File size: 13,912 Bytes
942788e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
26dc037
c425b36
942788e
a4ea9dc
 
c425b36
 
 
 
 
 
 
 
 
 
 
 
 
 
a4ea9dc
 
942788e
 
 
26dc037
c425b36
 
85d7243
 
c425b36
 
942788e
c425b36
 
85d7243
c425b36
 
 
85d7243
c425b36
942788e
 
 
 
c425b36
942788e
 
 
 
 
 
 
 
 
 
c425b36
942788e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c425b36
 
 
942788e
 
 
 
 
 
 
 
 
 
 
c425b36
 
 
 
942788e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
"""Build GR00T dataset views (symlinked episodes + rewritten meta + stats) for the G1 tasks.

Usage: <groot python> prepare_groot_data.py <task> [<task> ...]   (tasks: b6 basket a3 a7)
B6 always uses the fixed first 100 episodes (0-99) of b6-v5t7-ready-start-20261109.
"""
import hashlib
import json
import os
import subprocess
import sys
from pathlib import Path

H = Path(__file__).resolve().parent
G = H.parent
P = G.parent
OUT = G / "groot-data"
TASKS = {
    "b6": (P / "b6-v5t7-ready-start-60ep-20261109/data/deliveries/g1/b6-v5t7-ready-start-20261109", 100),
    "basket": (P / "basket-current-dynamics-100-v21-b0ff522/data", None),  # HF basket-transfer-expert-v1 @current-dynamics-100-v21 b0ff522 (old ed77719 retired)
    "basketv2": (P / "basket-trimv2-c4df4d4/data", None),  # HF a20-basket-transfer-pre-reach-trim-expert-v2 @c4df4d40 (user 2026-11-14)
    "trolley": (P / "trolley-push-expert-v1-20261110/data/trolley_push", 100),  # HF trolley-push-expert-v1 @ef9024a: fixed first 100 of 103
    "b7": (P / "b7-black-clothes-hamper-v1-1bb002a/data", 100),  # HF b7-black-clothes-hamper-expert-v1 @1bb002a, first 100 of 107
    "a6": (P / "a6-door-pull-v1-cf3ee9b/data/door_pull", None),  # HF door-pull-expert-v1 @cf3ee9b (A6)
    "b7green": (P / "b7-green-clothes-hamper-v1-ada288c/data", None),  # HF b7-green-clothes-hamper-expert-v1 @ada288c, 100 episodes (user 2026-11-13)
    "a7v2": (P / "a7-cabinet-close-v2-aecf070/data", None),  # HF a7-cabinet-close-expert-v2 @aecf070, 100 episodes (user 2026-11-14)
    "b8": (P / "b8-table-wipe-v1-3b2e618/data", None),  # HF b8-table-wipe-expert-v1 @3b2e618, 100 episodes (user 2026-11-14)
    "b1": (P / "b1-riceball-microwave-v1-36da298/data", None),  # HF b1-riceball-microwave-expert-v1 @36da298, all 115 episodes (user 2026-11-14)
    "a10v3": (P / "a10-trolley-push-v3-6e410c5/data", None),  # HF a10-trolley-push-expert-v3 @6e410c5, all 148 episodes (user 2026-11-15)
    "a3v3": (P / "a3-fridge-open-v3-24882d5/data", None),  # HF a3-fridge-open-expert-v3 @24882d5, all 125 episodes, measured hands (user 2026-11-15)
    "a7v3": (P / "a7-cabinet-close-v3-c2e15b0/data", None),  # HF a7-cabinet-close-expert-v3 @c2e15b0, all 100 episodes, measured hands (user 2026-11-15)
    "b6v6": (P / "b6-pedal-bin-v6-ad2f5bf/data", None),  # HF b6-pedal-bin @ad2f5bf0 (v6 at the repo root), all 542 episodes (user 2026-11-15)
    "b5": (P / "b5-chair-push-v1-b85b7e9/data", None),  # HF b5-chair-push-expert-v1 @b85b7e9, all 100 episodes (user 2026-11-15)
    "a3v4": (P / "a3-fridge-open-v4-6393d4a/data", None),  # HF a3-fridge-open-expert-v4 @6393d4a, all 125 episodes (user 2026-11-15: v4 replaces v3)
