Registered classes

Everything a config can name with type:, registry by registry, generated from the registries, from what each class declares about itself and from the shipped configs. A chip says which methods the slot applies to; the toolbox's own registries have none. The same tables in a terminal:

grasppanda components                    # every registry
grasppanda components --slot backbones   # one registry, its columns explained
grasppanda components PTv3               # one class: options, configs, fits
grasppanda components --json             # for scripts

Which class can replace which: Swap rules. What was measured for every option under every method: Does it run?.

How to read the tables

• Contract columns (between the name and the status) are the class attributes the config checker compares before anything is built. Each registry explains its own.
• Status is what has been proven. tested: unit tests on CPU cover the class's contract and behaviour; for the stages of a ported method, the method's README says how it was compared with the reference implementation. weights-verified: additionally, weights released by the authors load strictly and the outputs match the official implementation; the measured difference is on Encoders checked against their authors' code. bridge: an adapter around a library installed as an optional extra; the library's own code runs, and the tests here cover the adapter (contract, shapes, gradients), not the library. —: the class declares no status. No status is a claim about grasp-detection accuracy.
• Source is the paper, method or library the class implements, and where the code comes from: written for GraspPanda from the paper and the behaviour of the reference code (the ported methods; licences and commits of the references are in the method READMEs), a bridge to a library, or adapted from permissively licensed code (listed in THIRD_PARTY_NOTICES.md).
• Needs lists optional dependencies (pip install -e ".[...]").
• Shipped configs are the runnable configs under configs/ whose resolved tree names the class.

Detectors

model — A complete method. Seed-based detectors fill the six slots below.

5 registered classes

• required_inputs: Observation fields the detector reads; the dataset must provide them.
• needs_supervision: Fields of the training supervision the component reads (objects are the dense label grids), each with the option that makes it read the field. The training dataset must serve every one; the config check compares.

Component	required_inputs	needs_supervision	Status	Source	What it is	Shipped configs
EconomicGrasp	(points)	—	tested	EconomicGrasp (Wu et al., ECCV 2024)	Economic-supervision grasp detector; see the module docstring and economicgrasp.yaml.	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e finegrasp/finegrasp_sunet14_g1b-realsense_10e
GSNet	(points)	—	tested	GSNet (Wang et al., ICCV 2021)	Graspness-based grasp detector; see the module docstring and gsnet.yaml.	experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e
GraspNetBaseline	(points)	—	tested	GraspNet-baseline (Fang et al., CVPR 2020)	The GraspNet-1Billion baseline; see the module docstring and graspnet_baseline.yaml.	experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e
HGGD	(depth, intrinsics, points, rgb)	extras.view_grasps (always)	tested	HGGD (Chen et al., RA-L 2023)	Heat-map guided grasp detector; see the module docstring and hggd.yaml.	experimental/hggd_dinov2-small-reg-depth-stem_g1b-realsense_15e hggd/hggd_r34w8_g1b-kinect_15e hggd/hggd_r34w8_g1b-realsense-to-gc6d-d435_15e hggd/hggd_r34w8_g1b-realsense_15e
ScaleBalancedGrasp	(points)	—	tested	Scale-Balanced-Grasp (Ma and Huang, CoRL 2022)	Scale-Balanced-Grasp; see the module docstring and scale_balanced_grasp.yaml.	experimental/sbg_ptv3_g1b-realsense_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e

Scene encoders

model.backbone — Scene encoder: points to features, on every point or on a subsample. Seed-based · 5 methods Options

33 registered classes

• output: Where the returned features live: dense = on every input point, subsampled = on a furthest-point subsample that records its parent. A proposal accepts one kind.

