File size: 21,204 Bytes
c42c446
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
"""Retarget LAFAN1 G1 motions to the Berkeley Lite humanoid.

Two-step pipeline per clip:

* **Step 1** — direct joint copy ``lite_q = sign * g1_q + offset`` using the
  static :data:`common.G1_TO_LITE` table. Adds a constant pelvis-z shift so
  Lite's feet stand on the ground.
* **Step 2** — per-frame ``mink`` IK that refines step 1 to match G1's
  pelvis-local EE poses (feet + hands, position + orientation). The
  per-DOF posture cost biases the solution toward step 1 so the refinement
  is a *tweak*, not a rearrangement.

Output is a HuggingFace LeRobotDataset written at the repository root.

Usage:
    uv run scripts/retarget.py
    uv run scripts/retarget.py --clip 'walk1_subject1'
    uv run scripts/retarget.py --validate-only
    uv run scripts/retarget.py --workers 8        # parallel across clips
    uv run scripts/retarget.py --workers -1       # use all CPU cores
"""

import os
import re
import shutil
import sys
from concurrent.futures import ProcessPoolExecutor
from pathlib import Path

import mujoco
import numpy as np
import tyro
from tqdm import tqdm

sys.path.insert(0, str(Path(__file__).resolve().parent))
from common import (  # noqa: E402
    FPS,
    G1_FOOT_BODIES,
    G1_HAND_BODIES,
    G1_LAFAN_JOINT_NAMES,
    G1_TO_LITE,
    LITE_DATASET_REPO_ID,
    LITE_FOOT_BODIES,
    LITE_HAND_BODIES,
    LITE_TASK_NAME,
    angular_velocity_from_quat,
    body_id,
    dataset_features,
    finite_diff,
    joint_qpos_addr,
    lite_joint_names,
    load_g1_model,
    load_lafan_csv,
    load_lite_model,
)

REPO_ROOT: Path = Path(__file__).resolve().parent.parent
LAFAN_ROOT: Path = REPO_ROOT / ".cache" / "lafan1_g1"
BUILD_ROOT: Path = REPO_ROOT / ".lerobot_build"
# LeRobotDataset.create refuses an existing destination, so the writer drops
# its meta/ + data/ tree into BUILD_ROOT and we move it up to REPO_ROOT on
# finalize.

# Validation frame stride for the EE-error summary.
SAMPLE_STRIDE: int = 50


# ── Step 1: direct copy with sign + offset ────────────────────────────────────


def _build_remap_indices(
    lite_model: mujoco.MjModel,
    lite_joint_addrs: np.ndarray,
) -> tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
    """Parallel arrays ``(lite_col, g1_col, sign, offset)`` over mapped joints."""
    addr_to_col = {int(a): i for i, a in enumerate(lite_joint_addrs.tolist())}
    lite_cols, g1_cols, signs, offsets = [], [], [], []
    for g1_col, g1_name in enumerate(G1_LAFAN_JOINT_NAMES):
        entry = G1_TO_LITE.get(g1_name)
        if entry is None:
            continue
        lite_name, sign, offset = entry
        lite_cols.append(addr_to_col[joint_qpos_addr(lite_model, lite_name)])
        g1_cols.append(g1_col)
        signs.append(float(sign))
        offsets.append(float(offset))
    return (
        np.asarray(lite_cols, dtype=np.int32),
        np.asarray(g1_cols, dtype=np.int32),
        np.asarray(signs, dtype=np.float32),
        np.asarray(offsets, dtype=np.float32),
    )


def _pelvis_z_offset(g1_model: mujoco.MjModel, lite_model: mujoco.MjModel) -> float:
    """Difference in standing leg length between Lite and G1 (welded pelvis).

