98/100 is the answer to does the filter destroy good data -- and that is all a set of real demonstrations can tell you. It cannot tell you whether the filter catches bad data, because there is no bad data here. That requires generated video.


F5. Reproducing the verdicts

world3d/world3d_<U>.json records the settings each verdict was produced with, so a reimplementation can be checked against them rather than trusted:

{"ulid": "...", "task_hand": "R", "static_box": true, "box_frame": 0,
 "box_trim_pct": 1.0, "boundary": "sphere", "grasp_frame": 194,
 "grasp_frame_recorded": 191, "grasp_source": "rollout",
 "tip_to_box_at_grasp_cm": 0.0, "curve_min_cm": 0.0, "n_frames_scored": 149}

grasp_source: "rollout" is worth noting: the verdict is computed from the rolled-out trajectory, not the recorded states, so that this set and the generated set are judged by the same machinery. grasp_frame_recorded is what the recorded states say, kept for comparison -- the two agree to a few frames.


F6. Known limitation: the boundary sphere absorbs depth leak, and here you cannot see it

The mask's border bleeds a little past the object; the lifted points then include a few at the wrong depth. On this set that is nearly harmless, and the reason is worth stating because it does not transfer:

  • The croissant's AABB is 13.1 x 7.0 x 7.1 cm. Its longest axis is real, and it dominates the half-diagonal -- depth is only 43 % of it. Adding 5 cm of leak to the depth axis grows the sphere 24 %.
  • On the generated set the object is a block, 5.8 x 5.0 x 3.2 cm -- nearly isotropic, and depth is its shortest true axis, so it has the most room to be corrupted and takes over the diagonal the moment it is. The same 5 cm grows that sphere 42 %.
  • Add the scene: here a border leak lands on the tablecloth at the pastry's own depth; there it lands on a neighbouring object, which is a large jump.

Measured on the generated set, clips whose object-region depth spans under 5 cm give a 4.02 cm sphere while those spanning over 15 cm give 11.46 cm -- same object class, so what varies is the measurement, not the object. The fix is to select depth as a band around the object-region median rather than by percentile; the percentile trim cannot remove a leak larger than 5 % of the mask, and on a small object the leak is 10-20 %.

Nothing was changed here. The verdicts in world3d/ use the 5-95 percentile trim, and on this set the difference is small enough not to matter. Anyone porting these constants to a cluttered scene or a smaller object should re-derive them first.



ActionNet subset100 โ€” GT depth ํฌํ•จ 100 ์—ํ”ผ์†Œ๋“œ

ํ•œ ์ค„ ์š”์•ฝ: LeRobot ๋ณ€ํ™˜๋ณธ์—๋Š” depth๊ฐ€ ์—†๊ณ  RGB๋„ 192ร—288๋กœ ์ค„์–ด ์žˆ์–ด์„œ, ์›๋ณธ tar์—์„œ 1280ร—800 RGB + ๋ฌด์†์‹ค GT depth๋ฅผ ๊ฐ€์ง„ 100 ์—ํ”ผ์†Œ๋“œ๋งŒ ๊ณจ๋ผ ์ž๋ฆฝํ˜•์œผ๋กœ ๋ฌถ์€ ๊ฒƒ์ž…๋‹ˆ๋‹ค.

์ž‘์„ฑ 2026-08-20. ์ƒ์œ„ ๋””๋ ‰ํ† ๋ฆฌ์˜ README_MAPPING.md(๋‚ด๋ถ€ ๋ฌธ์„œ, ๊ณต๊ฐœ๋˜์ง€ ์•Š์Œ)๊ฐ€ ์›๋ณธ โ†” LeRobot ๋Œ€์‘์˜ ๊ทผ๊ฑฐ ๋ฌธ์„œ์ด๊ณ , ์ด ๋ฌธ์„œ๋Š” ๊ทธ์ค‘ 100๊ฐœ ๋ถ€๋ถ„์ง‘ํ•ฉ๊ณผ ํŒŒ์ƒ ์‚ฐ์ถœ๋ฌผ๋งŒ ๋‹ค๋ฃน๋‹ˆ๋‹ค.

์ •์ • 2026-08-20 (๋‹น์ผ ๊ฐฑ์‹ ) โ€” derived/states๋ฅผ ๊ต์ฒดํ–ˆ์Šต๋‹ˆ๋‹ค

๋ฌด์—‡ ์ด์ „ ์ง€๊ธˆ
derived/states/ LeRobot parquet์—์„œ ๋ฝ‘์•„ stride 4๋กœ ์ •๋ ฌ ์›๋ณธ hdf5๋ฅผ ์˜์ƒ์˜ timestamps.json ์‹œ๊ณ„์— ์ตœ๊ทผ์ ‘ ๋ฆฌ์ƒ˜ํ”Œ
์ด์ „ ํŒ์˜ ์œ„์น˜ โ€” derived/states_lerobot_WRONG_do_not_use/ (์ง€์šฐ์ง€ ์•Š๊ณ  ๋‚จ๊น€)
LeRobot ๋ณ€ํ™˜๋ณธ fps (ยง4 ํ‘œ) "15 fps" 30 fps, ํ”„๋ ˆ์ž„ ์ˆ˜๊ฐ€ ์›๋ณธ๊ณผ ๋™์ผ (447/432/491 ์‹ค์ธก, meta/info.json fps: 30)

์›์ธ์€ ์ƒ์œ„ README_MAPPING.md ยง4(๋‚ด๋ถ€ ๋ฌธ์„œ)์˜ ์ •๋ ฌ ์˜ค๋ฅ˜์ž…๋‹ˆ๋‹ค โ€” LeRobot์˜ state ํ–‰์€ 30 Hz๊ฐ€ ์•„๋‹ˆ๋ผ 60 Hz ์ธ๋ฑ์Šค์ž…๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์˜์ƒ ํ”„๋ ˆ์ž„ ๋ฒˆํ˜ธ๋กœ state๋ฅผ ์ธ๋ฑ์‹ฑํ•œ ์‚ฐ์ถœ๋ฌผ์€ ์ „๋ถ€ ๋‹ค์‹œ ๋งŒ๋“ค์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค(์›€์ง์ด๋Š” ํŒ” FK๊ฐ€ /state/pose์—์„œ 100 ep ์ค‘์•™ 7.59 cm ๋ฒ—์–ด๋‚จ, ์ตœ๋Œ€ 18.66 cm, 99/100 ep๊ฐ€ 3 cm ์ดˆ๊ณผ. ์ •์ง€ํ•œ ํŒ”์€ 0.33 cm๋ผ ํ•œ์ชฝ๋งŒ ๋ณด๋ฉด ์•ˆ ๋ณด์ž„). ๋งˆ์Šคํฌยทpointsยทtracks๋Š” ์˜ํ–ฅ ์—†์Šต๋‹ˆ๋‹ค โ€” ์˜์ƒ ํ”„๋ ˆ์ž„์—์„œ ์ง์ ‘ ๋งŒ๋“  ๊ฒƒ์ด๊ณ , ์ƒˆ derived/states์˜ frame_indices๋„ ๊ฐ™์€ ์›๋ณธ ๋น„๋””์˜ค ํ”„๋ ˆ์ž„ ์ถ•์ด๋ผ ๊ทธ๋Œ€๋กœ ๋Œ€์‘ํ•ฉ๋‹ˆ๋‹ค.