Component	output	Status	Source	What it is	Needs	Shipped configs
ASSANet	dense	tested	ASSANet (Qian et al., NeurIPS 2021), adapted from guochengqian/openpoints@bc778ec, MIT	ASSANet: PointNet++ with separable set abstraction, features on every input point.	—	—
DGCNN	dense	tested	DGCNN (Wang et al., ACM TOG 2019), adapted from antao97/dgcnn.pytorch@07d534c, MIT	Dynamic Graph CNN: EdgeConv over k-nearest-neighbour graphs, features on every point.	—	—
DeLA	dense	tested	DeLA (Chen et al., arXiv 2308.16532), adapted from Matrix-ASC/DeLA@0f427c3, MIT	DeLA: decoupled local aggregation on a grid pyramid, features on every input point.	—	—
FeaturePropagation	dense	tested	PointNet++ feature propagation (Qi et al., NeurIPS 2017)	Turn a subsampled backbone into a dense one by learned interpolation.	—	experimental/gsnet_pointnet2-fp_g1b-realsense_10e
FurthestPointSubsample	subsampled	tested	PointNet++ furthest point sampling (Qi et al., NeurIPS 2017)	Turn a dense backbone into a subsampled one by furthest point sampling.	—	experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/sbg_ptv3_g1b-realsense_18e
ImagePainted	dense	tested	PointPainting (Vora et al., CVPR 2020), PointAugmenting (Wang et al., CVPR 2021); written here	Give a dense scene encoder image features sampled at the pixel of every point.	—	—
ImagePaintedSubsampled	subsampled	tested	PointPainting (Vora et al., CVPR 2020), PointAugmenting (Wang et al., CVPR 2021); written here	Give a subsampled scene encoder image features sampled at the pixel of every point.	—	—
LitePT	dense	weights-verified	LitePT (Yue et al., CVPR 2026)	LitePT encoder-decoder (official LitePT, Pointcept LitePT-v1).	spconv-cu118 (or another spconv-* build)	experimental/gsnet_litept_g1b-realsense_10e
MinkUNet	dense	tested	MinkUNet (Choy et al., CVPR 2019), adapted from NVIDIA/MinkowskiEngine@02fc608, MIT	Residual sparse U-Net of MinkowskiEngine (MinkUNet 14 to 101, Res16UNet), per-point output.	spconv-cu118 (or another spconv-* build)	—
OACNNs	dense	tested	OA-CNNs (Peng et al., CVPR 2024), Pointcept oacnns-v1m1	OA-CNNs encoder-decoder (Pointcept oacnns-v1m1), one feature per input point.	spconv-cu118 (or another spconv-* build)	experimental/gsnet_oacnns_g1b-realsense_10e
OctFormer	dense	bridge	OctFormer (Wang, ACM TOG / SIGGRAPH 2023); adapted from octree-nn/octformer@c2975ab, MIT; octree operators via ocnn	OctFormer encoder with a feature-pyramid head: features on every input point.	ocnn	—
PTv3	dense	weights-verified	Point Transformer V3 (Wu et al., CVPR 2024)	Point Transformer V3 encoder-decoder (Pointcept PT-v3m1).	spconv-cu118 (or another spconv-* build)	experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/sbg_ptv3_g1b-realsense_18e
PTv3Sonata	dense	weights-verified	Sonata (Wu et al., CVPR 2025), Concerto (Zhang et al., NeurIPS 2025)	The PTv3 variant of Sonata and Concerto (Pointcept PT-v3m2).	spconv-cu118 (or another spconv-* build)	—
PTv3Utonia	dense	weights-verified	Utonia (Zhang et al., ICML 2026)	The PTv3 variant of Utonia (Pointcept PT-v3m3): PTv3Sonata plus RoPE.	spconv-cu118 (or another spconv-* build)	—
PointMLP	dense	tested	PointMLP (Ma et al., ICLR 2022), dense form, adapted from ma-xu/pointMLP-pytorch@1ab7f04, Apache-2.0	PointMLP: residual MLP stages with a geometric affine module, features on every point.	—	—
PointMamba	subsampled	bridge	PointMamba (Liang et al., NeurIPS 2024), layout of LMD0311/PointMamba@80858e0 (Apache-2.0); mixer via mamba-ssm	Mamba state-space model over point-patch tokens in Hilbert order (PointMamba).	mamba-ssm	—
PointMetaBase	dense	tested	PointMetaBase (Lin et al., CVPR 2023), adapted from linhaojia13/PointMetaBase@c364a67, MIT	PointMetaBase: set abstraction with the building block of PointMeta, per-point output.	—	—
PointNeXt	dense	tested	PointNeXt (Qian et al., NeurIPS 2022)	PointNeXt: PointNet++ with inverted residual MLP blocks, features on every input point.	—	experimental/gsnet_pointnext_g1b-realsense_10e
PointNet2	subsampled	weights-verified	PointNet++ (Qi et al., NeurIPS 2017), GraspNet-baseline layout	PointNet++ (SSG) returning features on a furthest-point subsample of the scene.	—	experimental/gsnet_pointnet2-fp_g1b-realsense_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e
PointNet2Dense	dense	tested	PointNet++ (Qi et al., NeurIPS 2017), segmentation layout	PointNet++ (SSG) propagated back to every input point.	—	—
PointNet2MSG	subsampled	tested	PointNet++ MSG (Qi et al., NeurIPS 2017), adapted from facebookresearch/votenet@2f6d6d3, MIT	PointNet++ with multi-scale grouping, features on a furthest-point subsample.	—	—
PointNet2MSGDense	dense	tested	PointNet++ MSG (Qi et al., NeurIPS 2017), segmentation layout, adapted from facebookresearch/votenet@2f6d6d3, MIT	PointNet++ with multi-scale grouping, propagated back to every input point.	—	—
PointNet2Transformer	subsampled	weights-verified	Scale-Balanced-Grasp (Ma and Huang, CoRL 2022)	PointNet++ set abstraction interleaved with point-transformer layers.	—	scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
PointTransformerV1	dense	tested	Point Transformer (Zhao et al., ICCV 2021); adapted from Pointcept/Pointcept@1342eda, MIT	Point Transformer (V1) U-Net with features on every input point.	—	—
PointTransformerV2	dense	tested	Point Transformer V2 (Wu et al., NeurIPS 2022); adapted from Pointcept/Pointcept@1342eda, MIT	Point Transformer V2 U-Net with features on every input point.	—	—
PointVector	dense	tested	PointVector (Deng et al., CVPR 2023), adapted from guochengqian/openpoints@bc778ec, MIT	PointVector encoder-decoder (scene form) with features on every input point.	—	—
PointViT	subsampled	tested	Point-BERT (Yu et al., CVPR 2022) and Point-MAE (Pang et al., ECCV 2022) encoder; adapted from Pang-Yatian/Point-MAE@7445a68, MIT	Plain transformer over point-patch tokens (Point-BERT / Point-MAE encoder).	—	—
PyGPointEncoder	dense	bridge	PointNetConv / EdgeConv / PointTransformerConv / PointGNNConv via torch_geometric (Fey and Lenssen, ICLR-W 2019)	A stack of PyTorch Geometric point convolutions with features on every input point.	torch_geometric	—
RandLANet	dense	tested	RandLA-Net (Hu et al., CVPR 2020), adapted from isl-org/Open3D-ML@a3d0f7a, MIT	RandLA-Net: random sub-sampling with attentive local aggregation, per-point output.	—	—
SpUNet	dense	tested	SpUNet of Pointcept, MinkUNet-34C layout (Choy et al., CVPR 2019)	Sparse residual U-Net of Pointcept (SpUNet-v1m1), one feature per input point.	spconv-cu118 (or another spconv-* build)	experimental/gsnet_spunet_g1b-realsense_10e
SparseUNet14	dense	tested	MinkUNet14D (Choy et al., CVPR 2019), as GSNet, EconomicGrasp and FineGrasp use it	Sparse residual U-Net returning one feature per input point.	spconv-cu118 (or another spconv-* build)	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e finegrasp/finegrasp_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e
Uni3D	subsampled	bridge	Uni3D (Zhou et al., ICLR 2024), layout of baaivision/Uni3D@64e03c3 (MIT); transformer blocks via timm	Uni3D point encoder: point-patch tokens through the blocks of a timm EVA ViT.	timm	—
Volt	dense	tested	Volt (Yilmaz et al., ECCV 2026, arXiv:2604.19609); adapted from YilmazKadir/Volt@8dcd608, MIT	Volt: global attention over voxel-patch tokens, features on every input point.	—	—

Image encoders

model.image_encoder — Image encoder: an RGB or RGB-D image to feature maps (HGGD, point painting). HGGD only Options

4 registered classes

Component	Status	Source	What it is	Needs	Shipped configs
HFImageEncoder	bridge	Hugging Face Transformers (AutoBackbone / AutoModel), via transformers	Any Hugging Face vision model, through AutoBackbone or AutoModel, by its hub name.	transformers	—
HGGDResidualEncoder	tested	HGGD (Chen et al., RA-L 2023)	HGGD's own image encoder: ResNet-34 block layout at base width 8, four input channels.	—	—
TimmImageEncoder	bridge	PyTorch Image Models (the timm model zoo), via timm	Any timm model that returns feature maps (features_only), by its timm name.	timm	experimental/hggd_dinov2-small-reg-depth-stem_g1b-realsense_15e
TorchvisionImageEncoder	bridge	The torchvision model zoo, via torchvision	A torchvision classification network as a pyramid encoder, by its torchvision name.	torchvision	—

Seed proposals

model.proposal — Scene features to seeds. Seed-based · 5 methods Options

4 registered classes

• requires: The kind of scene features the proposal works on (dense, subsampled, any).
• seed_scores: Per-seed scores the stage predicts, supervises and stores on the seeds. A score has one owner per detector.
• needs_supervision: Fields of the training supervision the component reads (objects are the dense label grids), each with the option that makes it read the field. The training dataset must serve every one; the config check compares.