    "a7v4": (P / "a7-cabinet-close-v4-90cc2b5/data", None),  # HF a7-cabinet-close-expert-v4 @90cc2b5, all 100 episodes (user 2026-11-15: v4 replaces v3)
    "b7greenv2": (P / "b7-green-clothes-hamper-v2-c12757e/data", None),  # HF b7-green-clothes-hamper @c12757e (v2 at the repo root), all 120 episodes (user 2026-11-14)
    "a10v2": (P / "a10-trolley-push-v2-d3abb4c/data", None),  # HF a10-trolley-push-expert-v2 @d3abb4c, all 150 episodes (user 2026-11-13)
    "a1": (P / "a1-ground-object-pickup-v1-f47f07e/data", None),  # HF a1-ground-object-pickup-expert-v1 @f47f07e (A1)
    "c6t": (P / "c6-test8-06c231d/data", None),  # c6 8-episode adaptation test only
    "c6": (P / "c6-bathroom-door-fetch-v1-3afba69/data", None),  # HF c6-bathroom-door-fetch-expert-v1 @3afba69 (C6)
    "a3": (P / "a3-a7-20261109/a3-fridge-open-expert-v1/780fcf1e4bd7db947a8c2d8f7b92e61f24f71b45", None),
    "a7": (P / "a3-a7-20261109/a7-cabinet-close-expert-v1/1935a98e8f55d4d1d2d1615ddb97c2679713595c", None),
    "a18": (P / "a18-h2-dataset-audit-8e956a10/dataset", None),
    "a17": (P / "a17-unscrew-bottle-v1-407e6e2/data", None),  # HF a17-unscrew-bottle-expert-v1 @407e6e2
    "a18c2": (P / "a18-h2-dataset-audit-8e956a10/dataset", None),  # A18, 2 cameras head + chest (user 2026-11-14 18:2x)
    "a17c2": (P / "a17-unscrew-bottle-v1-407e6e2/data", None),  # A17, 2 cameras head + chest (user 2026-11-14 18:2x)
    "a18cr2": (P / "a18-pour-water-cropped-v2-44cdd57/data", None),  # HF a18-pour-water-expert-cropped-v2 @44cdd57, head + chest (user 2026-11-15)
    "a17cr2": (P / "a17-unscrew-bottle-cropped-v2-3520b70/data", None),  # HF a17-unscrew-bottle-expert-cropped-v2 @3520b70, head + chest (user 2026-11-15)
    "a14": (P / "a14-unscrew-bulb-v1-b4390fc/data", None),  # HF a14-unscrew-bulb-expert-v1 @b4390fc (H2, 3 cameras)
    "a14v2": (P / "a14-unscrew-bulb-v2-8411a6d/data", None),  # HF a14-unscrew-bulb-expert-v2 @8411a6da: + action.planner (H2 guide 2026-11-14)
}
CONFIG = {"a18": "h2w3_body31_hand44_h16.py", "a17": "h2w3_body31_hand44_h16.py", "a14": "h2w3_body31_hand44_h16.py",
          "a18c2": "h2_body31_hand44_h16.py", "a17c2": "h2_body31_hand44_h16.py",
          "a18cr2": "h2_body31_hand44_h16.py", "a17cr2": "h2_body31_hand44_h16.py",
          "a14v2": "h2_body31_hand44_planner7_h16.py"}  # 2026-11-14 H2 guide: action82 incl. the planner command
# model-input cameras per H2 task (user 2026-11-12 22:0x: EVERY H2 task uses head + chest + wrist; third_person never)
CAMS = {"a18": ("head", "chest", "wrist"), "a17": ("head", "chest", "wrist"), "a14": ("head", "chest", "wrist"),
        "a18c2": ("head", "chest"), "a17c2": ("head", "chest"), "a18cr2": ("head", "chest"), "a17cr2": ("head", "chest"), "a14v2": ("head", "chest")}  # 2026-11-14: 2-camera re-train (user)
PLANNER = {"a14v2"}  # tasks whose action adds action.planner (7) after body_target + hands (H2 guide a14 2.2)
H2_GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29),
             "head": (29, 31), "left_hand": (31, 53), "right_hand": (53, 75)}