ยง5์— ์ถ”๊ฐ€๋œ ์‚ฌ์‹ค ํ•˜๋‚˜ ๋”: ๊ณต์œ  ๋งˆ์Šคํฌ๋Š” ์›๋ณธ ์–ด์•ˆ ์˜์ƒ ์ขŒํ‘œ๊ณ„๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค โ€” ์–ธ๋””์Šคํ† ํŠธ๋œ pinhole ๋ Œ๋”๋ง ์œ„์—์„œ ๊ณ„์‚ฐ๋์Šต๋‹ˆ๋‹ค. raw rgb.mp4์— ๊ทธ๋Œ€๋กœ ์–น์œผ๋ฉด ์ฃผ๋ณ€๋ถ€์—์„œ 100 px ์ด์ƒ ํ‹€์–ด์ง‘๋‹ˆ๋‹ค. ยง5 ์ฐธ์กฐ.


1. ๊ฒฝ๋กœ

๊ฒฝ๋กœ ํ‘œ๊ธฐ ์ฃผ์˜: ์•„๋ž˜ ํ‘œ๋Š” ์ด ํŒจํ‚ค์ง€๊ฐ€ ์šฐ๋ฆฌ ํด๋Ÿฌ์Šคํ„ฐ์— ๋†“์ธ ์œ„์น˜์ด๊ณ , ์™ธ๋ถ€ ๋…์ž์—๊ฒŒ๋Š” ์˜๋ฏธ๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. HF ์ €์žฅ์†Œ์—์„œ๋Š” ๊ฐ™์€ ๊ฒƒ๋“ค์ด ์ €์žฅ์†Œ ๋ฃจํŠธ์— ์žˆ๊ณ , ํŒŒ์ƒ ์‚ฐ์ถœ๋ฌผ์€ derived/๊ฐ€ ์•„๋‹ˆ๋ผ ์ตœ์ƒ์œ„ gtdepth/ masks/ states/ handmasks/ world3d/ occlusion/์ž…๋‹ˆ๋‹ค.

๋ฌด์—‡ ๊ฒฝ๋กœ (๋‚ด๋ถ€)
์ด ๋””๋ ‰ํ† ๋ฆฌ <๋‚ด๋ถ€ ์•„ํ‹ฐํŒฉํŠธ ๋งˆ์šดํŠธ>/action_net_og/subset100_gtdepth/
โ”œ ํŽ˜์ด๋กœ๋“œ tar (100 ep, ~4.4 GB) subset100_gtdepth.tar
โ”œ ๋ช…๋‹จ (ํ•œ ์ค„์— ULID ํ•˜๋‚˜) ulids.txt
โ”œ ๋Œ€์‘ํ‘œ mapping_subset100.csv / .json
โ”œ ์žฌํ˜„ ์Šคํฌ๋ฆฝํŠธ build_subset100.py
โ”” ํŒŒ์ƒ ์‚ฐ์ถœ๋ฌผ (์šฐ๋ฆฌ Stage-5 ์ž‘์—…๋ฌผ) derived/
์ƒ์œ„: ์›๋ณธ tar 4๊ฐœ + ์ „์ฒด ๋Œ€์‘ํ‘œ ..
HF ์›๋ณธ https://huggingface.co/datasets/FourierIntelligence/ActionNet

2. tar ์•ˆ์— ๋ฌด์—‡์ด ์žˆ๋Š”๊ฐ€

์—ํ”ผ์†Œ๋“œ๋‹น 4๊ฐœ ํŒŒ์ผ, ํ‰๊ท  44 MB (rgb 31 MB + depth 11.7 MB + hdf5 1 MB):

<ULID>/top/rgb.mp4          1280x800, 30 fps, h264
<ULID>/top/depth.mkv        1280x800, 16-bit, ffv1 / gray16le  (๋ฌด์†์‹ค)
<ULID>/top/timestamps.json  ํ”„๋ ˆ์ž„๋ณ„ ํƒ€์ž„์Šคํƒฌํ”„ ๋ฌธ์ž์—ด
<ULID>.hdf5                 /state/{robot(32),hand(12),pose(27)}, /action/{...}, /timestamp

์›๋ณธ tar๊ณผ ๋™์ผํ•œ ๋ ˆ์ด์•„์›ƒ์ด๋ผ, ๊ธฐ์กด์— tar์„ ๋‹ค๋ฃจ๋˜ ์ฝ”๋“œ๊ฐ€ ๊ทธ๋Œ€๋กœ ๋•๋‹ˆ๋‹ค.


3. ์ด 100๊ฐœ๊ฐ€ ๋ฌด์—‡์ธ๊ฐ€ โ€” ํฌ๋ฃจ์•„์ƒ ๊ณผ์ œ๋กœ ์˜๋„์ ์œผ๋กœ ๊ณ ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค

ํ•ญ๋ชฉ ๊ฐ’
์ถœ์ฒ˜ tar 48๊ฐœ โ† 01JMEX00XF-01JN0B0VXS.tar, 52๊ฐœ โ† 01JN0B1N7D-01JN0Y2AWR.tar
distinct prompt 3์ข…๋ฟ
โ”œ Pick up the croissant and put it on a plate 55
โ”œ Put the croissant in the container 42
โ”” Place the croissant into the container 3

