Datasets:

ckwolfe
/

benchmarks-viz-tiles

"""Standardized-framing video renderer — single source of truth.
Produces one mp4 per cell, all identical camera + resolution + codec so
the paper's tile grid lines up frame-for-frame. Bypasses each method's
native renderer (which differ in FOV, resolution, and what's onscreen)
by re-rendering a Spider-preprocessed MuJoCo scene with BENCH_CAMERA.
Inputs (choose highest-fidelity available):
  1. trajectory_mjwp.npz       — Spider's MJWP sampler output (closed-ish loop)
  2. trajectory_ikrollout.npz  — IK rollout of the kinematic trajectory
  3. trajectory_kinematic.npz  — pure MANO→hand IK (open-loop reference)
Output: outputs/videos_std/<method>_<hand>_<dataset>_<traj>_seed<N>_<ws>.mp4
All mp4s: 720×480, 30 fps, h264 crf 20, yuv420p.
Camera: BENCH_CAMERA["front"] pos=[0,-1.6,2.2] lookat=[0,-0.1,1.2] fov=30.
Run outside any container — needs only mujoco + numpy + ffmpeg.
"""
from __future__ import annotations

import argparse
import os
import sys
import subprocess
from pathlib import Path
from typing import Optional

import numpy as np

REPO = Path("/home/azureuser/ckwolfe/bench-dm-fix")
sys.path.insert(0, str(REPO))
sys.path.insert(0, str(REPO / "shared"))

OAKINK_TRAJS = [
    "lift_board", "pick_spoon_bowl", "pour_tube", "stir_beaker",
    "uncap_alcohol_burner", "unplug", "wipe_board",
]
HANDS = ["allegro", "inspire", "schunk", "xhand"]

PROCESSED = Path("/mnt/external/data/oakink_v2/processed/oakink")
OUT_DIR = Path("/mnt/external/outputs/videos_std")
MT_QPOS_DIR = Path("/mnt/external/outputs/mt_qpos")
OUT_DIR.mkdir(parents=True, exist_ok=True)


def resolve_cell(hand: str, traj: str):
    """Return (scene_xml, rollout_npz, source_tag) or (None, None, reason).
    Prefers `trajectory_kinematic.npz` over mjwp because kinematic is a
    clean 2-D (T, N) qpos array across all cells (~259 frames each),
    giving uniform rollout length in the tile grid. mjwp is a
    (T_plan, K_knots, N_dof) receding-horizon tensor and would need
    per-method unrolling logic to visualize.
    """
    cell = PROCESSED / hand / "bimanual" / traj
    scene = cell / "scene.xml"
    if not scene.is_file():
        return None, None, "no scene.xml"
    for nm, tag in [
        ("trajectory_kinematic.npz", "kinematic"),
        ("trajectory_ikrollout.npz", "ikrollout"),
        ("trajectory_mjwp.npz", "mjwp"),
    ]:
        p = cell / "0" / nm
        if p.is_file():
            return scene, p, tag
    return None, None, "no trajectory npz"


def load_qpos(npz_path: Path) -> Optional[np.ndarray]:
    """Return qpos as (T, N) float32. Common key aliases."""
    with np.load(npz_path, allow_pickle=True) as z:
        for k in ("qpos", "q", "data", "trajectory", "states"):
            if k in z.files:
                arr = z[k]
                if arr.ndim == 1 and arr.dtype == object:
                    # might be a packed dict
                    arr = arr.item() if callable(getattr(arr, "item", None)) else arr
                if isinstance(arr, dict):
                    for sub in ("qpos", "q", "data"):
                        if sub in arr:
                            arr = arr[sub]; break
                arr = np.asarray(arr)
                if arr.ndim == 2:
                    return arr.astype(np.float32)
                if arr.ndim == 3:
                    # (T, B, N) — take first batch
                    return arr[:, 0, :].astype(np.float32)
        # Fall back to first 2-D numeric array
        for k in z.files:
            arr = z[k]
            if getattr(arr, "ndim", 0) == 2:
                return arr.astype(np.float32)
    return None


def render(scene_xml: Path, qpos: np.ndarray, out_mp4: Path,
           camera_name: str = "front") -> tuple[bool, str]:
    """MuJoCo render + ffmpeg pipe. Returns (ok, note)."""
    try:
        import mujoco
    except ImportError:
        return False, "mujoco import failed"
    from bench.video_recorder import BENCH_CAMERA, BENCH_VIDEO, VideoRecorder

    cam_spec = BENCH_CAMERA[camera_name]
    w, h = BENCH_VIDEO["resolution"]
    fps = BENCH_VIDEO["fps"]

