WanhaoX25/GCDG-Benchmark

GCDG Benchmark

Dataset accompanying Minimal-Intervention Target Retrieval in Cluttered Scenes with Grasp-Conditioned Dependency Graphs.

Code and evaluation · Project page · Paper

The viewer above is a target index, not the full annotation content. The eleven archives contain the scene configurations, grasp proposals and dependency annotations. Download all shards for a split and extract them into the same directory.

GCDG dense-scene benchmark

The benchmark supports grasp-conditioned dependency prediction, stored-label minimal-intervention planning and physical target retrieval. Its 500 dense scenes each contain 25–27 objects. Training, validation and test are separated by scene, not by target.

Split	Scenes	Targets	Parallel-jaw grasps	Suction grasps
Train	350	9,090	112,521	106,428
Validation	75	1,950	24,462	22,792
Test	75	1,945	24,321	22,704
Total	500	12,985	161,304	151,924

Download and verify

Use release v1.0.0, or run:

python scripts/download_benchmark.py --split all --output data --extract

The download script checks SHA-256 before extraction. Shards contain up to 50 scenes and can be extracted into the same directory. The full uncompressed source annotation set is approximately 17.6 GB; released JSON is compacted. Download only the test shards to evaluate an existing checkpoint. Download train and validation as well to reproduce training.

Schema

benchmark/
  splits.json
  class_map.json
  scenes/scene_<split>_<id>/
    scene.json
    targets/obj_<id>/
      manifest.json
      proposals.json
      labels.json

scene.json stores object identities, physical specifications and poses. manifest.json provides observation-derived graph support and artifact metadata. proposals.json stores parallel-jaw and suction candidates with their pose and modality. labels.json stores staged validation outcomes, single-object intervention labels and bounded minimal blocker sets.

One sample is one scene–target pair, with a 36-dimensional feature vector per non-target object, a 28-dimensional vector per candidate grasp and a 22-dimensional vector per object–grasp edge. The predictor outputs four logits per edge:

Output	Meaning	Stored annotation
dep_progress_any	Approach or lift progress	Union of the two stage labels
dep_sufficient	Removing this object alone makes the fixed grasp feasible	dep_any
dep_approach	Approach-stage progress after removal	dep_collision_approach
dep_lift	Lift-stage progress after removal	dep_collision_lift

The all-zero vector is the no-dependency state. The supervision mask identifies actual modeled edges. Targets with no proposals remain in the benchmark; they are not discarded. A target with no proposals or no solution within the depth-three blocker search is not treated as a zero-blocker success.

The fixed feature settings disable privileged scene geometry. Ground-truth poses and intervention outcomes are supplied for labeling, physics and oracle evaluation; they must not be introduced as learned-predictor inputs. Exact feature names are in src/grasp_dependency_dataset/hetero_gnn/graph_features.py.

Evaluation tasks

1. Prediction: 75 test scenes / 1,945 targets; per-label AP, validation-calibrated F1 and blocker IoU. Report mean and sample SD over seeds 7/11/23. AP groups tied scores using scikit-learn's definition.
2. Stored-label planning: the same 75 test scenes / 1,945 targets, with 42 frozen configurations. Validation selection uses success minus 0.08 times removals. Keep the oracle separate from learned policies.
3. Single-target physical retrieval: the first ten specified test roots, 256 targets, five methods, 1,280 episodes. Reset to the initial scene for each target. Allow at most five non-target removals, including shared fallback.
4. Ordered retrieval: four additional development scenes with 15 objects, ten targets per order and four orders per scene. Each of nine conditions has 16 sequences / 160 target trials. These are repeated orders of four scenes. Continue after ordinary grasp failures; stop on invalid physical states and retain unattempted targets in the denominator.
5. Hardware: three existing batches of twenty trials per method. The release reaggregates recorded counts; it does not add robot trials.

See protocol.json and all tables. Ordered retrieval uses all five Ours seeds. The accepted representative seed 11 was selected after observing the original-order development results and fixed before the additional orders.

Assets and portability

The annotation release contains the stored graph inputs needed to train and evaluate dependency predictors without running AnyGrasp or MuJoCo. It does not contain third-party mesh/texture files or native grasp SDK binaries. See third-party dependencies for asset acquisition and simulation setup. Public JSON relocates machine-specific paths while preserving all numeric values. The source and released SHA-256 values are recorded in the release manifest.

The objects derive from the CEPB object collection. The local mesh asset provenance and redistribution terms need to be matched to their original providers before redistribution; a paper's open-access license does not itself license all underlying mesh files. The release therefore keeps external asset acquisition separate from the self-owned annotations.

License

Original annotations, scene configurations and results: CC BY 4.0. Cite GCDG and acknowledge the original object collections when using their assets. Code: MIT.