์ „๋ถ€ ํฌ๋ฃจ์•„์ƒ ๊ณผ์ œ์ด๊ณ , ์ด๋Š” ์˜๋„ํ•œ ์„ ํƒ์ž…๋‹ˆ๋‹ค. Stage-5 ํ•„ํ„ฐ๋ง์€ "์ƒ์„ฑ ์˜์ƒ์ด ์žก๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๋ณด์ผ ๋•Œ IDM action์ด 3D์—์„œ ์‹ค์ œ๋กœ ๋ฌผ์ฒด์— ๋‹ฟ์•˜๋Š”๊ฐ€"๋ฅผ ๊ฒ€์ฆํ•˜๋Š” ์ผ์ด๊ณ , ๊ทธ๊ฑธ ์ฒ˜์Œ ์„ธ์šธ ๋•Œ๋Š” ๊ณผ์ œ๋ฅผ ํ•˜๋‚˜๋กœ ๊ณ ์ •ํ•˜๋Š” ํŽธ์ด ํ›จ์”ฌ ์‰ฝ์Šต๋‹ˆ๋‹ค โ€” ๋ฌผ์ฒด๊ฐ€ ํ•˜๋‚˜๋ฟ์ด๋ฉด ๋ฌผ์ฒด ๋งˆ์Šคํฌยท3D ๋ฐ•์Šคยท์ ‘์ด‰ ํŒ์ •์ด ๊ณผ์ œ๋ณ„ ๋ณ€์ˆ˜ ์—†์ด ๋น„๊ต๋ฉ๋‹ˆ๋‹ค. ์„ธ ๋ฌธ์žฅ์€ ํ‘œํ˜„๋งŒ ๋‹ค๋ฅด๊ณ  ๊ฐ™์€ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค.

๊ณผ์ œ ๋‹ค์–‘์„ฑ์ด ํ•„์š”ํ•œ ์‹คํ—˜์—๋Š” ์ƒ์œ„ ๋””๋ ‰ํ† ๋ฆฌ์˜ ์ „์ฒด ๋Œ€์‘ํ‘œ์—์„œ ๋‹ค์‹œ ๊ณ ๋ฅด๋ฉด ๋ฉ๋‹ˆ๋‹ค (mapping_*.csv์˜ prompt / task_group ์—ด).


4. depth๋ฅผ ์“ธ ๋•Œ ์•Œ์•„์•ผ ํ•˜๋Š” ๊ฒƒ (์ธก์ •์œผ๋กœ ํ™•์ธํ•œ ์‚ฌ์‹ค)

์‚ฌ์‹ค ๊ทผ๊ฑฐ
์ ˆ๋Œ€ metric์ด๋‹ค (์ƒ๋Œ€ยทaffine ๋ชจํ˜ธ์„ฑ ์—†์Œ) ํŒŒ์ผ์€ uint16 ๋ฐ€๋ฆฌ๋ฏธํ„ฐ์ด๊ณ  ๋กœ๋”๊ฐ€ /1000.0์œผ๋กœ ๋ฏธํ„ฐ๋กœ ๋ฐ”๊พผ๋‹ค. ์›์‹œ ๋ถ„ํฌ min 251~260, ์ค‘์•™ 771~799, max 2698 mm. ๊ธฐํ•˜๋กœ ๊ต์ฐจ๊ฒ€์ฆ: ์นด๋ฉ”๋ผ 65 cm ๋†’์ดยทํ…Œ์ด๋ธ” 9 cmยทtilt 42ยฐ โ†’ ํ…Œ์ด๋ธ” ์ค‘์•™ ์‚ฌ๊ฑฐ๋ฆฌ โ‰ˆ 82 cm, ์‹ค์ธก 7780 cm. ๋˜ depth๋กœ ์ ํ•ฉํ•œ ๊ด‘ํ•™์ค‘์‹ฌ(base_link ์œ„ 65.14 cm)์ด URDF FK์˜ head_pitch_link(64.43 cm)์™€ 0.7 cm ์ฐจ
ํฌํ™” ์ƒํ•œ 2698 mm ๋ชจ๋“  ํ”„๋ ˆ์ž„์—์„œ p99 = max = 2698. 2.7 m ๋ฐ–์€ ์ž˜๋ ค ์žˆ์œผ๋‹ˆ ๋ฐฐ๊ฒฝ ๊ฑฐ๋ฆฌ๋กœ ์“ฐ๋ฉด ์•ˆ ๋œ๋‹ค
๋ฌดํšจ ํ”ฝ์…€ 17~18% ๊ฐ’ 0. ๋งˆ์Šคํฌยทํ‰๋ฉด ์ ํ•ฉ ์ „์— depth > 0์œผ๋กœ ๊ฑธ๋Ÿฌ์•ผ ํ•œ๋‹ค
z-depth์ด๋‹ค (๊ด‘์„  ๋ฐฉํ–ฅ ๊ฑฐ๋ฆฌ๊ฐ€ ์•„๋‹ˆ๋‹ค) ํ…Œ์ด๋ธ”๋ณด ํ‰๋ฉด ์ž”์ฐจ๋ฅผ ์ „ ํ™”๋ฉด ๋ฐ˜๊ฒฝ์—์„œ ์žฌ๋ฉด z ํ•ด์„์€ 0.83 cm ์ด๋‚ด๋กœ ํ‰ํ‰ํ•˜๊ณ , range ํ•ด์„์€ โˆ’2.85 cm๋กœ ํœœ๋‹ค. ์ค‘์•™ ROI๋งŒ ๋ณด๋ฉด ๋‘ ํ•ด์„์ด ๊ตฌ๋ถ„๋˜์ง€ ์•Š๋Š”๋‹ค (cos ฮธ โ‰ˆ 1)
ํ‰๋ฉด์„ฑ ์ „์ฒด ํ…Œ์ด๋ธ”๋ณด๋กœ ๊ฐ•๊ฑด ์ ํ•ฉ ์‹œ ์ž”์ฐจ๊ฐ€ ์ „ ๋ฐ˜๊ฒฝ 0.63 cm ์ด๋‚ด
์นด๋ฉ”๋ผ ์บ˜๋ฆฌ๋ธŒ๋ ˆ์ด์…˜์€ ๋ฐ์ดํ„ฐ์…‹์— ์—†๋‹ค tar ์—”ํŠธ๋ฆฌ ์ „์ˆ˜ยทhdf5 attrsยทmeta/info.json์„ ๋‹ค ํ™•์ธํ–ˆ๋‹ค. hdf5 root์—๋Š” camera_names=['top']๋ฟ์ด๊ณ  intrinsicยทdistortionยทextrinsic์€ ์–ด๋””์—๋„ ์—†๋‹ค
์–ด์•ˆ ์ดˆ์ ๊ฑฐ๋ฆฌ f_native = 577.47 px ์ถœ์ฒ˜ ์—†๋Š” ๋ฐ์ดํ„ฐ์‹œํŠธ ์ถ”์ •์น˜. ๋‹ค๋งŒ ์ธ์šฉ๋œ ์„ธ FOV(H 127ยฐ, V 79.5ยฐ, D 150ยฐ)๊ฐ€ r = fยทฮธ ์•„๋ž˜ 0.16%๋กœ ์ผ์น˜ํ•˜๊ณ  pinhole๋กœ๋Š” 83% ๋ฒŒ์–ด์ง€๋ฏ€๋กœ, ๋ชจ๋ธ์€ ์ž˜ ๋’ท๋ฐ›์นจ๋œ๋‹ค. ์ƒ์„ธ๋Š” filtering/docs/camera.md (๋‚ด๋ถ€ RoboCurate_V2, ๋น„๊ณต๊ฐœ)
์‹œ๊ณ„๊ฐ€ ๋‘ ๊ฐœ๋‹ค (2026-08-20 ์ •์ •) hdf5 59.9 Hz / ์›๋ณธ ๋น„๋””์˜ค 30 fps / LeRobot ์˜์ƒ๋„ 30 fps(ํ”„๋ ˆ์ž„ ์ˆ˜๊ฐ€ ์›๋ณธ๊ณผ ๋™์ผ). ์ด์ „ ํŒ์˜ "LeRobot 15 fps"๋Š” ํ‹€๋ ธ๋‹ค โ€” meta/info.json fps: 30, ffprobe ์‹ค์ธก 447/432/491 ํ”„๋ ˆ์ž„์ด ์›๋ณธ ์นด๋ฉ”๋ผ ํ”„๋ ˆ์ž„ ์ˆ˜์™€ ๊ฐ™๋‹ค. ํ–‰ ์ˆ˜๋Š” 30 fps์ธ๋ฐ LeRobot state ๊ฐ’์€ 60 Hz ์ธ๋ฑ์Šค์ด๋ฏ€๋กœ ํ”„๋ ˆ์ž„ ๋Œ€์‘์€ ๋ฐ˜๋“œ์‹œ ๊ฐ ํŒŒ์ผ์˜ frame_indicesยทtimestamp๋กœ ๋งž์ถœ ๊ฒƒ