    Both robots default with the pelvis at z=0 and feet hanging at negative z,
    so ``-min(foot_z)`` is each robot's standing leg length; the difference
    is the z-shift to apply to LAFAN1's base trajectory so Lite stays grounded.
    """

    def _leg_length(model: mujoco.MjModel, foot_names: tuple[str, str]) -> float:
        data = mujoco.MjData(model)
        mujoco.mj_kinematics(model, data)
        return -min(float(data.xpos[body_id(model, n), 2]) for n in foot_names)

    return _leg_length(lite_model, LITE_FOOT_BODIES) - _leg_length(g1_model, G1_FOOT_BODIES)


def step1_direct_remap(
    motion: dict[str, np.ndarray],
    lite_joint_addrs: np.ndarray,
    lite_model: mujoco.MjModel,
    z_offset: float,
) -> dict[str, np.ndarray]:
    """Apply ``lite_q = sign * g1_q + offset`` to every mapped joint.

    Returns ``base_pos`` (with the pelvis-z shift), ``base_quat`` (WXYZ,
    unchanged from the source), and a ``(T, 74)`` ``joint_pos`` in Lite MJCF
    order. Joints with no G1 source (neck, fingers, ankle_yaw) stay at 0.
    """
    lite_cols, g1_cols, signs, offsets = _build_remap_indices(lite_model, lite_joint_addrs)
    frames = motion["base_pos"].shape[0]
    joint_pos = np.zeros((frames, lite_joint_addrs.shape[0]), dtype=np.float32)
    joint_pos[:, lite_cols] = signs * motion["g1_joint_pos"][:, g1_cols] + offsets

    base_pos = motion["base_pos"].astype(np.float32, copy=True)
    base_pos[:, 2] += z_offset
    return {
        "base_pos": base_pos,
        "base_quat": motion["base_quat_wxyz"].astype(np.float32),
        "joint_pos": joint_pos,
    }


# ── Step 2: per-frame IK refinement ───────────────────────────────────────────

_LIMB_TOKENS: tuple[str, ...] = ("hip", "knee", "ankle", "shoulder", "elbow", "wrist")
_TRUNK_TOKENS: tuple[str, ...] = ("waist",)


def _posture_cost_vector(lite_model: mujoco.MjModel) -> np.ndarray:
    """Per-DOF posture cost — three tiers so step 2 only *tweaks* step 1.

    The PostureTask pulls each joint toward step 1 with cost ``c``; the EE
    FrameTasks pull joints away with cost 1.0 + 1.0 (position + orientation).
    Costs:

    * ``1e3`` — locked. Neck, fingers, and ankle_yaw have no G1 source, so
      we keep them at step 1's zero.
    * ``10.0`` — stiff but adjustable. Waist rotates the torso (and head),
      so a high cost prevents the IK from twisting the whole upper body to
      satisfy hand targets; arm joints are recruited first.
    * ``1.0`` — same magnitude as the EE tasks. Arms + legs get a balanced
      trade-off; per-frame corrections come out small.
    """
    cost = np.full(lite_model.nv, 1e3, dtype=np.float64)
    for jid in range(lite_model.njnt):
        name = mujoco.mj_id2name(lite_model, mujoco.mjtObj.mjOBJ_JOINT, jid)
        if not name:
            continue
        dof = int(lite_model.jnt_dofadr[jid])
        if any(tok in name for tok in _LIMB_TOKENS):
            cost[dof] = 1.0
        elif any(tok in name for tok in _TRUNK_TOKENS):
            cost[dof] = 10.0
    return cost


def _rest_frame_conversions(
    g1_model: mujoco.MjModel,
    lite_model: mujoco.MjModel,
) -> dict[str, np.ndarray]:
    """Per-body constant ``R_conv`` such that ``R_target = R_g1_actual @ R_conv``.