    # mujoco sometimes rejects the per-texture colorspace attr on old scene.xml.
    # Strip it as a defensive fallback.
    # Defensive XML preprocessing: (1) strip unrecognized texture colorspace
    # attr, (2) inject <visual><global offwidth=... offheight=.../></visual>
    # so the offscreen framebuffer is big enough for our 720×480 target.
    def _prep_xml(src: Path) -> Path:
        import re
        txt = src.read_text()
        txt = txt.replace(' colorspace="linear"', '').replace(' colorspace="srgb"', '')
        # Ensure exactly one <visual> block with the offscreen framebuffer
        # declared. mujoco rejects duplicate <global> siblings.
        new_global = f'<global offwidth="{w}" offheight="{h}"/>'
        if "<global" in txt:
            # rewrite any existing global (add offwidth/height if missing)
            def sub(m):
                tag = m.group(0)
                if "offwidth" in tag and "offheight" in tag:
                    return tag
                # insert attributes before the trailing "/>" or ">"
                closing = "/>" if tag.rstrip().endswith("/>") else ">"
                body = tag[:len(tag) - len(closing)].rstrip()
                return f'{body} offwidth="{w}" offheight="{h}"{closing}'
            txt = re.sub(r"<global[^>]*/?>", sub, txt, count=1)
        elif "<visual>" in txt:
            txt = txt.replace("<visual>", f"<visual>\n    {new_global}", 1)
        else:
            txt = re.sub(r"(<mujoco[^>]*>)", rf"\1\n  <visual>\n    {new_global}\n  </visual>", txt, count=1)
        # Write the patched XML in a writable sibling of the original
        # so that `../../../assets/robots/...` relative paths resolve.
        # /mnt/external is owned by root, so use /tmp/scene_patches/<depth>/
        # with a symlink tree mirroring the relative structure.
        patch_root = Path("/tmp/scene_patches")
        # Depth of scene.xml below oakink root: .../oakink/<hand>/bimanual/<task>/scene.xml
        # The xml's relative refs go up 3 levels to reach .../oakink/assets/
        hand = src.parents[2].name  # .../oakink/<hand>/bimanual/<task>/scene.xml
        task = src.parent.name
        staging = patch_root / hand / "bimanual" / task
        staging.mkdir(parents=True, exist_ok=True)
        # Mirror the assets dir via symlink so relpaths work
        assets_src = src.parents[3] / "assets"  # .../oakink/assets
        assets_link = patch_root / "assets"
        if not assets_link.exists():
            assets_link.symlink_to(assets_src)
        patched = staging / "scene.xml"
        patched.write_text(txt)
        return patched

    try:
        model = mujoco.MjModel.from_xml_path(str(_prep_xml(scene_xml)))
    except ValueError as exc:
        return False, f"mjcf load failed: {exc}"

    data = mujoco.MjData(model)
    nq = model.nq
    k = min(qpos.shape[1], nq)
    T = int(qpos.shape[0])

    try:
        renderer = mujoco.Renderer(model, height=h, width=w)
    except Exception as exc:
        return False, f"renderer init failed: {exc}"

    cam = mujoco.MjvCamera()
    pos = np.array(cam_spec["pos"], dtype=np.float64)
    look = np.array(cam_spec["lookat"], dtype=np.float64)
    cam.lookat[:] = look
    d = pos - look
    cam.distance = float(np.linalg.norm(d))
    cam.azimuth = float(np.degrees(np.arctan2(d[0], -d[1])))
    cam.elevation = float(np.degrees(np.arcsin(d[2] / (cam.distance + 1e-9))))

    out_mp4.parent.mkdir(parents=True, exist_ok=True)
    rec = VideoRecorder(out_mp4, fps=fps, size=(w, h))
    frames_written = 0
    try:
        for t in range(T):
            data.qpos[:k] = qpos[t, :k]
            mujoco.mj_forward(model, data)
            renderer.update_scene(data, camera=cam)
            rec.log_frame(renderer.render())
            frames_written += 1
        rec.close()
    except Exception as exc:
        try: rec.close()
        except: pass
        return False, f"render failed at frame {frames_written}: {exc}"
    finally:
        try: renderer.close()
        except Exception: pass

    return frames_written > 0, f"{frames_written} frames"


def _resolve_mt_cell(hand: str, traj: str, seed: int, warmstart: str):
    """Find the MT captured-qpos npz + matching scene.xml.
    Returns (scene_xml, mt_npz, run_id) or (None, None, reason).
    """
    # MT's run_id convention matches methods/maniptrans/scripts/run_eval.py.
    run_id = f"maniptrans_{hand}_oakink_v2_{traj}_bimanual_seed{seed}_{warmstart}"
    mt_npz = MT_QPOS_DIR / f"{run_id}.npz"
    if not mt_npz.is_file():
        return None, None, f"no MT qpos at {mt_npz.name}"
    scene = PROCESSED / hand / "bimanual" / traj / "scene.xml"
    if not scene.is_file():
        return None, None, "no scene.xml"
    return scene, mt_npz, run_id