def h2_modality(info, cams=("head", "chest")):
    f = info["features"]
    names = f["observation.state"]["names"]
    assert len(names) == 75 and names[:31] == f["action.body_target"]["names"] and names[31:] == f["action.hands"]["names"]
    assert f["observation.projected_gravity"]["shape"] == [3]
    state = {g: {"original_key": "observation.state", "start": a, "end": b} for g, (a, b) in H2_GROUPS.items()}
    state["projected_gravity"] = {"original_key": "observation.projected_gravity", "start": 0, "end": 3}
    return {"state": state,
            "action": {"body_target": {"original_key": "action.body_target", "start": 0, "end": 31},
                       "left_hand_target": {"original_key": "action.hands", "start": 0, "end": 22},
                       "right_hand_target": {"original_key": "action.hands", "start": 22, "end": 44}},
            "video": {c: {"original_key": f"observation.images.{c}"} for c in cams},
            "annotation": {"human.task_description": {"original_key": "task_index"}}}
BODY = ["left_hip_pitch_joint", "left_hip_roll_joint", "left_hip_yaw_joint", "left_knee_joint", "left_ankle_pitch_joint",
        "left_ankle_roll_joint", "right_hip_pitch_joint", "right_hip_roll_joint", "right_hip_yaw_joint",
        "right_knee_joint", "right_ankle_pitch_joint", "right_ankle_roll_joint", "waist_yaw_joint", "waist_roll_joint",
        "waist_pitch_joint", "left_shoulder_pitch_joint", "left_shoulder_roll_joint", "left_shoulder_yaw_joint",
        "left_elbow_joint", "left_wrist_roll_joint", "left_wrist_pitch_joint", "left_wrist_yaw_joint",
        "right_shoulder_pitch_joint", "right_shoulder_roll_joint", "right_shoulder_yaw_joint", "right_elbow_joint",
        "right_wrist_roll_joint", "right_wrist_pitch_joint", "right_wrist_yaw_joint"]
GROUPS = {"left_leg": (0, 6), "right_leg": (6, 12), "waist": (12, 15), "left_arm": (15, 22), "right_arm": (22, 29)}


def jl(path):
    return [json.loads(x) for x in Path(path).read_text().splitlines() if x.strip()]


def wl(path, rows):
    Path(path).write_text("".join(json.dumps(r) + "\n" for r in rows))