    Derivation: we want Lite's world-frame motion delta (relative to its
    matched rest) to equal G1's world-frame motion delta (relative to G1
    rest). Solving for the target orientation gives
    ``R_target = R_g1_actual @ (R_g1_rest^-1 @ R_lite_matched_rest)``.
    """
    g1_data = mujoco.MjData(g1_model)
    mujoco.mj_kinematics(g1_model, g1_data)
    lite_data = mujoco.MjData(lite_model)
    for _, (lite_name, _, offset) in G1_TO_LITE.items():
        lite_data.qpos[joint_qpos_addr(lite_model, lite_name)] = offset
    mujoco.mj_kinematics(lite_model, lite_data)

    out: dict[str, np.ndarray] = {}
    for lite_name, g1_name in zip(
        (*LITE_FOOT_BODIES, *LITE_HAND_BODIES),
        (*G1_FOOT_BODIES, *G1_HAND_BODIES),
        strict=True,
    ):
        R_g1 = g1_data.xmat[body_id(g1_model, g1_name)].reshape(3, 3)
        R_lite = lite_data.xmat[body_id(lite_model, lite_name)].reshape(3, 3)
        out[lite_name] = R_g1.T @ R_lite
    return out


def step2_ik_refine(
    motion: dict[str, np.ndarray],
    step1_joint_pos: np.ndarray,
    g1_model: mujoco.MjModel,
    lite_model: mujoco.MjModel,
    lite_joint_addrs: np.ndarray,
    iters: int = 15,
    show_progress: bool = True,
) -> np.ndarray:
    """Refine step-1 joint positions to match G1's pelvis-local EE poses.

    Per-frame IK is independent — each frame seeds from step 1 and converges
    on its own. ``show_progress`` controls the inner tqdm bar; worker
    processes set it to False so their bars don't interleave in the terminal.
    """
    import mink

    g1_data = mujoco.MjData(g1_model)
    g1_addrs = np.asarray(
        [joint_qpos_addr(g1_model, n) for n in G1_LAFAN_JOINT_NAMES], dtype=np.int32
    )
    R_conv = _rest_frame_conversions(g1_model, lite_model)

    configuration = mink.Configuration(lite_model)
    foot_tasks = [
        mink.FrameTask(name, "body", position_cost=1.0, orientation_cost=1.0, lm_damping=1.0)
        for name in LITE_FOOT_BODIES
    ]
    hand_tasks = [
        mink.FrameTask(name, "body", position_cost=1.0, orientation_cost=1.0, lm_damping=1.0)
        for name in LITE_HAND_BODIES
    ]
    posture_task = mink.PostureTask(lite_model, cost=_posture_cost_vector(lite_model))
    all_tasks = [*foot_tasks, *hand_tasks, posture_task]
    limits = [mink.ConfigurationLimit(lite_model)]

    ee_pairs = tuple(zip(
        (*LITE_FOOT_BODIES, *LITE_HAND_BODIES),
        (*G1_FOOT_BODIES, *G1_HAND_BODIES),
        strict=True,
    ))
    out = step1_joint_pos.copy()
    seed_qpos = np.zeros(lite_model.nq, dtype=np.float64)
    frames = step1_joint_pos.shape[0]
    frame_iter = tqdm(range(frames), desc="    IK", leave=False, unit="frame") if show_progress else range(frames)

    for t in frame_iter:
        g1_data.qpos[g1_addrs] = motion["g1_joint_pos"][t]
        mujoco.mj_kinematics(g1_model, g1_data)
        for task, (lite_name, g1_name) in zip([*foot_tasks, *hand_tasks], ee_pairs, strict=True):
            bid = body_id(g1_model, g1_name)
            mat = np.eye(4)
            mat[:3, :3] = g1_data.xmat[bid].reshape(3, 3) @ R_conv[lite_name]
            mat[:3, 3] = g1_data.xpos[bid]
            task.set_target(mink.SE3.from_matrix(mat))

        seed_qpos[lite_joint_addrs] = step1_joint_pos[t]
        configuration.q[:] = seed_qpos
        posture_task.set_target(seed_qpos.copy())
        for _ in range(iters):
            vel = mink.solve_ik(
                configuration, all_tasks, 1.0, solver="daqp", damping=1e-1, limits=limits,
            )
            configuration.integrate_inplace(vel, 1.0)
        out[t] = configuration.q[lite_joint_addrs]

    return out


# ── Validation ────────────────────────────────────────────────────────────────


def _rotation_angle_error(R_a: np.ndarray, R_b: np.ndarray) -> float:
    cos_theta = np.clip((np.trace(R_a @ R_b.T) - 1.0) * 0.5, -1.0, 1.0)
    return float(np.arccos(cos_theta))


def validate_ee_tracking(
    motion: dict[str, np.ndarray],
    lite_joint_pos: np.ndarray,
    g1_model: mujoco.MjModel,
    lite_model: mujoco.MjModel,
    lite_joint_addrs: np.ndarray,
) -> None:
    """Print per-EE position + rotation error vs. G1 at every SAMPLE_STRIDE frames.