def render_mt(scene_xml: Path, mt_npz: Path, out_mp4: Path,
              camera_name: str = "front") -> tuple[bool, str]:
    """Render MT's captured DOF trajectory by name-permutation into scene.xml.
    MT dumps IG-ordered `q` plus `dof_names`. MJ scene.xml has its own
    per-joint ordering. We build a permutation `mj_qpos_adr[j] = ig_q[k]`
    where `dof_names[k] == mj_joint_names[j]`.
    """
    try:
        import mujoco
    except ImportError:
        return False, "mujoco import failed"
    from bench.video_recorder import BENCH_CAMERA, BENCH_VIDEO, VideoRecorder

    with np.load(mt_npz, allow_pickle=True) as z:
        if "q" not in z.files:
            return False, f"npz missing 'q' key (has {z.files})"
        mt_q = z["q"].astype(np.float32)  # (T, n_ig_dof)
        ig_names = list(z["dof_names"]) if "dof_names" in z.files else None

    T = int(mt_q.shape[0])
    if T == 0:
        return False, "empty mt_q"

    cam_spec = BENCH_CAMERA[camera_name]
    w, h = BENCH_VIDEO["resolution"]
    fps = BENCH_VIDEO["fps"]

    # Reuse the XML prep logic from render() via a tiny local copy.
    import re as _re
    patched = scene_xml  # placeholder; assigned below
    try:
        import subprocess  # noqa: F401  - for ffmpeg via VideoRecorder
        txt = scene_xml.read_text()
        txt = txt.replace(' colorspace="linear"', '').replace(' colorspace="srgb"', '')
        new_global = f'<global offwidth="{w}" offheight="{h}"/>'
        if "<global" in txt:
            def _sub(m):
                tag = m.group(0)
                if "offwidth" in tag and "offheight" in tag:
                    return tag
                closing = "/>" if tag.rstrip().endswith("/>") else ">"
                body = tag[:len(tag) - len(closing)].rstrip()
                return f'{body} offwidth="{w}" offheight="{h}"{closing}'
            txt = _re.sub(r"<global[^>]*/?>", _sub, txt, count=1)
        elif "<visual>" in txt:
            txt = txt.replace("<visual>", f"<visual>\n    {new_global}", 1)
        else:
            txt = _re.sub(r"(<mujoco[^>]*>)", rf"\1\n  <visual>\n    {new_global}\n  </visual>",
                          txt, count=1)
        patch_root = Path("/tmp/scene_patches")
        hand = scene_xml.parents[2].name
        task = scene_xml.parent.name
        staging = patch_root / hand / "bimanual" / task
        staging.mkdir(parents=True, exist_ok=True)
        assets_src = scene_xml.parents[3] / "assets"
        assets_link = patch_root / "assets"
        if not assets_link.exists():
            assets_link.symlink_to(assets_src)
        patched = staging / "scene.xml"
        patched.write_text(txt)
    except Exception as exc:
        return False, f"xml prep failed: {exc}"

    try:
        model = mujoco.MjModel.from_xml_path(str(patched))
    except ValueError as exc:
        return False, f"mjcf load failed: {exc}"

    # Build MJ-joint -> IG-DOF permutation by name.
    perm = None
    if ig_names is not None:
        ig_idx = {str(n): i for i, n in enumerate(ig_names)}
        perm = []
        for j in range(model.njnt):
            jn = mujoco.mj_id2name(model, mujoco.mjtObj.mjOBJ_JOINT, j)
            qadr = int(model.jnt_qposadr[j])
            jtype = int(model.jnt_type[j])
            # Skip free + ball joints (nq>1 per joint) — we don't map obj free joint
            if jtype != mujoco.mjtJoint.mjJNT_HINGE and jtype != mujoco.mjtJoint.mjJNT_SLIDE:
                continue
            if jn in ig_idx:
                perm.append((qadr, ig_idx[jn]))
        note_perm = f"perm size={len(perm)}/{model.njnt}"
    else:
        # Fallback: write first min(nq, n_ig_dof) values directly.
        k = min(model.nq, mt_q.shape[1])
        perm = [(i, i) for i in range(k)]
        note_perm = f"no dof_names; fallback linear k={k}"

    data = mujoco.MjData(model)
    try:
        renderer = mujoco.Renderer(model, height=h, width=w)
    except Exception as exc:
        return False, f"renderer init failed: {exc}"