Component	requires	seed_scores	needs_supervision	Status	Source	What it is	Shipped configs
GraspnessFPS	dense	(graspness, objectness)	point_labels.objectness (always), point_labels.graspness (always)	tested	GSNet (Wang et al., ICCV 2021)	Objectness + graspness per point, threshold, furthest point sampling.	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e finegrasp/finegrasp_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e
SampledSeeds	dense	—	—	tested	Dense-feature seed sampling as in GraspNet-baseline (Fang et al., CVPR 2020) with the samplers of 3DSSD (Yang et al., CVPR 2020) and others as options	Seeds picked by a sampler among all points of a dense encoder; nothing is scored.	—
ScoredSeeds	dense	(graspness, objectness)	point_labels.objectness (always), point_labels.graspness (always)	tested	GSNet (Wang et al., ICCV 2021) scorer with the selection rules of MultiGraspNet (product), DexGraspNet 2.0 (top fraction) and a swappable sampler	Objectness and graspness per point; the selecting score, the cut and the sampler are options.	—
SubsampledSeeds	subsampled	—	—	tested	GraspNet-baseline (Fang et al., CVPR 2020)	The backbone's subsample is the seed set; all seeds are real.	experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/sbg_ptv3_g1b-realsense_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e

Seed samplers

model.proposal.sampler — Part of a proposal: which of the candidate points become seeds. Seed-based · 5 methods Options

11 registered classes

• needs: What the sampler reads besides coordinates: scores (one value per candidate) and/or feats (the scene features). The proposal that holds it must provide them.

Component	needs	Status	Source	What it is	Needs	Shipped configs
FPS	—	tested	PointNet++ (Qi et al., NeurIPS 2017); grasppanda.ops.furthest_point_sample	Furthest point sampling on coordinates (D-FPS): seeds spread evenly over the candidates.	—	—
FPSThenTopK	(scores)	tested	Contact-GraspNet (Sundermeyer et al., ICRA 2021), selection rule; written from it	Furthest point sampling for coverage, then the best-scored of the rest.	—	—
FeatureFPS	(feats)	tested	3DSSD (Yang et al., CVPR 2020); written from the definition, conventions read in github.com/dvlab-research/3DSSD@8bc7605 (MIT)	Furthest point sampling in feature space (F-FPS): seeds that differ in what the encoder sees.	—	—
FpsampleFPS	—	bridge	QuickFPS (Han et al., IEEE TCAD 2023) via fpsample 1.0 (github.com/leonardodalinky/fpsample, MIT)	Furthest point sampling on the CPU through fpsample, with the bucket-based QuickFPS.	fpsample	—
FusionFPS	(feats)	tested	3DSSD (Yang et al., CVPR 2020); written from the definition, conventions read in github.com/dvlab-research/3DSSD@8bc7605 (MIT)	Fusion sampling (FS): part of the seeds by F-FPS, the rest by D-FPS.	—	—
RandomSampler	—	tested	GSNet (Wang et al., ICCV 2021), sampling ablation; RandLA-Net (Hu et al., CVPR 2020)	Uniformly random candidates without replacement: the baseline of every sampling ablation.	—	—
ScoreNMS	(scores)	tested	S4G (Qin et al., CoRL 2019), Algorithm 1; written from the paper	Greedy suppression on points: best score first, no two seeds closer than a radius.	—	—
ScoreWeightedFPS	(scores)	tested	SASA (Chen et al., AAAI 2022); written from the definition, conventions read in github.com/blakechen97/SASA@cd79f60 (Apache-2.0)	Semantics-guided furthest point sampling (S-FPS): spread seeds, prefer high scores.	—	—
TopK	(scores)	tested	IA-SSD (Zhang et al., CVPR 2022) and GraNet (Wang et al., IROS 2023); torch.sort	The best-scored candidates, wherever they are (top-k): no spatial spread at all.	—	—
TorchClusterFPS	—	bridge	PointNet++ (Qi et al., NeurIPS 2017) via torch_cluster.fps (github.com/rusty1s/pytorch_cluster, MIT; Fey & Lenssen, PyTorch Geometric)	Furthest point sampling through torch_cluster.fps, the kernel PyTorch Geometric uses.	torch-cluster	—
VoxelGridSampler	—	tested	Grid subsampling as in KPConv (Thomas et al., ICCV 2019); grasppanda.ops.voxelize	One candidate per occupied voxel: seeds on a regular grid, their number set by the scene.	—	—

View stages

model.view — One approach direction per seed. Seed-based · 5 methods Options

8 registered classes

• seed_scores: Per-seed scores the stage predicts, supervises and stores on the seeds. A score has one owner per detector.
• label_formats: Kinds of matched labels the view stage can compute its targets from.
• needs_matched: Optional fields of the matched labels the stage cannot work without; the assigner must return them.
• applies_seed_weights: Whether the losses honour a per-seed weight from the assigner. An assigner that returns weights is refused next to a stage that would ignore them.

