Hugging Face's logo

Datasets:
wanlilll
/
WeaveBench

Tasks:
Other
Languages:
English
Size:
n<1K
ArXiv:
Tags:
benchmark
computer-use-agent
gui
cli
hybrid-interface
long-horizon
License:
Dataset card Files
xet
 Community
WeaveBench / tasks /SPA /SPA_task_0_blender_room_arrange.md
wanlilll's picture
wanlilll
Add 114 tasks under 8-domain flat layout
71e9dba verified
preview code
|
Raw History Blame Contribute Delete
17.9 kB
metadata 
id: SPA_task_0_blender_room_arrange
name: Blender 室内 3D 摆放小场景
category: SPA
timeout_seconds: 1800

Prompt

⚙️ Execution convention: this task is graded purely on the deliverable files — there is no human approver. Just execute, produce the files, and don't write "if you would like me to..." style questions or wait for approval. All deliverables must land in the /tmp_workspace/ root directory, and once finished, run ls -la /tmp_workspace/ as a self-check before exiting.

Background: under /tmp_workspace/assets/ there are 6 .blend furniture assets (sofa / coffee table / TV stand / dining table / chairs ×2 / nightstand). Note that assets/floor_lamp.blend is a legacy name — the internal object is actually ClassicNightstand_01 (a nightstand). Place it according to "nightstand" semantics; keep the filename unchanged.

Goal: build a 30 m² rectangular living room (approximately 5 m × 6 m × 2.8 m, with 4 walls + floor + ceiling), arrange the 6 pieces of furniture per the home-layout common sense below, add basic lighting and a camera, render the scene, and produce the primary deliverable room.blend plus the supporting deliverables. The implementation method is up to you (Blender GUI / bpy script / headless blender -b -P are all fine); only the final deliverable fields are graded.

Placement rules:

  • Sofa against the north wall, ~100 mm clearance from the wall;
  • TV stand against the south wall, centered horizontally;
  • Coffee table sits between the sofa and the TV stand, centered;
  • Dining table against the east wall with ~800 mm clearance, with the two chairs on either side of it;
  • Nightstand (from floor_lamp.blend) in the corner at the right end of the sofa, flush against the sofa's short side as a side table.

Hard constraints on the primary deliverable room.blend

  • The filename must be exactly room.blend (not room_scene.blend or similar).
  • Must contain at least 6 MESH objects whose names can be recognized as synonyms of the following 6 furniture categories: sofa / coffee(table) / tv(stand or console) / dining(table) / chair / nightstand (bedside / sidetable / lamp and similar synonyms are accepted).
  • The xy bounding boxes of the 6 pieces must pairwise not overlap (no clipping); the overall overlap-compliance rate must be ≥ 0.85.
  • Placement orientation and wall clearances must conform to the rules above (tolerance ~ ±50–100 mm).
  • At least 1 piece of furniture must have distance_to_nearest_wall_mm within [0, 200] (i.e., genuinely flush against a wall).

Supporting deliverables (under /tmp_workspace/ root)

File Requirement
render.png Perspective render, resolution strictly 1920×1080, with real materials and shadows (grayscale std > 35), file size ≥ 30 KB
top_view.png Top-down view screenshot, resolution ≥ 1280×720, file size ≥ 15 KB; md5 must differ from render.png (prevents renaming the same image)
layout.csv ≥ 6 rows; header must be object,x,y,z,rotation_z,distance_to_nearest_wall_mm; the 6 object names must all be distinct

Overall, the scene should visibly demonstrate: a complete room, no furniture clipping, ≥ 3 large pieces against walls, a walkable aisle through the center, and normal lighting (not all-black / not over-exposed).

Expected Behavior

设计意图与典型解题路径（仅供出题人参考，不发给 agent）：

  1. 推荐用 Blender 4.0 GUI 操作（也可走 bpy 脚本或 headless blender -b -P，三种通道判分等价）。
  2. 用 Cube/Plane 建 4 面墙 + 地板 + 天花板，内部净尺寸约 5m×6m×2.8m，墙厚向外延伸不要侵占内部空间。
  3. File → Append 从 /tmp_workspace/assets/ 中逐个导入 6 个 .blend 的 Object。
  4. 在 N 面板按摆放规则输入精确坐标（沙发 y 贴北墙、电视柜 y 贴南墙居中且与茶几共线、餐桌 x 靠东墙、椅子分坐餐桌两侧、床头柜在沙发右端角落）。
  5. 添加 1 个 Sun 灯 + 1 个 Area 灯。
  6. 设置一个人眼高度 ~1.6m 的对角斜拍相机，切换到 Cycles 或 Eevee，渲染 1920×1080 → /tmp_workspace/render.png。
  7. 切顶视图（Numpad 7 或正交相机）截图 → /tmp_workspace/top_view.png，确保与 render.png 是不同视角的真实图（md5 不同）。
  8. 写 /tmp_workspace/layout.csv，6 行，表头齐全，至少 1 件家具 distance_to_nearest_wall_mm ≤ 200。
  9. 保存场景为 /tmp_workspace/room.blend，最后 ls -la /tmp_workspace/ 自检。