    cam = mujoco.MjvCamera()
    pos = np.array(cam_spec["pos"], dtype=np.float64)
    look = np.array(cam_spec["lookat"], dtype=np.float64)
    cam.lookat[:] = look
    d = pos - look
    cam.distance = float(np.linalg.norm(d))
    cam.azimuth = float(np.degrees(np.arctan2(d[0], -d[1])))
    cam.elevation = float(np.degrees(np.arcsin(d[2] / (cam.distance + 1e-9))))

    out_mp4.parent.mkdir(parents=True, exist_ok=True)
    rec = VideoRecorder(out_mp4, fps=fps, size=(w, h))
    frames = 0
    try:
        for t in range(T):
            q_t = mt_q[t]
            for qadr, idx in perm:
                if idx < q_t.shape[0]:
                    data.qpos[qadr] = q_t[idx]
            mujoco.mj_forward(model, data)
            renderer.update_scene(data, camera=cam)
            rec.log_frame(renderer.render())
            frames += 1
        rec.close()
    except Exception as exc:
        try: rec.close()
        except: pass
        return False, f"render failed at frame {frames}: {exc}"
    finally:
        try: renderer.close()
        except Exception: pass

    return frames > 0, f"{frames} frames, {note_perm}"


def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("--hand", choices=HANDS, default=None)
    ap.add_argument("--traj", choices=OAKINK_TRAJS, default=None)
    ap.add_argument("--source", choices=["kinematic", "mt"], default="kinematic",
                    help="'kinematic' = Oracle/reference via Spider preprocess; "
                         "'mt' = ManipTrans captured qpos via bench_hooks.patch")
    ap.add_argument("--mt-seed", type=int, default=42,
                    help="Seed tag on MT run_id (default 42)")
    ap.add_argument("--mt-warmstart", default="default",
                    help="Warmstart tag on MT run_id (default 'default')")
    ap.add_argument("--method-label", default=None,
                    help="Override filename prefix (default: 'std' for kinematic, 'mt' for mt)")
    ap.add_argument("--force", action="store_true")
    args = ap.parse_args()

    cells = []
    for h in HANDS if args.hand is None else [args.hand]:
        for t in OAKINK_TRAJS if args.traj is None else [args.traj]:
            cells.append((h, t))

    label = args.method_label or ("mt" if args.source == "mt" else "std")
    n_ok = n_fail = n_skip = 0
    print(f"{'='*72}\nstandardized render ({args.source}) — {len(cells)} cells\n{'='*72}")
    print(f"camera:   front pos=[0,-1.6,2.2] lookat=[0,-0.1,1.2] fov=30")
    print(f"video:    720×480 @ 30fps, h264 crf=20")
    print(f"output:   {OUT_DIR}\n")

    for hand, traj in cells:
        if args.source == "mt":
            scene, npz, tag_or_rid = _resolve_mt_cell(hand, traj, args.mt_seed, args.mt_warmstart)
            if scene is None:
                print(f"  [fail] {hand}/{traj}: {tag_or_rid}")
                n_fail += 1; continue
            # Preserve seed + warmstart in output name so each MT run renders distinctly.
            out = OUT_DIR / f"{label}_{hand}_oakink_v2_{traj}_bimanual_seed{args.mt_seed}_{args.mt_warmstart}.mp4"
            if out.is_file() and not args.force:
                print(f"  [skip] {out.name} (exists)"); n_skip += 1; continue
            ok, note = render_mt(scene, npz, out)
            status = "✅pass" if ok else "❌fail"
            print(f"  [{status}] {hand}/{traj} (src=mt_qpos): {note}")
            if ok: n_ok += 1
            else: n_fail += 1
            continue

        run_id = f"{label}_{hand}_oakink_v2_{traj}_bimanual_seed0_default"
        out = OUT_DIR / f"{run_id}.mp4"
        if out.is_file() and not args.force:
            print(f"  [skip] {run_id} (exists)")
            n_skip += 1; continue

        scene, npz, tag = resolve_cell(hand, traj)
        if scene is None:
            print(f"  [fail] {hand}/{traj}: {tag}")
            n_fail += 1; continue

        qpos = load_qpos(npz)
        if qpos is None or qpos.shape[0] == 0:
            print(f"  [fail] {hand}/{traj}: no qpos in {npz.name}")
            n_fail += 1; continue

        ok, note = render(scene, qpos, out)
        status = "✅pass" if ok else "❌fail"
        print(f"  [{status}] {hand}/{traj} (src={tag}): {note}")
        if ok:
            n_ok += 1
        else:
            n_fail += 1

    print(f"\n{'='*72}\nsummary: {n_ok} rendered, {n_fail} failed, "
          f"{n_skip} skipped (already present)\n{'='*72}")
    return 0 if n_fail == 0 else 1


if __name__ == "__main__":
    sys.exit(main())

	"""Standardized-framing video renderer — single source of truth.