Component	seed_scores	label_formats	needs_matched	applies_seed_weights	Status	Source	What it is	Needs	Shipped configs
DirectionViewStage	—	(dense, compact)	—	yes	tested	direct regression of the approach vector (ablation of GraspNet-1Billion, Fang et al., CVPR 2020; Generalizing-Grasp, Ma et al., CVPR 2024), optional von Mises-Fisher likelihood (vMF-Contact, Shi et al., ICRA 2025); written for GraspPanda	Regress the approach vector of every seed and snap it to the label lattice.	—	—
HealpixViewStage	—	(dense, compact)	—	yes	bridge	HEALPix (Górski et al., ApJ 2005) via healpix (BSD-3-Clause)	Score the pixel centres of a HEALPix grid (through healpix).	healpix	—
IcosphereViewStage	—	(dense, compact)	—	yes	bridge	icosphere lattice via trimesh (trimesh.creation.icosphere, MIT)	Score the vertices of a subdivided icosahedron (through trimesh).	trimesh	—
LatticeViewStage	—	(dense, compact)	—	yes	tested	view scoring of GraspNet-1Billion (Fang et al., CVPR 2020) and GSNet (Wang et al., ICCV 2021) on a free lattice, written for GraspPanda; hemisphere template and angular class loss of RGB Matters (Gou et al., ICRA 2021), the loss adapted from GouMinghao/rgb_matters@2e176728f6, MIT	Score a Fibonacci lattice of any size; labels stay on their own lattice.	—	—
ObjectnessViewStage	(objectness)	(dense)	(on_object, view_width)	yes	tested	GraspNet-baseline (Fang et al., CVPR 2020)	Classify every seed as object or background and choose its best template view.	—	experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/sbg_ptv3_g1b-realsense_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
RomaRotationViewStage	—	(dense, compact)	—	yes	bridge	rotation mappings via RoMa (Brégier, 3DV 2021; BSD-3-Clause): 6-D (Zhou et al., CVPR 2019), 9-D + SVD (Levinson et al., arXiv 2006.14616), quaternion, rotation vector	Regress a full rotation per seed through a mapping of the RoMa library.	roma	—
TemplateViewStage	—	(dense, compact)	—	yes	tested	GSNet (Wang et al., ICCV 2021)	Regress a score for each view of a Fibonacci template and choose one view.	—	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e finegrasp/finegrasp_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e
TwoVectorViewStage	—	(dense, compact)	—	yes	tested	two-vector Gram-Schmidt frame: 6-D rotation representation (Zhou et al., CVPR 2019) and the approach/baseline parameterisation of Contact-GraspNet (Sundermeyer et al., ICRA 2021); written for GraspPanda	Regress two vectors per seed and orthonormalise them into a grasp frame.	—	—

Assigners

model.assigner — Training only: seeds and chosen views to matched grasp labels. Seed-based · 5 methods Options

3 registered classes

• label_format: Kind of matched labels: dense (angle × depth grids per view) or compact (one grasp per point and view). Assigner and head bundle must agree.
• optional_outputs: Optional fields of the matched labels the assigner can return, each with the option that switches it on.
• needs_supervision: Fields of the training supervision the component reads (objects are the dense label grids), each with the option that makes it read the field. The training dataset must serve every one; the config check compares.

Component	label_format	optional_outputs	needs_supervision	Status	Source	What it is	Shipped configs
CompactLabelAssigner	compact	—	extras.compact_labels (always)	tested	EconomicGrasp (Wu et al., ECCV 2024)	Nearest annotated point per seed; its one grasp at the chosen view, if usable.	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e finegrasp/finegrasp_sunet14_g1b-realsense_10e
NearestLabelAssigner	dense	on_object (with supervise: object_seeds), view_width (with view_widths: true)	objects (always), point_labels.objectness (with supervise: object_seeds)	tested	Label matching of GraspNet-baseline and GSNet	Nearest annotated grasp point per seed, labels sliced at the chosen view.	experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e
ScaleBalancedAssigner	dense	on_object (with supervise: object_seeds), view_width (always), weight (always)	objects (always), point_labels.objectness (with supervise: object_seeds), extras.scale_histogram (always)	tested	Scale-Balanced-Grasp (Ma and Huang, CoRL 2022)	Nearest-label matching plus a scale-balanced loss weight per seed.	experimental/sbg_ptv3_g1b-realsense_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e

Groupers

model.grouper — A local crop per seed, aligned with its approach. Seed-based · 5 methods Options

14 registered classes

• support: What a crop gathers: seeds (the seeds and their learned features) or scene (raw scene geometry). Not the same crop at another resolution.
• out_axes: Named axes of the crop features. Must equal the head bundle's in_axes.

Component	support	out_axes	Status	Source	What it is	Needs	Shipped configs
CylinderGroup	seeds	(seed, channel)	tested	GSNet (Wang et al., ICCV 2021)	One approach-aligned cylinder per seed, grouping among the seeds.	—	experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e
CylinderInteractionGroup	seeds	(seed, channel)	tested	EconomicGrasp (Wu et al., ECCV 2024)	CylinderGroup with self-attention among the neighbours of each crop.	—	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e finegrasp/finegrasp_sunet14_g1b-realsense_10e
DepthCylinderGroup	scene	(seed, depth, channel)	tested	GraspNet-baseline (Fang et al., CVPR 2020)	Approach-aligned cylinders of increasing reach over the scene cloud.	—	experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
DepthCylinderProjectionGroup	scene	(seed, depth, channel)	tested	Scale-Balanced-Grasp (Ma and Huang, CoRL 2022) crop with a projection of the seed feature added here	DepthCylinderGroup with a gated seed residual for an encoder of any width.	—	experimental/sbg_ptv3_g1b-realsense_18e
DepthEmbedGroup	seeds	(seed, depth, channel)	tested	GraspPanda (axis adapter; no published model)	Give a (seed, channel) crop among the seeds a depth axis by a learned embedding.	—	—
DepthLocalGroup	scene	(seed, depth, channel)	tested	The depth-sliced crop of GraspNet-baseline (Fang et al., CVPR 2020) with the shapes and poolings of LocalGroup, written here on grasppanda.ops	A crop of the scene cloud, pooled once per insertion depth.	—	—
DepthPoolGroup	scene	(seed, channel)	tested	GraspPanda (axis adapter; no published model)	Reduce the depth axis of a (seed, depth, channel) crop of the scene.	—	—
LocalGroup	seeds	(seed, channel)	tested	Parts from PointNet++ (Qi et al., NeurIPS 2017), DGCNN (Wang et al., TOG 2019), PointRCNN (Shi et al., CVPR 2019), RandLA-Net (Hu et al., CVPR 2020), Point Transformer (Zhao et al., ICCV 2021) and PointNeXt (Qian et al., NeurIPS 2022), written here on grasppanda.ops	A crop among the seeds whose shape, encoder and pooling are options.	—	—
MultiRadiusGroup	seeds	(seed, channel)	tested	Multi-scale grouping of PointNet++ (Qi et al., NeurIPS 2017) on the seed crops of the grasp methods, written here	One crop per radius among the seeds, fused without attention.	—	—
MultiScaleGroup	seeds	(seed, channel)	tested	FineGrasp (arXiv 2507.05978)	One crop per radius among the seeds, fused by ScaleAttentionFusion.	—	experimental/finegrasp_ptv3_g1b-realsense_10e finegrasp/finegrasp_sunet14_g1b-realsense_10e
PyGConvGroup	seeds	(seed, channel)	bridge	PyTorch Geometric (Fey and Lenssen, 2019), MIT: PointNetConv, PointTransformerConv, EdgeConv, GATv2Conv, TransformerConv; via torch_geometric	A crop among the seeds, encoded by a message-passing layer of PyTorch Geometric.	torch_geometric	—
RoIAwareGroup	seeds	(seed, channel)	tested	Part-A² (Shi et al., TPAMI 2020), RoI-aware point cloud pooling; adapted from open-mmlab/OpenPCDet@233f849, Apache-2.0	RoI-aware pooling (Part-A²) on the gripper-aligned box of every seed.	—	—
RoIGridGroup	seeds	(seed, channel)	tested	PV-RCNN (Shi et al., CVPR 2020), RoI-grid pooling; adapted from open-mmlab/OpenPCDet@233f849, Apache-2.0	RoI-grid pooling (PV-RCNN) on the gripper-aligned box of every seed.	—	—
VoxelSetAbstractionGroup	scene	(seed, channel)	tested	PV-RCNN (Shi et al., CVPR 2020), voxel set abstraction; adapted from open-mmlab/OpenPCDet@233f849, Apache-2.0	The keypoint feature of PV-RCNN: scene points at several radii, fused with the seed.	—	—