hdf5์˜ /state/pose๋Š” ๊ทธ๋ƒฅ ์ง€๋‚˜์น˜๊ธฐ ์•„๊น์Šต๋‹ˆ๋‹ค

(T, 27) = 3 ์ž์„ธ ร— (์œ„์น˜ 3 + ํšŒ์ „ 6) ์ด๊ณ , ์ˆœ์„œ๋Š” ์™ผ EEF / ์˜ค๋ฅธ EEF / ๋จธ๋ฆฌ์ž…๋‹ˆ๋‹ค. GR1T1 URDF๋กœ FK๋ฅผ ๋Œ๋ ค ๋Œ€์กฐํ•˜๋ฉด ์ด๋ ‡๊ฒŒ ๋งž์Šต๋‹ˆ๋‹ค:

๋ฐ์ดํ„ฐ์…‹ /state/pose URDF FK ์ฐจ์ด
์™ผ EEF [0.2143, 0.2437, 0.1735] left_end_effector_link [0.2132, 0.2451, 0.1697] 0.4 cm
์˜ค๋ฅธ EEF [0.2268, โˆ’0.2085, 0.1733] right_end_effector_link [0.2284, โˆ’0.2071, 0.1752] 0.3 cm
๋จธ๋ฆฌ [โˆ’0.0041, โˆ’0.001, 0.6443] head_pitch_link [โˆ’0.0021, 0.005, 0.6443] 0.6 cm

์ฆ‰ ์šด๋™ํ•™๊ณผ state ๋งคํ•‘์„ ๋ฐ์ดํ„ฐ์…‹ ์ž์ฒด๋กœ ๊ฒ€์ฆํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋งŒ ์„ธ ๊ฐ€์ง€๋ฅผ ์ง€ํ‚ค์„ธ์š”.