约束说明：摆放精度约 ±50mm；render.png 与 top_view.png 必须是真实不同的图，禁止复制同一张图改名（grader 会比对 md5）。

评分要点（hard gates，从 grader cap 反推）：

  • room.blend 不存在 → 上限 0.30；
  • bbox no-overlap < 0.85 → 上限 0.40；
  • 摆放规则符合度 < 0.6 → 上限 0.45；
  • mesh_count 不足 6 → 上限 0.50；
  • render.png 非 1920×1080 或 std 不达标 → 上限 0.45；
  • top_view.png 缺失或分辨率不达标 → 上限 0.50；
  • 两张截图 md5 相同（同图复用） → 上限 0.45；
  • layout.csv 表头不全或 6 个 object 名不唯一 → 上限 0.50；
  • VLM 综合评分 < 0.6 → 上限 0.50；< 0.4 → 上限 0.35；VLM 不可用 → 上限 0.60。

Source

Grading Criteria

  • room.blend 存在，含 ≥ 6 mesh object 命名匹配 sofa/coffee/tv/dining/chair/nightstand(或 lamp 别名)
  • 6 物体的 bounding box 两两 IoU = 0（无穿模），no_overlap ≥ 0.85 才不被 cap
  • render.png 1920×1080，非纯色（std > 35），文件大小 ≥ 30 KB
  • top_view.png 存在，分辨率 ≥ 1280×720，文件大小 ≥ 15 KB
  • layout.csv 6 行，列齐全；6 个 object 名两两不同
  • 至少 1 个家具 distance_to_nearest_wall_mm 落在 [0, 200] 区间内
  • render.png 与 top_view.png md5 不相同（防同图复用）

Automated Checks 

from pathlib import Path
import csv, re, json, subprocess
from PIL import Image
import numpy as np