Head bundles

model.head — Head bundle: prediction layers, target encoder, loss and decoder as one unit. Seed-based · 5 methods Options

4 registered classes

• in_axes: Named axes of the crop features the head bundle expects.
• label_format: Kind of matched labels: dense (angle × depth grids per view) or compact (one grasp per point and view). Assigner and head bundle must agree.
• score_kind: Meaning of the decoded grasp score. Scores of different kinds are not comparable.
• applies_seed_weights: Whether the losses honour a per-seed weight from the assigner. An assigner that returns weights is refused next to a stage that would ignore them.
• needs_supervision: Fields of the training supervision the component reads (objects are the dense label grids), each with the option that makes it read the field. The training dataset must serve every one; the config check compares.

Component	in_axes	label_format	score_kind	applies_seed_weights	needs_supervision	Status	Source	What it is	Shipped configs
DecoupledGridHead	(seed, channel)	dense	predicted_quality	yes	—	tested	after the grasp head of FGC-GraspNet (Lu et al., ICRA 2022), written here; width bins with the distribution focal loss of Generalized Focal Loss (Li et al., NeurIPS 2020)	Score grid plus separate angle, depth and width outputs per seed; one grasp per seed.	—
EconomicHead	(seed, channel)	compact	unspecified	yes	—	tested	EconomicGrasp (Wu et al., ECCV 2024)	Angle, depth and quality classes plus a width per seed; one decoded grasp per seed.	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e finegrasp/finegrasp_sunet14_g1b-realsense_10e
OperationHead	(seed, depth, channel)	dense	predicted_quality_x_tolerance	yes	objects.tolerance (always)	tested	GraspNet-baseline (Fang et al., CVPR 2020)	Score, angle class, width and tolerance per (angle, depth); one grasp per seed.	experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/sbg_ptv3_g1b-realsense_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
SWADHead	(seed, channel)	dense	predicted_quality	yes	—	tested	GSNet (Wang et al., ICCV 2021)	Score and width grids per seed, regressed; one decoded grasp per seed.	experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e

Primitive losses

<stage>.<name>_loss — Primitive loss inside a stage. Seed-based · 5 methods Options

22 registered classes

• accepts: Kinds of loss term the primitive fits: class (logits against class indices), unit (a target quality in [0, 1]), value (an unbounded target) or bins (logits against a fractional bin position). A stage names the kind of each of its loss options, and the config check refuses any other pairing.
• prediction: How the primitive reads the output of a unit term: as the value itself or as a logit. A stage hands a logit on through its sigmoid.