  1. ๋งž๋Š” ๋งํฌ๋Š” {left,right}_end_effector_link์ž…๋‹ˆ๋‹ค. {L,R}_hand_base_link๋กœ ๋น„๊ตํ•˜๋ฉด ์•ฝ 2.1 cm๊ฐ€ ๋‚จ์Šต๋‹ˆ๋‹ค(์—ํ”ผ์†Œ๋“œ ์ „์ฒด์—์„œ ๊ฑฐ์˜ ์ผ์ • = ๋งํฌ ์„ ํƒ ์ฐจ์ด, ์˜ค์ฐจ๊ฐ€ ์•„๋‹˜). ์ฆ‰ ์ด ๋ฐ์ดํ„ฐ์…‹์ด ๋งํ•˜๋Š” "EEF"๋Š” hand_base๊ฐ€ ์•„๋‹ˆ๋ผ end_effector ๋งํฌ์ž…๋‹ˆ๋‹ค.
  2. /state/pose๋Š” 60 Hz์ž…๋‹ˆ๋‹ค โ€” ํ–‰ ์ˆ˜๊ฐ€ mp4 ํ”„๋ ˆ์ž„ ์ˆ˜์˜ ์•ฝ 2๋ฐฐ์ž…๋‹ˆ๋‹ค(892/447, 799/401, 1133/567 โ€ฆ). ๊ทธ๋ž˜์„œ P[์˜์ƒํ”„๋ ˆ์ž„]์œผ๋กœ ์ธ๋ฑ์‹ฑํ•˜๋ฉด ํ‹€๋ฆฝ๋‹ˆ๋‹ค. frame 0๋ผ๋ฆฌ๋Š” ๊ทธ๋ƒฅ ๋งž์ง€๋งŒ, ๊ทธ ๋ฐ–์˜ ํ”„๋ ˆ์ž„์€ /timestamp๋กœ ๋ฆฌ์ƒ˜ํ”Œํ•˜์„ธ์š”. โš ๏ธ LeRobot์˜ state ํ–‰๋„ ๋˜‘๊ฐ™์ด 60 Hz ์ธ๋ฑ์Šค์ž…๋‹ˆ๋‹ค (2026-08-20 ์‹ค์ธก: observation.robot_joints[k] == /state/robot[k], 100 ep ์ค‘ 97๊ฑด์ด max|diff| 1e-5 ์ด๋‚ด, ์ค‘์•™ 5.949e-08; [2k] ๊ฐ€์„ค์€ 0/100, ์ค‘์•™ 0.740). LeRobot์˜ ํ–‰ ์ˆ˜๋Š” ์˜์ƒ ํ”„๋ ˆ์ž„ ์ˆ˜์™€ ๊ฐ™์•„์„œ "๊ทธ๋Ÿฌ๋‹ˆ ์˜์ƒ ํ”„๋ ˆ์ž„์œผ๋กœ ์ธ๋ฑ์‹ฑํ•ด๋„ ๋œ๋‹ค"๊ณ  ์ฐฉ๊ฐํ•˜๊ธฐ ์‰ฌ์šด๋ฐ, ํ–‰ ์ˆ˜๊ฐ€ ๋งž๋Š” ๊ฒƒ๊ณผ ๋‚ด์šฉ์˜ ์‹œ๊ณ„๊ฐ€ ๋งž๋Š” ๊ฒƒ์€ ๋‹ค๋ฅธ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. hdf5๋“  LeRobot์ด๋“  ์˜์ƒ ํ”„๋ ˆ์ž„ ๋ฒˆํ˜ธ๋กœ state๋ฅผ ์ธ๋ฑ์‹ฑํ•˜์ง€ ๋งˆ์„ธ์š” โ€” ์‹œ๊ณ„๋กœ ๋ฆฌ์ƒ˜ํ”Œํ•˜์„ธ์š”. ๊ทผ๊ฑฐ๋Š” ์ƒ์œ„ README_MAPPING.md ยง4(๋‚ด๋ถ€ ๋ฌธ์„œ).
  3. /state/robot์˜ ์ฑ„๋„ ์ˆœ์„œ๋Š” LeRobot 44์ฐจ์›๊ณผ ๋‹ค๋ฆ…๋‹ˆ๋‹ค โ€” ๋‹ค๋งŒ ๊ทธ ์ˆœ์„œ๋Š” ์ด์ œ ์ธก์ •์œผ๋กœ ํ™•์ •๋ผ ์ƒ์œ„ README_MAPPING.md ยง6-2์— ํ‘œ๋กœ ์ ํ˜€ ์žˆ์Šต๋‹ˆ๋‹ค(์ ์šฉํ•˜๋ฉด 100/100 ์—ํ”ผ์†Œ๋“œ์—์„œ max|diff| = 0). derived/states์˜ 44์ฐจ์›๋„ ๊ทธ ์ˆœ์„œ์ž…๋‹ˆ๋‹ค.

5. derived/ โ€” ์šฐ๋ฆฌ๊ฐ€ ๋งŒ๋“  ํŒŒ์ƒ ์‚ฐ์ถœ๋ฌผ (์ง„ํ–‰ ์ค‘)

๊ฒฝ๋กœ ๊ฐœ์ˆ˜ ๋‚ด์šฉ
derived/states/states_<ULID>.npz 100 ์›๋ณธ hdf5์—์„œ ๋งŒ๋“  per-video-frame ์ƒํƒœ, ฯ„ ๋ณด์ • ์ „. ์•„๋ž˜ ์ฐธ์กฐ (2026-08-20 ์žฌ์ƒ์„ฑ)
derived/states_tau/ โ†’ states/ 100 ฯ„ ๋ณด์ • ์ƒํƒœ. ์˜์ƒ ํ”„๋ ˆ์ž„๊ณผ ์ง์ง€์„ ๋•Œ๋Š” ์ด๊ฒƒ์„ ์“ฐ์„ธ์š”. ๊ทผ๊ฑฐ๋Š” states/README.md
derived/states_lerobot_WRONG_do_not_use/ 100 ์“ฐ์ง€ ๋งˆ์„ธ์š”. ์ •๋ ฌ์ด ํ‹€๋ฆฐ ์ด์ „ ํŒ(LeRobot ์œ ๋ž˜, stride 4). ํŒ์ •์šฉ์œผ๋กœ๋งŒ ๋‚จ๊ฒจ ๋‘ก๋‹ˆ๋‹ค
derived/points/vlm_points_<ULID>.json 99 ๋ฌผ์ฒด point prompt (positive/negative), ์—ํ”ผ์†Œ๋“œ๋ณ„ ์ˆ˜๋™ ์ง€์ •
derived/masks/ โ†’ masks/ 100 SAM3 ๋ฌผ์ฒด ๋งˆ์Šคํฌ, ์ „ ํ”„๋ ˆ์ž„. 43๊ฐœ๋ฟ์ด๋˜ derived/masks/๋Š” ์‚ญ์ œํ–ˆ๊ณ  ์ตœ์ƒ์œ„ masks/๊ฐ€ 100ํŽธ ์ „๋ถ€์ž…๋‹ˆ๋‹ค
derived/hand_points2/points_<ULID>.json 10 ํŒ”+์† point prompt: ์†๋‹น 3์ (์†๊ฐ€๋ฝยท์†๋ฐ”๋‹ฅยทํŒ”๋š) + negative
derived/hand_tracks3/tracks_<ULID>.npz 10 SAM3 ํŒ”+์† ๋งˆ์Šคํฌ, ์ „ ํ”„๋ ˆ์ž„, ์ขŒ/์šฐ ๋ถ„๋ฆฌ (left, right, frame_indices)

derived/states โ€” ์›๋ณธ hdf5์—์„œ ๋‹ค์‹œ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค (2026-08-20)