	Produces one mp4 per cell, all identical camera + resolution + codec so
	the paper's tile grid lines up frame-for-frame. Bypasses each method's
	native renderer (which differ in FOV, resolution, and what's onscreen)
	by re-rendering a Spider-preprocessed MuJoCo scene with BENCH_CAMERA.

	Inputs (choose highest-fidelity available):
	1. trajectory_mjwp.npz — Spider's MJWP sampler output (closed-ish loop)
	2. trajectory_ikrollout.npz — IK rollout of the kinematic trajectory
	3. trajectory_kinematic.npz — pure MANO→hand IK (open-loop reference)

	Output: outputs/videos_std/<method>_<hand>_<dataset>_<traj>_seed<N>_<ws>.mp4

	All mp4s: 720×480, 30 fps, h264 crf 20, yuv420p.
	Camera: BENCH_CAMERA["front"] pos=[0,-1.6,2.2] lookat=[0,-0.1,1.2] fov=30.

	Run outside any container — needs only mujoco + numpy + ffmpeg.
	"""
	from __future__ import annotations

	import argparse
	import os
	import sys
	import subprocess
	from pathlib import Path
	from typing import Optional

	import numpy as np

	REPO = Path("/home/azureuser/ckwolfe/bench-dm-fix")
	sys.path.insert(0, str(REPO))
	sys.path.insert(0, str(REPO / "shared"))

	OAKINK_TRAJS = [
	"lift_board", "pick_spoon_bowl", "pour_tube", "stir_beaker",
	"uncap_alcohol_burner", "unplug", "wipe_board",
	]
	HANDS = ["allegro", "inspire", "schunk", "xhand"]

	PROCESSED = Path("/mnt/external/data/oakink_v2/processed/oakink")
	OUT_DIR = Path("/mnt/external/outputs/videos_std")
	MT_QPOS_DIR = Path("/mnt/external/outputs/mt_qpos")
	OUT_DIR.mkdir(parents=True, exist_ok=True)


	def resolve_cell(hand: str, traj: str):
	"""Return (scene_xml, rollout_npz, source_tag) or (None, None, reason).

	Prefers `trajectory_kinematic.npz` over mjwp because kinematic is a
	clean 2-D (T, N) qpos array across all cells (~259 frames each),
	giving uniform rollout length in the tile grid. mjwp is a
	(T_plan, K_knots, N_dof) receding-horizon tensor and would need
	per-method unrolling logic to visualize.
	"""
	cell = PROCESSED / hand / "bimanual" / traj
	scene = cell / "scene.xml"
	if not scene.is_file():
	return None, None, "no scene.xml"
	for nm, tag in [
	("trajectory_kinematic.npz", "kinematic"),
	("trajectory_ikrollout.npz", "ikrollout"),
	("trajectory_mjwp.npz", "mjwp"),
	]:
	p = cell / "0" / nm
	if p.is_file():
	return scene, p, tag
	return None, None, "no trajectory npz"


	def load_qpos(npz_path: Path) -> Optional[np.ndarray]:
	"""Return qpos as (T, N) float32. Common key aliases."""
	with np.load(npz_path, allow_pickle=True) as z:
	for k in ("qpos", "q", "data", "trajectory", "states"):
	if k in z.files:
	arr = z[k]
	if arr.ndim == 1 and arr.dtype == object:
	# might be a packed dict
	arr = arr.item() if callable(getattr(arr, "item", None)) else arr
	if isinstance(arr, dict):
	for sub in ("qpos", "q", "data"):
	if sub in arr:
	arr = arr[sub]; break
	arr = np.asarray(arr)
	if arr.ndim == 2:
	return arr.astype(np.float32)
	if arr.ndim == 3:
	# (T, B, N) — take first batch
	return arr[:, 0, :].astype(np.float32)
	# Fall back to first 2-D numeric array
	for k in z.files:
	arr = z[k]
	if getattr(arr, "ndim", 0) == 2:
	return arr.astype(np.float32)
	return None


	def render(scene_xml: Path, qpos: np.ndarray, out_mp4: Path,
	camera_name: str = "front") -> tuple[bool, str]:
	"""MuJoCo render + ffmpeg pipe. Returns (ok, note)."""
	try:
	import mujoco
	except ImportError:
	return False, "mujoco import failed"
	from bench.video_recorder import BENCH_CAMERA, BENCH_VIDEO, VideoRecorder

	cam_spec = BENCH_CAMERA[camera_name]
	w, h = BENCH_VIDEO["resolution"]
	fps = BENCH_VIDEO["fps"]

