AssemblyWorldBench

AssemblyWorldBench is the benchmark of AssemblyWorld: Rethinking 3D Assembly with General-Purpose Agents. An agent receives the rigid parts of one object, scattered in an interactive 3D scene, and must assemble them by inspecting rendered views and manipulating parts through the environment's WebMCP tools, optionally guided by a reference image or a full assembly manual. The finished scene is scored geometrically against the ground-truth assembly.

The benchmark has 100 evaluations over 80 shapes from four sources, organized into five blocks of 20. Agent results for the systems in the paper are released separately in AssemblyWorld/AssemblyWorldBench-Results.

Block	Source dataset	Reference given to the agent	Objects	Parts per object
partnet-none	PartNet (Manual-PA split)	none	7 chairs, 7 tables, 6 storage	5–15
partnet-final-image	PartNet (Manual-PA split)	final assembly image	same 20 objects	5–15
ikea-manualbook	IKEA-Manual	full manual, original page order	20 furniture objects	3–19
assemblybench-manualbook	AssemblyBench	full step diagrams, in order	20 industrial objects	3–18
fantastic-breaks-none	Fantastic Breaks	none	20 broken objects	2

Objects were selected from task properties only (category and part-count bands, fixed seed), never from model results. The selection inputs are in selection/.

Layout

benchmark.json                     frozen definition: blocks, samples, protocol, aggregation, provenance
launch-commands.txt                one run command per block
index/samples.parquet              one row per evaluation (block, source, sample, category, band, parts, paths)
selection/                         spec, manifest, exclusions, exclusion candidates, overrides
<block>/task.txt                   the task text of the block
<block>/<dataset>/<config_id>/
    config.json                    preparation identity and per-sample initial episode records
    <sample>.episode.zip           initial episode: the only input an agent receives
    cache/<sample>/evaluation.json ground truth for scoring: part point clouds, GT poses, equivalence groups
    cache/<sample>/reference/<mode>/  reference pages attached to the prompt (final-image, manualbook)

An initial episode is a 3DWebAgent episode archive (3dwebagent-episode v1, runtime contract 3dwebagent-runtime-1): MuJoCo world, part meshes and the scattered initial state. The cache/ directories hold ground truth and must never be given to an agent; the runner reads only the reference pages from them, and the evaluator reads evaluation.json.

Running and scoring

The environment is served at https://assemblyworld.github.io/3DWebAgent/. Runs and scoring use assembly-world-agent:

# after downloading this repository to data/assemblyworldbench/
uv run --locked --extra episodes --group browser assembly-world-agent run \
  --dataset partnet-manualpa --config-id prep-v1-s0-i0-987553cbcd58 \
  --data data/assemblyworldbench/partnet-none --reference-mode none \
  --prompt-file data/assemblyworldbench/partnet-none/task.txt \
  --agent codex --model MODEL --headless
# ... one command per block, see launch-commands.txt
uv run --locked --extra episodes python scripts/evaluate_run.py \
  --benchmark data/assemblyworldbench/benchmark.json <run directories of the five blocks>

Scoring uses protocol assembly-evaluation-v2: the final assembly is aligned to the ground truth with one global SE(3) transform, and parts are matched to ground-truth parts within geometry equivalence groups. SCD is the whole-shape Chamfer distance (sum of bidirectional mean squared nearest-neighbor distances, ×1000); PA is the fraction of parts whose Chamfer distance to their matched part is at most 0.01; SR is 1 when every part is correct. The overall score is the mean over sources of the mean over that source's blocks; a sample without a scored episode counts SR = PA = 0, and a model refusal counts as a failed evaluation.

Sources and terms

Every block pins a source dataset revision in benchmark.json. The initial episodes embed the source part geometry, and the reference pages are the source images. These assets keep their original rights; see LICENSE.md and the source cards.

Source	Hugging Face dataset	Terms of the underlying assets
PartNet (Manual-PA)	AssemblyWorld/partnet-manualpa	PartNet and ImagePA / Manual-PA source rights; no blanket redistribution grant verified
IKEA-Manual	AssemblyWorld/ikea-manual	annotations CC BY-NC-SA 4.0; manuals and meshes keep their creators' rights
AssemblyBench	AssemblyWorld/assemblybench	modifications CC BY-SA 4.0 (MERL); Fusion 360 Gallery assets non-commercial research only under the Fusion 360 Gallery Dataset License
Fantastic Breaks	AssemblyWorld/fantastic-breaks	no dataset license specified; source assets keep their original rights

Please also cite the source datasets: Mo et al., PartNet, 2019; Zhang et al., Manual-PA, 2025; Wang et al., IKEA-Manual, 2022; Li et al., AssemblyBench, 2026; Lamb et al., Fantastic Breaks, 2023.