def grade(workspace_path=None, **kwargs):
    workspace = Path(workspace_path) if workspace_path else Path("/tmp_workspace")
    """Stricter SPA_5 grader: bounding-box rule conformance against gt/expected.json."""
    gt_path = (workspace.parent / "gt" / "expected.json")
    if not gt_path.exists():
        gt_path = Path("/tmp_workspace/gt/expected.json")
    gt = json.loads(gt_path.read_text()) if gt_path.exists() else {}
    s = {}
    bl = workspace/"room.blend"
    s["blend_exists"] = 1.0 if bl.exists() else 0.0
    objs = []
    blender_ok = False
    if bl.exists():
        try:
            sc=subprocess.run(["blender","-b",str(bl),"--python-expr",
                "import bpy,json;objs=[(o.name,list(o.location),list(o.dimensions),list(o.rotation_euler)) for o in bpy.data.objects if o.type=='MESH'];print('JSON_OUT',json.dumps(objs))"],
                capture_output=True,timeout=180,text=True)
            m=re.search(r"JSON_OUT (.+)", sc.stdout)
            if m: objs=json.loads(m.group(1)); blender_ok = True
        except Exception as e: s["blender_err"]=str(e)[:80]
    # Fallback: parse layout.csv if blender CLI not available / parse failed
    if not objs:
        lc = workspace/"layout.csv"
        if lc.exists():
            try:
                import csv as _csv
                with lc.open() as fh:
                    for row in _csv.DictReader(fh):
                        try:
                            x = float(row.get("x", 0) or 0)
                            y = float(row.get("y", 0) or 0)
                            z = float(row.get("z", 0) or 0)
                            rz = float(row.get("rotation_z", 0) or 0)
                            objs.append((row.get("object","obj"),
                                         [x, y, z],
                                         [0.5, 0.5, 0.5],  # default bbox dims
                                         [0, 0, rz]))
                        except Exception: pass
            except Exception: pass
    s["mesh_count"] = min(1.0, len(objs)/max(gt.get("min_meshes",6),6))
    # keys -> 同义词集合（覆盖 Polyhaven 真实对象名 ClassicConsole / WoodenTable /
    # ClassicNightstand / ArmChair / WoodenChair 等，以及常见英文别名 couch/desk/stand/light）
    key_synonyms = {
        "sofa":   ["sofa", "couch", "settee"],
        "coffee": ["coffee", "coffeetable", "cocktail"],
        "tv":     ["tv", "television", "console", "classicconsole", "tvstand", "stand", "media", "cabinet"],
        "dining": ["dining", "diningtable", "woodentable", "table"],
        "chair":  ["chair", "armchair", "woodenchair", "stool", "seat"],
        "nightstand": ["nightstand", "classicnightstand", "bedside", "sidetable", "endtable", "lamp", "light", "floorlamp", "torchiere"],
    }
    def _norm(n): return re.sub(r"[^a-z0-9]", "", n.lower())
    norm_names = [_norm(n) for n,_,_,_ in objs] if objs else []
    name_blob = " ".join(norm_names)
    matched_keys = set()
    for k, syns in key_synonyms.items():
        if any(s in name_blob for s in syns): matched_keys.add(k)
    s["furniture_named"] = len(matched_keys)/6
    def _obj_key(name):
        nn = _norm(name)
        for k, syns in key_synonyms.items():
            if any(s in nn for s in syns): return k
        return None
    # bbox xy non-overlap
    boxes=[]
    for n,loc,dim,_ in objs:
        if _obj_key(n) is not None:
            boxes.append((loc[0]-dim[0]/2,loc[1]-dim[1]/2,loc[0]+dim[0]/2,loc[1]+dim[1]/2))
    no_overlap = True; pairs=0; ok_pairs=0
    for i in range(len(boxes)):
        for j in range(i+1,len(boxes)):
            pairs+=1; a,b=boxes[i],boxes[j]
            if max(a[0],b[0])>=min(a[2],b[2]) or max(a[1],b[1])>=min(a[3],b[3]): ok_pairs+=1
            else: no_overlap=False
    s["no_overlap"] = (ok_pairs/pairs) if pairs else 0.0
    # rule conformance vs gt
    R = gt.get("room_dims_m",{}); tol = gt.get("tolerance_m",0.05)
    Y = R.get("y",6.0); X = R.get("x",5.0)
    rule_hits=0; rule_total=0
    obj_map = {}
    for n,loc,dim,_ in objs:
        k = _obj_key(n)
        if k is not None: obj_map.setdefault(k, (loc,dim))
    for rule in gt.get("rules",[]):
        rule_total+=1
        target = next((k for k in key_synonyms if k in rule.get("object","")), None)
        if not target or target not in obj_map: continue
        loc,dim = obj_map[target]
        if rule.get("wall")=="north":
            d = abs(Y - (loc[1]+dim[1]/2))
            if abs(d - rule.get("distance_m",0))<=tol+0.05: rule_hits+=1
        elif rule.get("wall")=="south":
            d = loc[1]-dim[1]/2
            if d<=tol+0.05: rule_hits+=1
        elif rule.get("wall")=="east":
            d = abs(X - (loc[0]+dim[0]/2))
            if abs(d - rule.get("distance_m",0))<=tol+0.1: rule_hits+=1
        if "x_center_m" in rule and abs(loc[0]-rule["x_center_m"])<=0.2: rule_hits+=0.5
    s["rule_conformance"] = min(1.0, rule_hits/max(rule_total,1))
    # render & top_view
    rp=workspace/"render.png"
    render_md5 = None
    if rp.exists():
        im=Image.open(rp); arr=np.array(im.convert("L"))