Component	accepts	prediction	Status	Source	What it is	Needs	Shipped configs
BalancedL1	(value, unit)	value	tested	Libra R-CNN (Pang et al., CVPR 2019); adapted from open-mmlab/mmdetection@cfd5d3a, Apache-2.0	Balanced L1: Smooth-L1 with a larger gradient for small errors.	—	—
BinaryCrossEntropy	(unit)	logit	tested	PyTorch binary_cross_entropy_with_logits	Sigmoid cross-entropy between a logit and a target in [0, 1], soft or binary.	—	—
ClassBalanced	(class)	value	tested	Class-Balanced Loss Based on Effective Number of Samples (Cui et al., CVPR 2019); the weight formula of the paper, written here	Class-balanced weighting: every class counts by the inverse of its effective number.	—	—
CrossEntropy	(class)	value	—	PyTorch cross_entropy	Softmax cross-entropy between (N, C) logits and (N,) class indices.	—	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e experimental/sbg_ptv3_g1b-realsense_18e finegrasp/finegrasp_sunet14_g1b-realsense_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
DiceLoss	(class)	value	bridge	V-Net (Milletari et al., arXiv 1606.04797), via kornia.losses	Dice loss: one minus the overlap of predicted and labelled classes over the set.	kornia	—
DistributionFocal	(bins)	value	tested	Generalized Focal Loss (Li et al., NeurIPS 2020); adapted from open-mmlab/mmdetection@cfd5d3a, Apache-2.0	Distribution focal loss: cross-entropy to the two bins around a continuous target.	—	—
FocalLoss	(class, unit)	logit	bridge	Focal Loss for Dense Object Detection (Lin et al., ICCV 2017), via torchvision.ops	Sigmoid focal loss: cross-entropy that turns down well-classified elements.	torchvision	—
GHMC	(unit, class)	logit	tested	Gradient Harmonized Single-stage Detector (Li et al., AAAI 2019); adapted from open-mmlab/mmdetection@cfd5d3a, Apache-2.0	Gradient harmonising, classification form: sigmoid cross-entropy re-weighted so that every range of gradient norms contributes alike.	—	—
GHMR	(value, unit)	value	tested	Gradient Harmonized Single-stage Detector (Li et al., AAAI 2019); adapted from open-mmlab/mmdetection@cfd5d3a, Apache-2.0	Gradient harmonising, regression form: a smooth L1 re-weighted so that every range of gradient norms contributes alike.	—	—
Huber	(value, unit)	value	tested	PyTorch huber_loss	Huber penalty: quadratic below delta, linear with slope delta above.	—	—
L1	(value, unit)	value	—	PyTorch l1_loss	Absolute error.	—	—
LabelSmoothingCrossEntropy	(class)	value	tested	Rethinking the Inception Architecture (Szegedy et al., CVPR 2016), via PyTorch cross_entropy	Softmax cross-entropy with label smoothing and, optionally, a factor per class.	—	—
LovaszSoftmaxLoss	(class)	value	bridge	The Lovász-Softmax loss (Berman et al., CVPR 2018), via kornia.losses	Lovász-Softmax: a convex surrogate of the mean intersection-over-union of the set.	kornia	—
MSE	(value, unit)	value	—	PyTorch mse_loss	Squared error.	—	experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/sbg_ptv3_g1b-realsense_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
Poly1CrossEntropy	(class)	value	tested	PolyLoss (Leng et al., ICLR; arXiv 2204.12511); the Poly-1 formula of the paper on PyTorch cross_entropy	Poly-1 loss: cross-entropy plus epsilon x (1 - p_t).	—	—
QualityFocal	(unit)	logit	tested	Generalized Focal Loss (Li et al., NeurIPS 2020); adapted from open-mmlab/mmdetection@cfd5d3a, Apache-2.0	Quality focal loss: sigmoid cross-entropy to a soft target, focused on large errors.	—	—
RobustLoss	(value, unit)	value	bridge	A General and Adaptive Robust Loss Function (Barron, CVPR 2019), special cases, via kornia.losses	Robust regression penalties of kornia: Charbonnier, Welsch, Cauchy, Geman-McClure.	kornia	—
SmoothL1	(value, unit)	value	—	PyTorch smooth_l1_loss	Huber-style penalty: quadratic below beta, linear above.	—	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e experimental/sbg_ptv3_g1b-realsense_18e finegrasp/finegrasp_sunet14_g1b-realsense_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
SoftmaxFocalLoss	(class)	value	bridge	Focal Loss for Dense Object Detection (Lin et al., ICCV 2017), via kornia.losses	Softmax focal loss: (1 - p_t)^gamma x cross-entropy over mutually exclusive classes.	kornia	—
TopK	(class, unit, value, bins)	value	tested	Online Hard Example Mining (Shrivastava et al., CVPR 2016); top-k of the penalties as in Contact-GraspNet (Sundermeyer et al., ICRA 2021), written here	Online hard example mining: only the hardest fraction of the elements counts.	—	—
UncertaintyWeighted	(class, unit, value, bins)	value	tested	Multi-Task Learning Using Uncertainty to Weigh Losses (Kendall et al., CVPR 2018); the weighting formula as in LibMTL (UW), one log-variance per wrapped term, written here	Uncertainty weighting: a learned factor on the term, which pays for being small.	—	—
Varifocal	(unit)	logit	tested	VarifocalNet (Zhang et al., CVPR 2021); adapted from open-mmlab/mmdetection@cfd5d3a, Apache-2.0	Varifocal loss: positives weighted by their target quality, negatives focused.	—	—

Post-processors

evaluation.variants.<name>[i] — Predictions to predictions at inference time. All 6 methods Options

14 registered classes

Component	Status	Source	What it is	Needs	Shipped configs
ApproachAngleFilter	tested	Generic; filter_approach_direction of GPD (ten Pas et al., arXiv 1706.09911), approach_steering of the AnyGrasp SDK, the gravity cone of ThinkGrasp (github.com/H-Freax/ThinkGrasp)	Keep grasps that approach within a cone about a reference direction.	—	—
FurthestGraspSampling	tested	Generic, on grasppanda.ops.furthest_point_sample; the idea is the grasp selection of Contact-GraspNet (Sundermeyer et al., ICRA 2021)	Pick grasps that are spread over the scene: furthest point sampling of grasp points.	—	—
GraspNMS	tested	Grasp NMS rule of graspnetAPI	Keep the best grasp of every cluster of near-duplicates.	—	experimental/hggd_dinov2-small-reg-depth-stem_g1b-realsense_15e hggd/hggd_r34w8_g1b-kinect_15e hggd/hggd_r34w8_g1b-realsense-to-gc6d-d435_15e hggd/hggd_r34w8_g1b-realsense_15e
GripperCentering	tested	Generic, written here from the description in AnyGrasp (Fang et al., arXiv 2212.08333, section IV-D)	Shift each grasp along its closing axis so that both fingers travel equally far.	—	—
GripperLimits	tested	Generic; max_gripper_width of the AnyGrasp SDK, min_aperture/max_aperture of GPD (arXiv 1706.09911)	Keep the grasps a real gripper can execute: opening and insertion depth in range.	—	—
GripperMeshCollisionFilter	bridge	Bridge to trimesh (Dawson-Haggerty et al., github.com/mikedh/trimesh, MIT): mesh loading, body split, watertightness, convex hull and box primitives; the point-cloud-against-gripper-mesh check of GraspGen (Murali et al., arXiv 2507.13097) as a containment test	Remove grasps at which a gripper mesh would contain observed points.	trimesh	—
MaskFilter	tested	Generic; filter_grasps of Contact-GraspNet (Sundermeyer et al., ICRA 2021) and apply_object_mask of the AnyGrasp SDK	Keep the grasps on a target: those near a point of a named mask of the observation.	—	—
ModelFreeCollisionFilter	tested	Collision filter of the GraspNet-baseline and HGGD test scripts	Remove grasps that collide with the observed cloud.	—	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e experimental/hggd_dinov2-small-reg-depth-stem_g1b-realsense_15e experimental/sbg_ptv3_g1b-realsense_18e finegrasp/finegrasp_sunet14_g1b-realsense_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e hggd/hggd_r34w8_g1b-kinect_15e hggd/hggd_r34w8_g1b-realsense-to-gc6d-d435_15e hggd/hggd_r34w8_g1b-realsense_15e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
ScoreThreshold	tested	Generic; choose_fn: better_than_threshold of 6-DOF GraspNet (ICCV 2019)	Keep grasps whose score reaches a threshold.	—	—
TablePlaneFilter	tested	Generic; the table test of the ZeroGrasp collision detector (Iwase et al., CVPR 2025) with a plane fitted to the cloud instead of a calibrated table frame	Remove grasps whose gripper would reach into the table.	—	—
ToFrame	tested	Generic; GraspGroup.transform with a calibration from the config	Express the grasps in another frame, for example the robot base.	—	—
TopK	tested	Generic; the top-k step of AnyGrasp (Fang et al., T-RO, arXiv 2212.08333, section IV-D) and of GPD	Keep the k best-scoring grasps, best first.	—	—
WidthFromCloud	tested	Generic; the idea of the contact-based width refinement of ZeroGrasp (Iwase et al., arXiv 2504.10857), on the sensor cloud alone	Set each opening to the extent of the scene between the fingers plus a clearance.	—	—
WorkspaceFilter	tested	Generic; the workspace limits of GPD (ten Pas et al., arXiv 1706.09911) and AnyGrasp SDK	Keep grasps whose grasp point lies inside an axis-aligned box.	—	—

Datasets

data.<split>.dataset — Frames with optional supervision.