    Rotation error is measured as the angle of "motion delta from matched
    rest" — each robot's EE rotation relative to its own matched rest.
    Lite's matched rest is step 1 applied at G1 ``q = 0`` (arms at offsets).
    """
    g1_data = mujoco.MjData(g1_model)
    lite_data = mujoco.MjData(lite_model)
    g1_addrs = np.asarray(
        [joint_qpos_addr(g1_model, n) for n in G1_LAFAN_JOINT_NAMES], dtype=np.int32
    )

    mujoco.mj_kinematics(g1_model, g1_data)
    pairs = tuple(zip(
        (*LITE_FOOT_BODIES, *LITE_HAND_BODIES),
        (*G1_FOOT_BODIES, *G1_HAND_BODIES),
        strict=True,
    ))
    rest_g1 = {g: g1_data.xmat[body_id(g1_model, g)].reshape(3, 3).copy() for _, g in pairs}

    matched_rest = np.zeros(lite_model.nq, dtype=np.float64)
    for _, (lite_name, _, offset) in G1_TO_LITE.items():
        matched_rest[joint_qpos_addr(lite_model, lite_name)] = offset
    lite_data.qpos[:] = matched_rest
    mujoco.mj_kinematics(lite_model, lite_data)
    rest_lite = {l: lite_data.xmat[body_id(lite_model, l)].reshape(3, 3).copy() for l, _ in pairs}

    frames = lite_joint_pos.shape[0]
    indices = list(range(0, frames, SAMPLE_STRIDE))
    stats: dict[str, dict[str, list[float]]] = {l: {"pos": [], "rot": []} for l, _ in pairs}

    for t in indices:
        lite_data.qpos[lite_joint_addrs] = lite_joint_pos[t]
        mujoco.mj_kinematics(lite_model, lite_data)
        g1_data.qpos[g1_addrs] = motion["g1_joint_pos"][t]
        mujoco.mj_kinematics(g1_model, g1_data)
        for lname, gname in pairs:
            lbid, gbid = body_id(lite_model, lname), body_id(g1_model, gname)
            stats[lname]["pos"].append(float(np.linalg.norm(lite_data.xpos[lbid] - g1_data.xpos[gbid])))
            R_l = lite_data.xmat[lbid].reshape(3, 3) @ rest_lite[lname].T
            R_g = g1_data.xmat[gbid].reshape(3, 3) @ rest_g1[gname].T
            stats[lname]["rot"].append(_rotation_angle_error(R_l, R_g))

    print(f"\nEE tracking error across {len(indices)} frames (stride={SAMPLE_STRIDE}, total={frames}):")
    print(f"  {'body':<14s} {'pos mean':>9s} {'pos max':>9s}   {'rot mean':>9s} {'rot max':>9s}")
    for lname, _ in pairs:
        pos = np.asarray(stats[lname]["pos"])
        rot = np.asarray(stats[lname]["rot"])
        print(
            f"  {lname:<14s} {pos.mean():>8.3f}m {pos.max():>8.3f}m   "
            f"{np.degrees(rot.mean()):>7.2f}° {np.degrees(rot.max()):>7.2f}°"
        )


# ── Per-clip pipeline + LeRobotDataset writer ─────────────────────────────────


def _frame_records(
    base_pos: np.ndarray,
    base_quat: np.ndarray,
    joint_pos: np.ndarray,
) -> dict[str, np.ndarray]:
    """Compose the six dataset-feature arrays from a per-clip trajectory."""
    base_pos = base_pos.astype(np.float32, copy=False)
    base_quat = base_quat.astype(np.float32, copy=False)
    joint_pos = joint_pos.astype(np.float32, copy=False)
    return {
        "base_pos": base_pos,
        "base_quat": base_quat,
        "base_lin_vel": finite_diff(base_pos, FPS),
        "base_ang_vel": angular_velocity_from_quat(base_quat, FPS).astype(np.float32),
        "joint_pos": joint_pos,
        "joint_vel": finite_diff(joint_pos, FPS),
    }