        s["render_size"] = 1.0 if im.size==(1920,1080) else 0.0
        s["render_nontrivial"] = 1.0 if arr.std()>35 else max(0.0, (arr.std()-10)/25)
        try:
            import hashlib
            render_md5 = hashlib.md5(rp.read_bytes()).hexdigest()
            s["render_filesize_ok"] = 1.0 if rp.stat().st_size >= 30*1024 else 0.0
        except Exception:
            s["render_filesize_ok"] = 0.0
    else:
        s["render_size"]=0.0; s["render_nontrivial"]=0.0; s["render_filesize_ok"]=0.0
    tv=workspace/"top_view.png"
    tv_md5 = None
    if tv.exists():
        _tvsz = Image.open(tv).size
        s["top_view_ok"] = 1.0 if (_tvsz[0]>=1280 and _tvsz[1]>=720) else 0.0
        try:
            import hashlib
            tv_md5 = hashlib.md5(tv.read_bytes()).hexdigest()
            s["top_view_filesize_ok"] = 1.0 if tv.stat().st_size >= 15*1024 else 0.0
        except Exception:
            s["top_view_filesize_ok"] = 0.0
    else:
        s["top_view_ok"] = 0.0; s["top_view_filesize_ok"] = 0.0
    s["screenshots_distinct"] = 1.0 if (render_md5 and tv_md5 and render_md5 != tv_md5) else 0.0
    lc=workspace/"layout.csv"
    if lc.exists():
        rows=list(csv.DictReader(lc.open()))
        if len(rows)>=6 and all(k in rows[0] for k in ["object","x","y","z","rotation_z","distance_to_nearest_wall_mm"]):
            s["layout_csv_schema"] = 1.0
            try:
                ok = sum(1 for row in rows if 0<=float(row["distance_to_nearest_wall_mm"])<=200)
                s["layout_csv_distance"] = 1.0 if ok>=1 else 0.0
            except: s["layout_csv_distance"]=0.0
            names = [r.get("object","").strip().lower() for r in rows[:6]]
            s["layout_csv_unique"] = 1.0 if len(set(names))==6 and all(names) else 0.0
        else:
            s["layout_csv_schema"]=0.5 if rows else 0.0; s["layout_csv_distance"]=0.0; s["layout_csv_unique"]=0.0
    else:
        s["layout_csv_schema"]=0.0; s["layout_csv_distance"]=0.0; s["layout_csv_unique"]=0.0
    try:
        from _judge_helper import vlm_score_rubric
    except Exception:
        vlm_score_rubric = None
    imgs = [str(p) for p in [workspace/"render.png", workspace/"top_view.png"] if p.exists()]
    if vlm_score_rubric and imgs:
        rubric = {
            "vlm_room_complete": "渲染图呈现一个完整的室内房间场景，含 ≥6 件家具（沙发/桌椅/柜灯等）",
            "vlm_no_overlap": "家具之间无明显穿模或重叠，摆放空间合理",
            "vlm_against_wall": "至少 3 件大件家具贴墙摆放（沙发靠墙、柜靠墙等）",
            "vlm_walkable_aisle": "中央或主要通道留有可行走的空间，未被家具完全占满",
            "vlm_lighting_ok": "渲染图光照正常（非全黑、非过曝），材质可辨",
        }
        vlm = vlm_score_rubric(imgs[:2], rubric, instruction="评估 Blender 室内家具布置的合理性。第一张为透视渲染，第二张为顶视图（如有）。")
        for k in rubric: s[k] = vlm.get(k, 0.0)
        s["judge_method"] = vlm.get("judge_method", "failed")
    nums=[v for v in s.values() if isinstance(v,(int,float))]
    flat = sum(nums)/len(nums) if nums else 0.0
    # 加权：核心交付 60% / GUI 证据 30% / 辅助 10%
    core_keys = ["blend_exists","mesh_count","furniture_named","no_overlap","rule_conformance","layout_csv_schema","layout_csv_distance","layout_csv_unique"]
    gui_keys  = ["render_size","render_nontrivial","render_filesize_ok","top_view_ok","top_view_filesize_ok","screenshots_distinct"]
    aux_keys  = [k for k in s if k.startswith("vlm_")]
    def _avg(keys):
        vals = [s[k] for k in keys if k in s and isinstance(s[k],(int,float))]
        return sum(vals)/len(vals) if vals else 0.0
    core = _avg(core_keys); gui = _avg(gui_keys); aux = _avg(aux_keys) if aux_keys else flat
    base = 0.6*core + 0.3*gui + 0.1*aux
    # 多层 hard gate（均上调）
    if s.get("blend_exists",0) < 1.0: base = min(base, 0.30)
    if s.get("no_overlap",0) < 0.85: base = min(base, 0.40)
    if s.get("rule_conformance",0) < 0.6: base = min(base, 0.45)
    if s.get("mesh_count",0) < 1.0: base = min(base, 0.50)
    if s.get("render_size",0) < 1.0 or s.get("render_nontrivial",0) < 0.8: base = min(base, 0.45)
    if s.get("top_view_ok",0) < 1.0: base = min(base, 0.50)
    if s.get("screenshots_distinct",0) < 1.0: base = min(base, 0.45)
    if s.get("layout_csv_schema",0) < 1.0 or s.get("layout_csv_unique",0) < 1.0: base = min(base, 0.50)
    vlm_keys = [k for k in s if k.startswith("vlm_")]
    if vlm_keys:
        vlm_avg = sum(s.get(k,0) for k in vlm_keys)/len(vlm_keys)
        if vlm_avg < 0.6: base = min(base, 0.50)
        if vlm_avg < 0.4: base = min(base, 0.35)
    else:
        # VLM 不可用时退化分上限封顶 0.6
        base = min(base, 0.60)
    s["overall_score"] = round(base, 3)
    return s

Workspace Path

workspace/SPA/task_0_blender_room_arrange/

Skills

Env

Warmup 

which blender >/dev/null 2>&1 || (curl -fsSLo /tmp/blender.tar.xz https://download.blender.org/release/Blender4.0/blender-4.0.2-linux-x64.tar.xz && mkdir -p /opt/blender && tar -xJf /tmp/blender.tar.xz -C /opt/blender --strip-components=1 && ln -sf /opt/blender/blender /usr/local/bin/blender) || true
apt-get install -y -qq tesseract-ocr || true
pip install -q numpy pillow pytesseract || true