	# mujoco sometimes rejects the per-texture colorspace attr on old scene.xml.
	# Strip it as a defensive fallback.
	# Defensive XML preprocessing: (1) strip unrecognized texture colorspace
	# attr, (2) inject <visual><global offwidth=... offheight=.../></visual>
	# so the offscreen framebuffer is big enough for our 720×480 target.
	def _prep_xml(src: Path) -> Path:
	import re
	txt = src.read_text()
	txt = txt.replace(' colorspace="linear"', '').replace(' colorspace="srgb"', '')
	# Ensure exactly one <visual> block with the offscreen framebuffer
	# declared. mujoco rejects duplicate <global> siblings.
	new_global = f'<global offwidth="{w}" offheight="{h}"/>'
	if "<global" in txt:
	# rewrite any existing global (add offwidth/height if missing)
	def sub(m):
	tag = m.group(0)
	if "offwidth" in tag and "offheight" in tag:
	return tag
	# insert attributes before the trailing "/>" or ">"
	closing = "/>" if tag.rstrip().endswith("/>") else ">"
	body = tag[:len(tag) - len(closing)].rstrip()
	return f'{body} offwidth="{w}" offheight="{h}"{closing}'
	txt = re.sub(r"<global[^>]*/?>", sub, txt, count=1)
	elif "<visual>" in txt:
	txt = txt.replace("<visual>", f"<visual>\n {new_global}", 1)
	else:
	txt = re.sub(r"(<mujoco[^>]*>)", rf"\1\n <visual>\n {new_global}\n </visual>", txt, count=1)
	# Write the patched XML in a writable sibling of the original
	# so that `../../../assets/robots/...` relative paths resolve.
	# /mnt/external is owned by root, so use /tmp/scene_patches/<depth>/
	# with a symlink tree mirroring the relative structure.
	patch_root = Path("/tmp/scene_patches")
	# Depth of scene.xml below oakink root: .../oakink/<hand>/bimanual/<task>/scene.xml
	# The xml's relative refs go up 3 levels to reach .../oakink/assets/
	hand = src.parents[2].name # .../oakink/<hand>/bimanual/<task>/scene.xml
	task = src.parent.name
	staging = patch_root / hand / "bimanual" / task
	staging.mkdir(parents=True, exist_ok=True)
	# Mirror the assets dir via symlink so relpaths work
	assets_src = src.parents[3] / "assets" # .../oakink/assets
	assets_link = patch_root / "assets"
	if not assets_link.exists():
	assets_link.symlink_to(assets_src)
	patched = staging / "scene.xml"
	patched.write_text(txt)
	return patched

	try:
	model = mujoco.MjModel.from_xml_path(str(_prep_xml(scene_xml)))
	except ValueError as exc:
	return False, f"mjcf load failed: {exc}"

	data = mujoco.MjData(model)
	nq = model.nq
	k = min(qpos.shape[1], nq)
	T = int(qpos.shape[0])

	try:
	renderer = mujoco.Renderer(model, height=h, width=w)
	except Exception as exc:
	return False, f"renderer init failed: {exc}"

	cam = mujoco.MjvCamera()
	pos = np.array(cam_spec["pos"], dtype=np.float64)
	look = np.array(cam_spec["lookat"], dtype=np.float64)
	cam.lookat[:] = look
	d = pos - look
	cam.distance = float(np.linalg.norm(d))
	cam.azimuth = float(np.degrees(np.arctan2(d[0], -d[1])))
	cam.elevation = float(np.degrees(np.arcsin(d[2] / (cam.distance + 1e-9))))

	out_mp4.parent.mkdir(parents=True, exist_ok=True)
	rec = VideoRecorder(out_mp4, fps=fps, size=(w, h))
	frames_written = 0
	try:
	for t in range(T):
	data.qpos[:k] = qpos[t, :k]
	mujoco.mj_forward(model, data)
	renderer.update_scene(data, camera=cam)
	rec.log_frame(renderer.render())
	frames_written += 1
	rec.close()
	except Exception as exc:
	try: rec.close()
	except: pass
	return False, f"render failed at frame {frames_written}: {exc}"
	finally:
	try: renderer.close()
	except Exception: pass

	return frames_written > 0, f"{frames_written} frames"


	def _resolve_mt_cell(hand: str, traj: str, seed: int, warmstart: str):
	"""Find the MT captured-qpos npz + matching scene.xml.

	Returns (scene_xml, mt_npz, run_id) or (None, None, reason).
	"""
	# MT's run_id convention matches methods/maniptrans/scripts/run_eval.py.
	run_id = f"maniptrans_{hand}_oakink_v2_{traj}_bimanual_seed{seed}_{warmstart}"
	mt_npz = MT_QPOS_DIR / f"{run_id}.npz"
	if not mt_npz.is_file():
	return None, None, f"no MT qpos at {mt_npz.name}"
	scene = PROCESSED / hand / "bimanual" / traj / "scene.xml"
	if not scene.is_file():
	return None, None, "no scene.xml"
	return scene, mt_npz, run_id


	def render_mt(scene_xml: Path, mt_npz: Path, out_mp4: Path,
	camera_name: str = "front") -> tuple[bool, str]:
	"""Render MT's captured DOF trajectory by name-permutation into scene.xml.