def build(task):
    src, limit = TASKS[task]
    info = json.loads((src / "meta/info.json").read_text())
    names = info["features"]["observation.state"]["names"]
    if task in CONFIG:
        idx = list(range(31))
    else:
        idx = [names.index(n) for n in BODY]
        for g, (a, b) in GROUPS.items():
            assert idx[a:b] == list(range(idx[a], idx[a] + (b - a))), (task, g, idx[a:b])
        for key, dim in (("action.motion_token", 64), ("observation.inspire_left_hand_state", 6),
                         ("observation.inspire_right_hand_state", 6), ("observation.projected_gravity", 3),
                         ("observation.images.ego_view", None), ("observation.images.chest", None)):
            assert key in info["features"], (task, key)
            if dim:
                assert info["features"][key]["shape"] == [dim], (task, key)
    episodes = jl(src / "meta/episodes.jsonl")
    stats = jl(src / "meta/episodes_stats.jsonl")
    n = len(episodes) if limit is None else limit
    episodes, stats = episodes[:n], stats[:n]
    assert [e["episode_index"] for e in episodes] == list(range(n)) == [s["episode_index"] for s in stats]
    frames = sum(int(e["length"]) for e in episodes)
    dst = OUT / task
    if dst.exists():
        raise SystemExit(f"{dst} exists")
    (dst / "meta").mkdir(parents=True)
    vkeys = [k for k, f in info["features"].items() if f.get("dtype") == "video"]
    for e in episodes:
        i = e["episode_index"]
        chunk = i // int(info["chunks_size"])
        rel = info["data_path"].format(episode_chunk=chunk, episode_index=i)
        (dst / rel).parent.mkdir(parents=True, exist_ok=True)
        os.symlink(src / rel, dst / rel)
        for k in vkeys:
            relv = info["video_path"].format(episode_chunk=chunk, video_key=k, episode_index=i)
            (dst / relv).parent.mkdir(parents=True, exist_ok=True)
            os.symlink(src / relv, dst / relv)
    if task in CONFIG:
        # stats.py computes stats for every low-dim feature in info.json and cannot handle e.g. observation.tactile
        # [10,24,24]; keep only what the H2 modality uses (parquet columns stay untouched behind the symlinks).
        keep = {"observation.state", "observation.projected_gravity", "action.body_target", "action.hands",
                "timestamp", "frame_index", "episode_index", "index", "task_index"}
        keep |= {f"observation.images.{c}" for c in CAMS[task]}
        keep |= {"action.planner"} if task in PLANNER else set()
        info["features"] = {k: v for k, v in info["features"].items() if k in keep}
        vkeys = [k for k in vkeys if k in keep]
        stats = [{**s, "stats": {k: v for k, v in s["stats"].items() if k in keep}} if "stats" in s else s
                 for s in stats]
    info.update(total_episodes=n, total_frames=frames, total_videos=n * len(vkeys),
                total_chunks=(n - 1) // int(info["chunks_size"]) + 1, splits={"train": f"0:{n}"})
    (dst / "meta/info.json").write_text(json.dumps(info, indent=2))
    wl(dst / "meta/episodes.jsonl", episodes)
    wl(dst / "meta/episodes_stats.jsonl", stats)
    (dst / "meta/tasks.jsonl").write_text((src / "meta/tasks.jsonl").read_text())
    if task in CONFIG:
        modality = h2_modality(info, CAMS[task])
        if task in PLANNER:
            assert info["features"]["action.planner"]["shape"] == [7]
            modality["action"]["planner"] = {"original_key": "action.planner", "start": 0, "end": 7}
    else:
        state = {g: {"original_key": "observation.state", "start": idx[a], "end": idx[a] + (b - a)}
                 for g, (a, b) in GROUPS.items()}
        state.update(left_hand={"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
                     right_hand={"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6},
                     projected_gravity={"original_key": "observation.projected_gravity", "start": 0, "end": 3})
        modality = {"state": state,
                    "action": {"motion_token": {"original_key": "action.motion_token", "start": 0, "end": 64},
                               "left_hand_joints": {"original_key": "observation.inspire_left_hand_state", "start": 0, "end": 6},
                               "right_hand_joints": {"original_key": "observation.inspire_right_hand_state", "start": 0, "end": 6}},
                    "video": {"ego_view": {"original_key": "observation.images.ego_view"},
                              "chest": {"original_key": "observation.images.chest"}},
                    "annotation": {"human.task_description": {"original_key": "task_index"}}}
    (dst / "meta/modality.json").write_text(json.dumps(modality, indent=2))
    py = sys.executable
    src_g = G / "HumanWAM/third_party/Isaac-GR00T"
    env = dict(os.environ, PYTHONPATH=f"{src_g}:{H}")
    done = None
    for tag in ("NEW_EMBODIMENT", "new_embodiment"):
        r = subprocess.run([py, str(src_g / "gr00t/data/stats.py"), "--dataset-path", str(dst), "--embodiment-tag", tag,
                            "--modality-config-path", str(H / CONFIG.get(task, "g1_feedback76_h16.py"))], env=env, cwd=src_g,
                           capture_output=True, text=True)
        if r.returncode == 0:
            done = tag
            break
        last = r.stdout[-1500:] + r.stderr[-1500:]
    if done is None:
        raise SystemExit(f"stats failed for {task}:\n{last}")
    assert (dst / "meta/stats.json").is_file()
    receipt = dict(task=task, source=str(src), episodes=n, frames=frames, video_keys=vkeys,
                   state_dim_raw=len(names), body_index=idx, stats_tag=done,
                   modality_sha256=hashlib.sha256((dst / "meta/modality.json").read_bytes()).hexdigest(),
                   stats_sha256=hashlib.sha256((dst / "meta/stats.json").read_bytes()).hexdigest())
    (dst / "hb-dataset-receipt.json").write_text(json.dumps(receipt, indent=2) + "\n")
    print("DATASET_OK", json.dumps(receipt)[:400], flush=True)


for t in sys.argv[1:]:
    build(t)