2 registered classes

Component	Status	Source	What it is	Shipped configs
GraspClutter6D	tested	GraspClutter6D (arXiv 2504.06866)	GraspClutter6D frames as point clouds with optional supervision.	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_gc6d-d435_10e hggd/hggd_r34w8_g1b-realsense-to-gc6d-d435_15e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e
GraspNet1B	tested	GraspNet-1Billion (Fang et al., CVPR 2020)	GraspNet-1Billion frames as point clouds with optional supervision.	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e experimental/hggd_dinov2-small-reg-depth-stem_g1b-realsense_15e experimental/sbg_ptv3_g1b-realsense_18e finegrasp/finegrasp_sunet14_g1b-realsense_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e hggd/hggd_r34w8_g1b-kinect_15e hggd/hggd_r34w8_g1b-realsense-to-gc6d-d435_15e hggd/hggd_r34w8_g1b-realsense_15e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e

Transforms

data.<split>.dataset.preprocess[i] / .transforms[i] — Preprocessing and augmentation steps of a dataset.

16 registered classes

Component	Status	Source	What it is	Shipped configs
ConstantFeatures	—	GSNet (Wang et al., ICCV 2021)	Give every point the same input feature vector (value repeated channels times).	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e
CoordinateFeatures	tested	FineGrasp (arXiv 2507.05978)	Put the coordinates of every point in front of its input features.	experimental/finegrasp_ptv3_g1b-realsense_10e finegrasp/finegrasp_sunet14_g1b-realsense_10e
DepthNormals	tested	FineGrasp (arXiv 2507.05978)	Give every point the surface normal of its pixel as input features.	experimental/finegrasp_ptv3_g1b-realsense_10e finegrasp/finegrasp_sunet14_g1b-realsense_10e
DepthToCloud	—	GraspPanda	Back-project the depth image into an organised cloud; valid where depth > 0.	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e experimental/hggd_dinov2-small-reg-depth-stem_g1b-realsense_15e experimental/sbg_ptv3_g1b-realsense_18e finegrasp/finegrasp_sunet14_g1b-realsense_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e hggd/hggd_r34w8_g1b-kinect_15e hggd/hggd_r34w8_g1b-realsense-to-gc6d-d435_15e hggd/hggd_r34w8_g1b-realsense_15e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
FilterObjects	—	GSNet (Wang et al., ICCV 2021)	Drop object instances that are barely visible in the current point set.	experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e experimental/sbg_ptv3_g1b-realsense_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
NoWorkspace	—	GraspPanda	Keep every pixel with valid depth.	experimental/hggd_dinov2-small-reg-depth-stem_g1b-realsense_15e hggd/hggd_r34w8_g1b-kinect_15e hggd/hggd_r34w8_g1b-realsense-to-gc6d-d435_15e hggd/hggd_r34w8_g1b-realsense_15e
RandomDropout	tested	Qi et al., PointNet++ (arXiv 1706.02413); adapted from Pointcept/Pointcept@1342eda, MIT	With probability prob remove a fraction ratio of the points, labels included.	—
RandomFlip	—	GraspNet-1Billion reference pipelines	Mirror the scene along a camera axis with probability prob.	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e experimental/sbg_ptv3_g1b-realsense_18e finegrasp/finegrasp_sunet14_g1b-realsense_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
RandomJitter	tested	Qi et al., PointNet++ (arXiv 1706.02413); adapted from Pointcept/Pointcept@1342eda, MIT	Add clipped Gaussian noise to every input point (sensor noise).	—
RandomRotation	—	GraspNet-1Billion reference pipelines	Rotate the scene about a camera axis by a uniform angle in [-max_angle, max_angle).	experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e experimental/sbg_ptv3_g1b-realsense_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
RandomSample	—	GraspNet-1Billion reference pipelines	Reduce the cloud to exactly num_points points.	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e experimental/hggd_dinov2-small-reg-depth-stem_g1b-realsense_15e experimental/sbg_ptv3_g1b-realsense_18e finegrasp/finegrasp_sunet14_g1b-realsense_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e hggd/hggd_r34w8_g1b-kinect_15e hggd/hggd_r34w8_g1b-realsense-to-gc6d-d435_15e hggd/hggd_r34w8_g1b-realsense_15e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
RandomShift	tested	Pointcept (Wu et al.), RandomShift; adapted from Pointcept/Pointcept@1342eda, MIT	Translate the scene by a uniform offset per camera axis; object poses move along.	—
SampleGraspLabels	—	GraspNet-1Billion reference pipelines	Read the dense grasp labels of every remaining object, on a subset of its points.	experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e experimental/sbg_ptv3_g1b-realsense_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e gsnet/gsnet_sunet14_gc6d-d435_10e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
SegmentationBoxWorkspace	—	GraspNet-baseline (Fang et al., CVPR 2020)	Crop to the bounding box of the ground-truth foreground.	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e experimental/sbg_ptv3_g1b-realsense_18e finegrasp/finegrasp_sunet14_g1b-realsense_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_g1b-realsense_10e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
SegmentationRectWorkspace	—	graspclutter6dAPI	Crop to the image rectangle around the ground-truth foreground.	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_gc6d-d435_10e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e
TableBoxWorkspace	—	GraspPanda	Crop to a fixed box in the table frame.	—

Evaluators

evaluation.evaluator — Benchmark protocol: prediction dumps to scores.