	MT dumps IG-ordered `q` plus `dof_names`. MJ scene.xml has its own
	per-joint ordering. We build a permutation `mj_qpos_adr[j] = ig_q[k]`
	where `dof_names[k] == mj_joint_names[j]`.
	"""
	try:
	import mujoco
	except ImportError:
	return False, "mujoco import failed"
	from bench.video_recorder import BENCH_CAMERA, BENCH_VIDEO, VideoRecorder

	with np.load(mt_npz, allow_pickle=True) as z:
	if "q" not in z.files:
	return False, f"npz missing 'q' key (has {z.files})"
	mt_q = z["q"].astype(np.float32) # (T, n_ig_dof)
	ig_names = list(z["dof_names"]) if "dof_names" in z.files else None

	T = int(mt_q.shape[0])
	if T == 0:
	return False, "empty mt_q"

	cam_spec = BENCH_CAMERA[camera_name]
	w, h = BENCH_VIDEO["resolution"]
	fps = BENCH_VIDEO["fps"]

	# Reuse the XML prep logic from render() via a tiny local copy.
	import re as _re
	patched = scene_xml # placeholder; assigned below
	try:
	import subprocess # noqa: F401 - for ffmpeg via VideoRecorder
	txt = scene_xml.read_text()
	txt = txt.replace(' colorspace="linear"', '').replace(' colorspace="srgb"', '')
	new_global = f'<global offwidth="{w}" offheight="{h}"/>'
	if "<global" in txt:
	def _sub(m):
	tag = m.group(0)
	if "offwidth" in tag and "offheight" in tag:
	return tag
	closing = "/>" if tag.rstrip().endswith("/>") else ">"
	body = tag[:len(tag) - len(closing)].rstrip()
	return f'{body} offwidth="{w}" offheight="{h}"{closing}'
	txt = _re.sub(r"<global[^>]*/?>", _sub, txt, count=1)
	elif "<visual>" in txt:
	txt = txt.replace("<visual>", f"<visual>\n {new_global}", 1)
	else:
	txt = _re.sub(r"(<mujoco[^>]*>)", rf"\1\n <visual>\n {new_global}\n </visual>",
	txt, count=1)
	patch_root = Path("/tmp/scene_patches")
	hand = scene_xml.parents[2].name
	task = scene_xml.parent.name
	staging = patch_root / hand / "bimanual" / task
	staging.mkdir(parents=True, exist_ok=True)
	assets_src = scene_xml.parents[3] / "assets"
	assets_link = patch_root / "assets"
	if not assets_link.exists():
	assets_link.symlink_to(assets_src)
	patched = staging / "scene.xml"
	patched.write_text(txt)
	except Exception as exc:
	return False, f"xml prep failed: {exc}"

	try:
	model = mujoco.MjModel.from_xml_path(str(patched))
	except ValueError as exc:
	return False, f"mjcf load failed: {exc}"

	# Build MJ-joint -> IG-DOF permutation by name.
	perm = None
	if ig_names is not None:
	ig_idx = {str(n): i for i, n in enumerate(ig_names)}
	perm = []
	for j in range(model.njnt):
	jn = mujoco.mj_id2name(model, mujoco.mjtObj.mjOBJ_JOINT, j)
	qadr = int(model.jnt_qposadr[j])
	jtype = int(model.jnt_type[j])
	# Skip free + ball joints (nq>1 per joint) — we don't map obj free joint
	if jtype != mujoco.mjtJoint.mjJNT_HINGE and jtype != mujoco.mjtJoint.mjJNT_SLIDE:
	continue
	if jn in ig_idx:
	perm.append((qadr, ig_idx[jn]))
	note_perm = f"perm size={len(perm)}/{model.njnt}"
	else:
	# Fallback: write first min(nq, n_ig_dof) values directly.
	k = min(model.nq, mt_q.shape[1])
	perm = [(i, i) for i in range(k)]
	note_perm = f"no dof_names; fallback linear k={k}"

	data = mujoco.MjData(model)
	try:
	renderer = mujoco.Renderer(model, height=h, width=w)
	except Exception as exc:
	return False, f"renderer init failed: {exc}"