์ด์ „ ํŒ์€ LeRobot parquet์˜ state ์—ด์„ stride 4๋กœ ์˜์ƒ ํ”„๋ ˆ์ž„์— ์–น์€ ๊ฒƒ์ด์—ˆ๋Š”๋ฐ, ์ „์ œ("LeRobot 15 fps ร— stride 4 = 60 Hz")๊ฐ€ ๋‘˜ ๋‹ค ํ‹€๋ ธ์Šต๋‹ˆ๋‹ค โ€” LeRobot์€ 30 fps์ด๊ณ  state ํ–‰์€ 60 Hz ์ธ๋ฑ์Šค์ž…๋‹ˆ๋‹ค(ยง4, ์ƒ์œ„ README_MAPPING.md ยง4(๋‚ด๋ถ€ ๋ฌธ์„œ)). ์ง€๊ธˆ ํŒ์€ LeRobot์„ ์•„์˜ˆ ๊ฑฐ์น˜์ง€ ์•Š๊ณ , ์›๋ณธ hdf5๋ฅผ ์˜์ƒ ์ž์‹ ์˜ timestamps.json ์‹œ๊ณ„์— ์ตœ๊ทผ์ ‘ ๋ฆฌ์ƒ˜ํ”Œํ•ฉ๋‹ˆ๋‹ค (๋ณด๊ฐ„ ๊ธˆ์ง€ โ€” ์† 12์ฑ„๋„์ด ๊ณ„๋‹จ์‹ ์›์‹œ ๋‹จ์œ„๋ผ ๋ณด๊ฐ„ํ•˜๋ฉด ๋กœ๋ด‡์ด ๋ณด๊ณ ํ•œ ์  ์—†๋Š” ๊ฐ’์ด ์ƒ๊น๋‹ˆ๋‹ค). ๋นŒ๋”๋Š” ๋‚ด๋ถ€ RoboCurate_V2์˜ tools/build_states_hdf5.py(๋น„๊ณต๊ฐœ).

ํ‚ค shape (์˜ˆ: 01JMXF99BYFPWU5NOL, T=447) ๋‚ด์šฉ
states / actions (447, 44) float64 44์ฐจ์› GR1-T1 ์ˆœ์„œ (์ƒ์œ„ ๋ฌธ์„œ ยง6-2 ํ‘œ)
frame_indices (447,) int64 ์›๋ณธ ๋น„๋””์˜ค ํ”„๋ ˆ์ž„ ๋ฒˆํ˜ธ (0..Tโˆ’1, ๋งˆ์Šคํฌยทtracks์™€ ๊ฐ™์€ ์ถ•)
hdf5_rows (447,) int64 ๊ฐ ํ”„๋ ˆ์ž„์ด ๊ณ ๋ฅธ hdf5 ํ–‰
pose (447, 3, 9) float64 /state/pose ๊ฐ™์€ ํ–‰ (์™ผ EEF / ์˜ค๋ฅธ EEF / ๋จธ๋ฆฌ)
video_t, lag_s, clock_offset_s (447,), (447,), scalar ์˜์ƒ ์‹œ๊ณ„ ์ดˆ, ํ”„๋ ˆ์ž„โˆ’ํ–‰ ์ž”์ฐจ, hdf5โ†”์˜์ƒ ์‹œ๊ณ„ ์˜คํ”„์…‹

๊ฒ€์ฆ (100 ep ์ „์ˆ˜, FK vs ๋ฐ์ดํ„ฐ์…‹ ์ž์‹ ์˜ /state/pose):

์ค‘์•™ ์ตœ์•… 3 cm ์ดˆ๊ณผ
์™ผ EEF 0.29 cm 0.37 cm 0/100
์˜ค๋ฅธ EEF 0.27 cm 0.32 cm 0/100
๋จธ๋ฆฌ 0.59 cm 0.60 cm 0/100

ํ”„๋ ˆ์ž„โ†”ํ–‰ ์ž”์ฐจ(lag_s)๋Š” 46,927 ํ”„๋ ˆ์ž„ ์ค‘ 98.6%๊ฐ€ 8.34 ms(60 Hz ๋ฐ˜ ์นธ) ์ด๋‚ด (์ค‘์•™ 2.51 ms, p95 6.42 ms, 16.7 ms ์ดˆ๊ณผ 1 ํ”„๋ ˆ์ž„, ์ตœ์•… 24.35 ms โ€” ํ•œ ์—ํ”ผ์†Œ๋“œ 01JN0XAGCRDDJ33ML4์— ๋ชฐ๋ ค ์žˆ์Œ). ๊ฐ ์—ํ”ผ์†Œ๋“œ์˜ ๋งˆ์ง€๋ง‰ ํ”„๋ ˆ์ž„๋งŒ hdf5๊ฐ€ ์˜์ƒ๋ณด๋‹ค ๋จผ์ € ๋๋‚˜์„œ โˆ’12.0 ~ โˆ’57.7 ms(์ค‘์•™ โˆ’29.7 ms) ๋ฒŒ์–ด์ง‘๋‹ˆ๋‹ค. ์‹œ๊ณ„ ์˜คํ”„์…‹์€ โˆ’21.6 ~ +11.1 ms. ๋งˆ์ง€๋ง‰ ํ”„๋ ˆ์ž„์„ ์“ธ ๊ฑฐ๋ฉด ์ด ๊ฐ’์„ ๊ฐ์•ˆํ•˜์„ธ์š”.

โš ๏ธ ๋งˆ์Šคํฌ๋Š” ์›๋ณธ ์–ด์•ˆ ์ขŒํ‘œ๊ณ„๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค โ€” ์–ธ๋””์Šคํ† ํŠธ๋œ pinhole ๊ณต๊ฐ„์ž…๋‹ˆ๋‹ค

derived/hand_tracks3/์™€ derived/masks/๋Š” raw rgb.mp4 ์œ„์—์„œ ๊ณ„์‚ฐ๋œ ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. filtering/src/depth/undistort_fisheye.py --hfov 90์œผ๋กœ ๋‹ค์‹œ ๋ Œ๋”ํ•œ rectilinear(pinhole) ์˜์ƒ ์œ„์—์„œ SAM3๋ฅผ ๋Œ๋ฆฐ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ์ฆ‰ ๋‘ npz์˜ ์ขŒํ‘œ๊ณ„๋Š”:

ํ•ญ๋ชฉ ๊ฐ’
ํ•ด์ƒ๋„ 1280ร—800 (์›๋ณธ๊ณผ ๊ฐ™์Œ โ€” ๊ทธ๋ž˜์„œ ์œก์•ˆ์œผ๋กœ ๊ตฌ๋ถ„๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค)
K [[640, 0, 639.5], [0, 640, 399.5], [0, 0, 1]] (hfov 90ยฐ)
ํˆฌ์˜ pinhole (r = fยทtanฮธ). ์›๋ณธ์€ equidistant ์–ด์•ˆ (r = fยทฮธ)
npz ์•ˆ์˜ ์ฆ๊ฑฐ video ํ‚ค = /tmp/rc_undist_<ULID>.mp4 / /tmp/rc_full_<jobid>_<ULID>.mp4, ๋‘˜ ๋‹ค ์œ„ ์Šคํฌ๋ฆฝํŠธ ์ถœ๋ ฅ

raw rgb.mp4์— ์ด ๋งˆ์Šคํฌ๋ฅผ ๊ทธ๋Œ€๋กœ ์–น์œผ๋ฉด ์ฃผ๋ณ€๋ถ€์—์„œ 100 px ์ด์ƒ ํ‹€์–ด์ง‘๋‹ˆ๋‹ค. ๊ฐ™์€ ํ”ฝ์…€์ด ๋‘ ๊ณต๊ฐ„์—์„œ ์–ผ๋งˆ๋‚˜ ๋–จ์–ด์ ธ ์žˆ๋Š”์ง€(์œ„ ๋‘ ๋ชจ๋ธ๋กœ ๊ณ„์‚ฐ):