# ── Multiprocess worker (across-clip parallelism) ─────────────────────────────

_WORKER_STATE: dict[str, object] = {}


def _worker_init() -> None:
    """ProcessPoolExecutor initializer: compile MuJoCo models once per worker."""
    g1_model = load_g1_model(LAFAN_ROOT)
    lite_model = load_lite_model()
    addrs = np.asarray(
        [joint_qpos_addr(lite_model, n) for n in lite_joint_names(lite_model)], dtype=np.int32
    )
    _WORKER_STATE.update(
        g1_model=g1_model,
        lite_model=lite_model,
        lite_joint_addrs=addrs,
        z_offset=_pelvis_z_offset(g1_model, lite_model),
    )


def _worker_retarget(args: tuple[str, bool, int]) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
    """Retarget a single clip; returns ``(base_pos, base_quat, joint_pos)``."""
    csv_path_str, do_ik, ik_iters = args
    g1_model: mujoco.MjModel = _WORKER_STATE["g1_model"]  # type: ignore[assignment]
    lite_model: mujoco.MjModel = _WORKER_STATE["lite_model"]  # type: ignore[assignment]
    lite_joint_addrs: np.ndarray = _WORKER_STATE["lite_joint_addrs"]  # type: ignore[assignment]
    z_offset: float = _WORKER_STATE["z_offset"]  # type: ignore[assignment]

    motion = load_lafan_csv(Path(csv_path_str))
    step1 = step1_direct_remap(motion, lite_joint_addrs, lite_model, z_offset)
    joint_pos = step1["joint_pos"]
    if do_ik:
        joint_pos = step2_ik_refine(
            motion, joint_pos, g1_model, lite_model, lite_joint_addrs,
            iters=ik_iters, show_progress=False,
        )
    return step1["base_pos"], step1["base_quat"], joint_pos


# ── CLI ───────────────────────────────────────────────────────────────────────


def main(
    repo_id: str = LITE_DATASET_REPO_ID,
    clip: str | None = None,
    ik: bool = True,
    ik_iters: int = 15,
    workers: int = 1,
    validate_only: bool = False,
) -> None:
    """Retarget LAFAN1 G1 clips to Lite and write a LeRobotDataset.

    Args:
        repo_id: HF dataset repo id, recorded in dataset metadata.
        clip: Optional regex to retarget only matching CSVs.
        ik: If True, run step 2 IK to refine step 1. If False, output step 1 only.
        ik_iters: Newton-step iterations per frame in step 2.
        workers: Worker processes for across-clip parallelism. ``1`` (default)
            keeps the sequential path with per-frame tqdm. ``-1`` uses every
            CPU core. ``>1`` spawns a ``ProcessPoolExecutor`` and suppresses
            inner tqdm bars to keep the terminal readable.
        validate_only: Run on the first matching clip and stop without writing
            the dataset. Prints the step-1 (and step-2 if ``ik=True``) EE error
            table.
    """
    if workers == -1:
        workers = os.cpu_count() or 1
    # Each save_episode runs an HFDataset.map pass that prints its own bar
    # — 218 of those interleave badly with our outer clip bar.
    os.environ.setdefault("HF_DATASETS_DISABLE_PROGRESS_BARS", "1")

    csvs = sorted((LAFAN_ROOT / "g1").glob("*.csv"))
    if clip is not None:
        csvs = [p for p in csvs if re.search(clip, p.stem)]
    if not csvs:
        raise SystemExit(
            f"No CSVs to retarget under {LAFAN_ROOT / 'g1'} (clip={clip!r}). "
            f"Run scripts/download_lafan.py first."
        )