	cam = mujoco.MjvCamera()
	pos = np.array(cam_spec["pos"], dtype=np.float64)
	look = np.array(cam_spec["lookat"], dtype=np.float64)
	cam.lookat[:] = look
	d = pos - look
	cam.distance = float(np.linalg.norm(d))
	cam.azimuth = float(np.degrees(np.arctan2(d[0], -d[1])))
	cam.elevation = float(np.degrees(np.arcsin(d[2] / (cam.distance + 1e-9))))

	out_mp4.parent.mkdir(parents=True, exist_ok=True)
	rec = VideoRecorder(out_mp4, fps=fps, size=(w, h))
	frames = 0
	try:
	for t in range(T):
	q_t = mt_q[t]
	for qadr, idx in perm:
	if idx < q_t.shape[0]:
	data.qpos[qadr] = q_t[idx]
	mujoco.mj_forward(model, data)
	renderer.update_scene(data, camera=cam)
	rec.log_frame(renderer.render())
	frames += 1
	rec.close()
	except Exception as exc:
	try: rec.close()
	except: pass
	return False, f"render failed at frame {frames}: {exc}"
	finally:
	try: renderer.close()
	except Exception: pass

	return frames > 0, f"{frames} frames, {note_perm}"


	def main() -> int:
	ap = argparse.ArgumentParser()
	ap.add_argument("--hand", choices=HANDS, default=None)
	ap.add_argument("--traj", choices=OAKINK_TRAJS, default=None)
	ap.add_argument("--source", choices=["kinematic", "mt"], default="kinematic",
	help="'kinematic' = Oracle/reference via Spider preprocess; "
	"'mt' = ManipTrans captured qpos via bench_hooks.patch")
	ap.add_argument("--mt-seed", type=int, default=42,
	help="Seed tag on MT run_id (default 42)")
	ap.add_argument("--mt-warmstart", default="default",
	help="Warmstart tag on MT run_id (default 'default')")
	ap.add_argument("--method-label", default=None,
	help="Override filename prefix (default: 'std' for kinematic, 'mt' for mt)")
	ap.add_argument("--force", action="store_true")
	args = ap.parse_args()

	cells = []
	for h in HANDS if args.hand is None else [args.hand]:
	for t in OAKINK_TRAJS if args.traj is None else [args.traj]:
	cells.append((h, t))

	label = args.method_label or ("mt" if args.source == "mt" else "std")
	n_ok = n_fail = n_skip = 0
	print(f"{'='72}\nstandardized render ({args.source}) — {len(cells)} cells\n{'='72}")
	print(f"camera: front pos=[0,-1.6,2.2] lookat=[0,-0.1,1.2] fov=30")
	print(f"video: 720×480 @ 30fps, h264 crf=20")
	print(f"output: {OUT_DIR}\n")

	for hand, traj in cells:
	if args.source == "mt":
	scene, npz, tag_or_rid = _resolve_mt_cell(hand, traj, args.mt_seed, args.mt_warmstart)
	if scene is None:
	print(f" [fail] {hand}/{traj}: {tag_or_rid}")
	n_fail += 1; continue
	# Preserve seed + warmstart in output name so each MT run renders distinctly.
	out = OUT_DIR / f"{label}_{hand}_oakink_v2_{traj}_bimanual_seed{args.mt_seed}_{args.mt_warmstart}.mp4"
	if out.is_file() and not args.force:
	print(f" [skip] {out.name} (exists)"); n_skip += 1; continue
	ok, note = render_mt(scene, npz, out)
	status = "✅pass" if ok else "❌fail"
	print(f" [{status}] {hand}/{traj} (src=mt_qpos): {note}")
	if ok: n_ok += 1
	else: n_fail += 1
	continue

	run_id = f"{label}_{hand}_oakink_v2_{traj}_bimanual_seed0_default"
	out = OUT_DIR / f"{run_id}.mp4"
	if out.is_file() and not args.force:
	print(f" [skip] {run_id} (exists)")
	n_skip += 1; continue

	scene, npz, tag = resolve_cell(hand, traj)
	if scene is None:
	print(f" [fail] {hand}/{traj}: {tag}")
	n_fail += 1; continue

	qpos = load_qpos(npz)
	if qpos is None or qpos.shape[0] == 0:
	print(f" [fail] {hand}/{traj}: no qpos in {npz.name}")
	n_fail += 1; continue

	ok, note = render(scene, qpos, out)
	status = "✅pass" if ok else "❌fail"
	print(f" [{status}] {hand}/{traj} (src={tag}): {note}")
	if ok:
	n_ok += 1
	else:
	n_fail += 1

	print(f"\n{'='*72}\nsummary: {n_ok} rendered, {n_fail} failed, "
	f"{n_skip} skipped (already present)\n{'='*72}")
	return 0 if n_fail == 0 else 1


	if __name__ == "__main__":
	sys.exit(main())