์–ธ๋””์Šคํ† ํŠธ ์˜์ƒ์˜ ํ”ฝ์…€ ๊ด‘์ถ•์—์„œ ๋Œ€์‘ํ•˜๋Š” ์›๋ณธ ์–ด์•ˆ ํ”ฝ์…€ ์–ด๊ธ‹๋‚จ
(1279, 399.5) ์˜ค๋ฅธ์ชฝ ๋ ์ค‘์•™ 44.98ยฐ (1092.8, 399.5) 186 px
(1279, 799) ์˜ค๋ฅธ์•„๋ž˜ ์ฝ”๋„ˆ 49.68ยฐ (1064.1, 664.8) 253 px
(959.5, 399.5) ์˜ค๋ฅธ์ชฝ ์ ˆ๋ฐ˜ ์ง€์  26.57ยฐ (907.2, 399.5) 52 px
ํ™”๋ฉด ์ค‘์•™ 0ยฐ ํ™”๋ฉด ์ค‘์•™ 0 px

์ค‘์•™์—์„œ๋Š” 0์ด๋ผ ์ค‘์•™ ROI๋งŒ ๋ณด๋ฉด ๋‘ ๊ณต๊ฐ„์ด ๊ตฌ๋ถ„๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค โ€” ์ด ์˜ค๋ฅ˜๊ฐ€ ๋ˆˆ์— ๋„์ง€ ์•Š๋Š” ์ด์œ ์ž…๋‹ˆ๋‹ค. ๋˜ hfov 90ยฐ๋Š” ์–ด์•ˆ์ด ๋‹ด์€ ํ™”๊ฐ(127ยฐ)๋ณด๋‹ค ์ข์œผ๋ฏ€๋กœ ์–ด์•ˆ์˜ ๋ฐ”๊นฅ ์ฃผ๋ณ€๋ถ€๋Š” ์••์ถ•๋˜์ง€ ์•Š๊ณ  ๋ฒ„๋ ค์ง‘๋‹ˆ๋‹ค(1280ร—800์—์„œ ๊นจ๋—ํ•˜๊ฒŒ ๋ฝ‘์„ ์ˆ˜ ์žˆ๋Š” ์ตœ๋Œ€ hfov๋Š” 95.9ยฐ).

๋งž๋Š” ์˜์ƒ์„ ๋‹ค์‹œ ๋งŒ๋“œ๋Š” ๋ฒ•:

cd <๋‚ด๋ถ€ RoboCurate_V2>/filtering/src
python depth/undistort_fisheye.py <๋ฐ์ดํ„ฐ>/data_og/$U/top/rgb.mp4 --hfov 90 -o /tmp/rc_undist_$U.mp4
# ํ”„๋ ˆ์ž„ ์ˆ˜ยท์ˆœ์„œ๋Š” ์›๋ณธ๊ณผ ๋™์ผํ•˜๋ฏ€๋กœ mask์˜ frame_indices๊ฐ€ ๊ทธ๋Œ€๋กœ ๋งž์Šต๋‹ˆ๋‹ค

์ƒ‰์ƒ ํŒ์ •ยทdepth ์กฐํšŒ๋„ ๋ฐ˜๋“œ์‹œ ๊ฐ™์€ ์–ธ๋””์Šคํ† ํŠธ ๊ณต๊ฐ„์—์„œ ํ•˜์„ธ์š”. depth๋Š” depth/gt_depth_npz.py(๊ธฐ๋ณธ --hfov 90)๊ฐ€ rgb์™€ depth๋ฅผ ๊ฐ™์€ ์–ด์•ˆโ†’pinhole ๋งต์œผ๋กœ ๋ฆฌ์ƒ˜ํ”Œํ•˜๋ฏ€๋กœ(RGB ์ด์ค‘์„ ํ˜•, depth ์ตœ๊ทผ์ ‘) ๊ทธ ์ถœ๋ ฅ๊ณผ ๋งˆ์Šคํฌ๊ฐ€ ํ”ฝ์…€ ๋‹จ์œ„๋กœ ๋Œ€์‘ํ•ฉ๋‹ˆ๋‹ค. raw depth.mkv๋ฅผ ์ง์ ‘ ์ฝ์–ด ๋งˆ์Šคํฌ ํ”ฝ์…€๋กœ ์ธ๋ฑ์‹ฑํ•˜๋ฉด ๋‹ค๋ฅธ ํ”ฝ์…€์˜ ๊ฑฐ๋ฆฌ๊ฐ’์„ ์ง‘๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.

๋งˆ์Šคํฌ๋ฅผ ์“ธ ๋•Œ์˜ ์ฃผ์˜ (์ˆจ๊ธฐ์ง€ ์•Š๊ณ  ์ ์Šต๋‹ˆ๋‹ค): ํŒ”+์† ๋งˆ์Šคํฌ๋Š” 10๊ฐœ ์—ํ”ผ์†Œ๋“œ ์ค‘ 8๊ฐœ์—์„œ ํ•œ์ชฝ์ด 61,3xx~61,43x px์— ๋ญ‰์ณ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ถ”์ถœ ์ƒํ•œ(max_obj_frac 0.06 ร— 1,024,000 = 61,440)์— ๊นŽ์—ฌ ๋ถ™์€ ํ…Œ์ด๋ธ”๋ณด ๋ˆ„์ถœ์ž…๋‹ˆ๋‹ค. ๊นจ๋—ํ•œ ๋งˆ์Šคํฌ๋Š” 14k33k px(ํ”„๋ ˆ์ž„์˜ 1.43.3%)์ด๋‹ˆ ๋ฉด์  ์ƒํ•œ 0.045๋กœ ๊ฑธ๋Ÿฌ์„œ ์“ฐ์„ธ์š”. ํ”„๋ ˆ์ž„ ์„ ๋ณ„๊ธฐ๋Š” ๋‚ด๋ถ€ RoboCurate_V2(๋น„๊ณต๊ฐœ)์˜ filtering/src/align/filter_mask_frames.py์— ์žˆ๊ณ , ํŒ์ •์— ์นด๋ฉ”๋ผ ์ž์„ธ๋ฅผ ์ผ์ ˆ ์“ฐ์ง€ ์•Š์Šต๋‹ˆ๋‹ค (ํ˜„์žฌ extrinsic๊ณผ ์ž˜ ๋งž๋Š” ํ”„๋ ˆ์ž„๋งŒ ๊ณ ๋ฅด๋ฉด ๊ทธ extrinsic์„ ํ™•์ฆํ•ด ๋ฒ„๋ฆฌ๊ธฐ ๋•Œ๋ฌธ).