    print("Loading models …")
    g1_model = load_g1_model(LAFAN_ROOT)
    lite_model = load_lite_model()
    lite_jnames = lite_joint_names(lite_model)
    lite_joint_addrs = np.asarray(
        [joint_qpos_addr(lite_model, n) for n in lite_jnames], dtype=np.int32
    )
    z_offset = _pelvis_z_offset(g1_model, lite_model)
    flipped = sum(1 for _, s, _ in G1_TO_LITE.values() if s < 0)
    nonzero = sum(1 for _, _, off in G1_TO_LITE.values() if abs(off) > 1e-6)
    print(f"  G1 nq={g1_model.nq}, Lite nq={lite_model.nq}, joints={len(lite_jnames)}")
    print(
        f"  {len(G1_TO_LITE)} joint pairs, {flipped} sign flips, "
        f"{nonzero} nonzero offsets, pelvis z-offset={z_offset * 1000:.2f} mm"
    )

    if validate_only:
        motion = load_lafan_csv(csvs[0])
        step1 = step1_direct_remap(motion, lite_joint_addrs, lite_model, z_offset)
        print(f"\nClip: {csvs[0].name} ({step1['joint_pos'].shape[0]} frames)")
        print("\n=== Step 1 (direct copy with sign + offset) ===")
        validate_ee_tracking(motion, step1["joint_pos"], g1_model, lite_model, lite_joint_addrs)
        if ik:
            step2 = step2_ik_refine(
                motion, step1["joint_pos"], g1_model, lite_model, lite_joint_addrs,
                iters=ik_iters,
            )
            print("\n=== Step 1 + Step 2 (per-frame IK refinement) ===")
            validate_ee_tracking(motion, step2, g1_model, lite_model, lite_joint_addrs)
        return

    from lerobot.datasets import LeRobotDataset  # deferred: heavy import

    if BUILD_ROOT.exists():
        shutil.rmtree(BUILD_ROOT)
    dataset = LeRobotDataset.create(
        repo_id=repo_id,
        fps=FPS,
        features=dataset_features(joint_count=len(lite_jnames)),
        root=BUILD_ROOT,
        robot_type="lite",
        use_videos=False,
    )

    def _write_clip(base_pos: np.ndarray, base_quat: np.ndarray, joint_pos: np.ndarray) -> None:
        records = _frame_records(base_pos, base_quat, joint_pos)
        for t in range(records["base_pos"].shape[0]):
            dataset.add_frame({"task": LITE_TASK_NAME, **{k: v[t] for k, v in records.items()}})
        dataset.save_episode()

    if workers > 1:
        args_list = [(str(p), bool(ik), int(ik_iters)) for p in csvs]
        with ProcessPoolExecutor(max_workers=workers, initializer=_worker_init) as executor:
            for base_pos, base_quat, joint_pos in tqdm(
                executor.map(_worker_retarget, args_list, chunksize=1),
                total=len(csvs), desc=f"Clips (workers={workers})", unit="clip",
            ):
                _write_clip(base_pos, base_quat, joint_pos)
    else:
        for csv_path in tqdm(csvs, desc="Clips", unit="clip"):
            motion = load_lafan_csv(csv_path)
            step1 = step1_direct_remap(motion, lite_joint_addrs, lite_model, z_offset)
            joint_pos = step1["joint_pos"]
            if ik:
                joint_pos = step2_ik_refine(
                    motion, joint_pos, g1_model, lite_model, lite_joint_addrs, iters=ik_iters,
                )
            _write_clip(step1["base_pos"], step1["base_quat"], joint_pos)

    dataset.finalize()
    for sub in ("meta", "data"):
        src = BUILD_ROOT / sub
        if src.exists():
            dst = REPO_ROOT / sub
            if dst.exists():
                shutil.rmtree(dst)
            shutil.move(str(src), str(dst))
    shutil.rmtree(BUILD_ROOT, ignore_errors=True)
    print(f"\nWrote dataset to {REPO_ROOT} ({len(csvs)} episodes)")


if __name__ == "__main__":
    tyro.cli(main)