2 registered classes

Component	Status	Source	What it is	Shipped configs
GraspClutter6DEvaluator	—	graspclutter6dAPI	GraspClutter6D single-view evaluation; see the module docstring for the protocol.	economicgrasp/economicgrasp_sunet14_g1b-kinect-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_g1b-realsense-to-gc6d-d435_10e economicgrasp/economicgrasp_sunet14_gc6d-d435_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e gsnet/gsnet_sunet14_g1b-realsense-to-gc6d-d435_10e gsnet/gsnet_sunet14_gc6d-d435_10e hggd/hggd_r34w8_g1b-realsense-to-gc6d-d435_15e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e
GraspNet1BEvaluator	—	graspnetAPI 1.2.x (Fang et al., CVPR 2020)	GraspNet-1Billion single-view evaluation, value-compatible with graspnetAPI 1.2.x.	economicgrasp/economicgrasp_sunet14_g1b-kinect_10e economicgrasp/economicgrasp_sunet14_g1b-realsense_10e experimental/economicgrasp_ptv3_g1b-realsense_10e experimental/finegrasp_ptv3_g1b-realsense_10e experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/gsnet_litept_g1b-realsense_10e experimental/gsnet_oacnns_g1b-realsense_10e experimental/gsnet_pointnet2-fp_g1b-realsense_10e experimental/gsnet_pointnext_g1b-realsense_10e experimental/gsnet_ptv3_g1b-realsense_10e experimental/gsnet_spunet_g1b-realsense_10e experimental/hggd_dinov2-small-reg-depth-stem_g1b-realsense_15e experimental/sbg_ptv3_g1b-realsense_18e finegrasp/finegrasp_sunet14_g1b-realsense_10e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e gsnet/gsnet_sunet14_g1b-realsense_10e hggd/hggd_r34w8_g1b-kinect_15e hggd/hggd_r34w8_g1b-realsense_15e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e

Hooks

hooks[i] — Extra behaviour around the training loop. All 6 methods Options

13 registered classes

Component	Status	Source	What it is	Needs	Shipped configs
AdaptiveGradClipHook	bridge	Brock et al. (arXiv 2102.06171), via timm.utils.adaptive_clip_grad	Adaptive gradient clipping: bound each unit's gradient by a fraction of its weight norm.	optim	—
BNMomentumHook	—	BN momentum schedule of the PointNet++-based methods	Decays the momentum of every batch-norm layer in steps of epochs.	—	experimental/graspnet-baseline_sunet14-fps_g1b-realsense_18e experimental/hggd_dinov2-small-reg-depth-stem_g1b-realsense_15e experimental/sbg_ptv3_g1b-realsense_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense-to-gc6d-d435_18e graspnet_baseline/graspnet-baseline_pointnet2_g1b-realsense_18e hggd/hggd_r34w8_g1b-kinect_15e hggd/hggd_r34w8_g1b-realsense-to-gc6d-d435_15e hggd/hggd_r34w8_g1b-realsense_15e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense-to-gc6d-d435_18e scale_balanced_grasp/sbg_pointnet2-transformer_g1b-realsense_18e
CheckpointHook	—	GraspPanda	Writes checkpoints/epoch_NNNN.pth and checkpoints/best.pth.	—	—
EMAHook	tested	Tarvainen and Valpola, NeurIPS 2017 (arXiv 1703.01780), via torch.optim.swa_utils	Exponential moving average of the weights, with warm-up of the decay.	—	—
EpochSignalHook	—	GraspPanda	Calls model.on_train_epoch(epoch, max_epochs) before every epoch.	—	experimental/hggd_dinov2-small-reg-depth-stem_g1b-realsense_15e hggd/hggd_r34w8_g1b-kinect_15e hggd/hggd_r34w8_g1b-realsense-to-gc6d-d435_15e hggd/hggd_r34w8_g1b-realsense_15e
EvaluationHook	—	GraspPanda	Predicts on data.val and scores every variant of evaluation.variants.	—	—
FreezeHook	tested	Howard and Ruder, ACL 2018 (arXiv 1801.06146): gradual unfreezing	Freeze sub-modules for the first epochs (or for the whole run), then release them.	—	—
SWAHook	tested	Izmailov et al., UAI 2018 (arXiv 1803.05407), via torch.optim.swa_utils	Stochastic weight averaging: the equal-weight mean of the weights after late epochs.	—	—
ScheduleFreeHook	tested	Defazio et al., The Road Less Scheduled (arXiv 2405.15682), via schedulefree	Switch a schedule-free optimizer between its training point and its average.	—	—
StageTimerHook	—	GraspPanda	Times the forward pass of each stage (direct child module) of the detector.	—	—
TensorBoardLoggerHook	—	GraspPanda	Writes scalars to <run dir>/tensorboard: train/* by iteration, eval/* by epoch.	tensorboard	—
TextLoggerHook	—	GraspPanda	Prints one line per interval to the console and to log.txt.	—	—
WandbLoggerHook	—	GraspPanda	Logs scalars to Weights & Biases; the run is named after the run directory.	wandb	—

Optimizers

optimizer — Every optimizer of torch.optim, by its class name. All 6 methods Options

Named by class, whatever the installed PyTorch provides; grasppanda components --slot optimizers lists them.

4 registered classes

Component	Status	Source	What it is	Needs	Shipped configs
ParamGroups	tested	Bao et al., BEiT (arXiv 2106.08254): layer-wise decay as in timm.optim	Parameter groups for any optimizer: per-module rates, layer-wise decay, no decay on 1-D.	—	—
PytorchOptimizer	bridge	Kim, pytorch_optimizer (kozistr/pytorch_optimizer), via the library	Any optimizer of pytorch_optimizer by its name (prodigy, ademamix, soap, …).	optim	—
ScheduleFree	bridge	Defazio et al., The Road Less Scheduled (arXiv 2405.15682), via schedulefree	A schedule-free optimizer: no learning-rate schedule, the iterates are averaged instead.	optim	—
TimmOptimizer	bridge	Wightman, PyTorch Image Models (timm.optim), via timm	Any optimizer of timm by its name (lamb, adan, adopt, lion, mars, …).	optim	—

Schedulers

scheduler — Every scheduler of torch.optim.lr_scheduler, by its class name. All 6 methods Options

Named by class, whatever the installed PyTorch provides; grasppanda components --slot schedulers lists them.

2 registered classes

Component	Status	Source	What it is	Needs	Shipped configs
TimmSchedule	bridge	Wightman, PyTorch Image Models (timm.scheduler), via timm	A schedule of timm.scheduler by name: cosine, tanh, polynomial, step, multi-step.	optim	—
WarmupCosineLR	tested	FineGrasp (arXiv 2507.05978)	Linear warm-up, then a cosine: the schedule of the FineGrasp reference.	—	experimental/finegrasp_ptv3_g1b-realsense_10e finegrasp/finegrasp_sunet14_g1b-realsense_10e