6. ์“ฐ๋Š” ๋ฒ•

D=.        # HF ์ €์žฅ์†Œ๋ฅผ ๋ฐ›์€ ๋””๋ ‰ํ† ๋ฆฌ (๋‚ด๋ถ€์—์„œ๋Š” ์•„ํ‹ฐํŒฉํŠธ ๋งˆ์šดํŠธ ๊ฒฝ๋กœ)

# ์ „์ฒด ํ’€๊ธฐ (4.4 GB)
mkdir -p ~/actionnet_subset100 && tar -xf $D/subset100_gtdepth.tar -C ~/actionnet_subset100

# ํ•œ ์—ํ”ผ์†Œ๋“œ๋งŒ
U=01JMXF99BYFPWU5NOL
tar -xf $D/subset100_gtdepth.tar -C ~/actionnet_subset100 $U/ $U.hdf5

# 16-bit depth๋ฅผ numpy๋กœ (ffv1/gray16le๋Š” OpenCV๋กœ ์—ด๋ฆฌ์ง€ ์•Š์Šต๋‹ˆ๋‹ค)
ffmpeg -i $U/top/depth.mkv -f rawvideo -pix_fmt gray16le - \
  | python -c "import sys,numpy as np; \
      a=np.frombuffer(sys.stdin.buffer.read(),np.uint16).reshape(-1,800,1280); \
      print(a.shape, a.dtype, a[0].max())"

depth์˜ ๋‹จ์œ„๋Š” mm(uint16) ์ด๊ณ , ๋‚ด๋ถ€ RoboCurate_V2(๋น„๊ณต๊ฐœ)์˜ filtering/src/depth/gt_depth_npz.py๋Š” ์ด mkv์™€ rgb๋ฅผ ๊ฐ™์€ ์–ด์•ˆโ†’pinhole ๋งต์œผ๋กœ ๋ฆฌ์ƒ˜ํ”Œํ•ด(RGB ์ด์ค‘์„ ํ˜•, depth ์ตœ๊ทผ์ ‘) npz๋กœ ๋งŒ๋“ญ๋‹ˆ๋‹ค. depth๋ฅผ ๋ณด๊ฐ„ํ•˜๋ฉด ๋ฌผ์ฒด ๊ฒฝ๊ณ„์—์„œ ์—†๋Š” ๊ฑฐ๋ฆฌ๊ฐ’์ด ์ƒ๊ธฐ๋ฏ€๋กœ ์ตœ๊ทผ์ ‘์ด์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.


7. ์žฌํ˜„ยท๊ฒ€์ฆ

cd $D
python build_subset100.py --verify     # ๋Œ€์‘ํ‘œ ์žฌ์ƒ์„ฑ + 400๊ฐœ ํŽ˜์ด๋กœ๋“œ ํŒŒ์ผ ์กด์žฌ ํ™•์ธ
sha256sum -c subset100_gtdepth.tar.sha256

build_subset100.py๋Š” ULIDโ†”LeRobot ๋Œ€์‘์„ ๋‹ค์‹œ ๊ณ„์‚ฐํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค โ€” ์ƒ์œ„ build_mapping.py๊ฐ€ ๋งŒ๋“  per-tar ๋Œ€์‘ํ‘œ์—์„œ ์šฐ๋ฆฌ 100๊ฐœ ํ–‰๋งŒ ๊ณจ๋ผ๋‚ด๋ฏ€๋กœ, ๋‘ ํ‘œ๊ฐ€ ์–ด๊ธ‹๋‚  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฆ ๊ฒฐ๊ณผ๋Š” ์ž‘์„ฑ ์‹œ์ ์— 400/400์ด์—ˆ์Šต๋‹ˆ๋‹ค.

์ถ”๊ฐ€ 2026-08-21 โ€” derived/states_tau/์™€ ์™ผํŒ” ๋ณด์ •๊ฐ’

์˜์ƒ ํƒ€์ž„์Šคํƒฌํ”„๊ฐ€ ๋…ธ์ถœ ์‹œ๊ฐ๋ณด๋‹ค ์•ฝ 167 ms ๋Šฆ๊ฒŒ ๊ธฐ๋ก๋ผ ์žˆ์Œ์„ ๋งˆ์Šคํฌ ๋Œ€๋น„ ์‹ค์ธก์œผ๋กœ ํ™•์ธํ–ˆ๊ณ , ๊ทธ ์‹œํ”„ํŠธ๋ฅผ ๋ฐ˜์˜ํ•œ ์ƒํƒœ๋ฅผ ์ตœ์ƒ์œ„ states/(์˜ˆ์ „ derived/states_tau/)์— ๋‘์—ˆ์Šต๋‹ˆ๋‹ค(derived/states/๋Š” ๊ทธ๋Œ€๋กœ). ์˜์ƒ ํ”„๋ ˆ์ž„๊ณผ ์ƒํƒœ๋ฅผ ์ง์ง€์–ด ์“ฐ๋Š” ์šฉ๋„์—๋Š” states_tau๋ฅผ ์“ฐ์„ธ์š”. ์™ผํŒ” ๊ด€์ ˆ ์˜์  ๋ฐ”์ด์–ด์Šค ๋ณด์ •๊ฐ’์€ ์ €์žฅ์†Œ ๋ฃจํŠธ์˜ left_arm_offset_20260821.json์— ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘˜ ๋‹ค ๊ทผ๊ฑฐ์™€ ์ˆ˜์น˜๋Š” states/README.md์— ์ ์—ˆ์Šต๋‹ˆ๋‹ค.

Downloads last month
226