Skip to content

Does it run?

Which option trains on real frames under which method? Every option was composed under each method it fits; those accepted were probed on real frames. The few without a record of their own (two adapters compose inserts, two point-painting wrappers, post=target-mask and loss=distribution-focal, measured inside decoupled-grid-bins), the refusals and the options that need your consent are named per option.

A probe shows that a combination trains on real frames of GraspNet-1Billion, not how good it is. Accuracy is claimed only for a method trained and evaluated here.

The answer, slot by method

Slot GraspNet-baseline GSNet Scale-Balanced-Grasp EconomicGrasp FineGrasp HGGD
Image encoder — — — — — 36/37
Scene encoder 46/46 41/46 46/46 41/46 38/46 —
Seed proposal 21/23 23/23 21/23 23/23 23/23 —
View 20/20 20/20 20/20 18/20 18/20 —
Assigner 2/3 2/3 2/3 0/3 0/3 —
Grouper 25/27 26/27 25/27 26/27 26/27 —
Head bundle 7/9 9/9 7/9 9/9 9/9 —
Loss 17/18 20/20 15/17 20/20 20/20 —
Post-processing 14/14 14/14 14/14 14/14 14/14 14/14
Optimizer 17/17 16/17 17/17 16/17 16/17 14/15
Schedule 7/7 7/7 7/7 7/7 1/7 7/7
Hooks 5/7 6/7 5/7 6/7 5/7 4/5
Augmentation 3/3 3/3 3/3 3/3 2/3 1/3
Recipe 3/3 3/3 3/3 3/3 2/3 1/1
Slot GraspNet-baseline GSNet Scale-Balanced-Grasp EconomicGrasp FineGrasp HGGD
Image encoder — — — — — 6/37
Scene encoder 46/46 36/46 45/46 33/46 29/46 —
Seed proposal 14/23 23/23 19/23 23/23 20/23 —
View 13/20 16/20 15/20 14/20 14/20 —
Assigner 2/3 2/3 2/3 0/3 0/3 —
Grouper 24/27 26/27 25/27 25/27 26/27 —
Head bundle 7/9 5/9 7/9 6/9 5/9 —
Loss 14/18 19/20 13/17 20/20 18/20 —
Post-processing 14/14 14/14 14/14 14/14 14/14 8/14
Optimizer 11/17 16/17 12/17 13/17 14/17 10/15
Schedule 7/7 7/7 7/7 7/7 1/7 5/7
Hooks 5/7 5/7 5/7 5/7 4/7 3/5
Augmentation 3/3 3/3 3/3 3/3 2/3 0/3
Recipe 3/3 2/3 3/3 2/3 2/3 1/1
Slot GraspNet-baseline GSNet Scale-Balanced-Grasp EconomicGrasp FineGrasp HGGD
Image encoder — — — — — 34/37
Scene encoder 41/46 24/46 11/46 21/46 6/46 —
Seed proposal 21/23 23/23 6/23 23/23 8/23 —
View 19/20 19/20 20/20 18/20 11/20 —
Assigner 2/3 2/3 0/3 0/3 0/3 —
Grouper 25/27 26/27 11/27 26/27 26/27 —
Head bundle 7/9 9/9 5/9 9/9 9/9 —
Loss 17/18 20/20 2/17 20/20 11/20 —
Post-processing 14/14 14/14 0/14 14/14 4/14 14/14
Optimizer 17/17 16/17 6/17 16/17 16/17 14/15
Schedule 7/7 7/7 5/7 7/7 1/7 7/7
Hooks 5/7 6/7 2/7 6/7 1/7 4/5
Augmentation 3/3 3/3 1/3 3/3 0/3 1/3
Recipe 3/3 3/3 0/3 3/3 2/3 1/1

A cell counts the options that reached the level, out of those tried there as a swap (a method's own component is not counted); — : no such slot. Hover a cell for who is left out; below, why.

How it was judged

  • One cell, one option under one method: grasppanda compose <method config> --set <slot>=<option>, then grasppanda probe on the composed config. Measured on: 20 training iterations on 8 RealSense frames of GraspNet-1Billion (one scene), seed 0, one run per cell; nothing was measured on GraspClutter6D or on Kinect frames.
  • Four levels, each needing the one before: composes (config accepted), runs (trains on real frames without an error), learns (its losses fall), predicts (it ranks grasps). Exact definitions.
  • A refusal is the config check's answer: nothing was built, and the kind says why (refusal words).
  • Undecided counts against: a method where a level could not be decided stays in the count, named with the reason.
  • One run per cell: the config hash and the first losses repeat exactly; later numbers may differ in the next run.
What was run, and how it was judged in detail

1262 recorded cells (1014 runs, 174 runs-empty, 29 runs-unstable, 41 refused, 4 fails): every shipped method config as it is, every registered scene encoder under every seed-based method, every variant of every other stage with one stage replaced at a time, every image-encoder preset under HGGD, every post-processing preset under every method (but target-mask: reads observation.masks, which no shipped dataset fills), every recipe preset under every method it applies to, every loss preset in one term of each kind it accepts (but distribution-focal: fits bins terms only, which no shipped method has; probed as the width loss of the head preset decoupled-grid-bins (stage cells)), a sample of the sets of stages and a few requests that replace several slots at once. A component that needs an option only a preset supplies (DepthEmbedGroup, DepthPoolGroup, MultiRadiusGroup, MultiScaleGroup) has no cell under its type name: that request is refused for the missing option before any contract is compared, which says nothing about a combination. Its presets have cells.

Group Cells recorded runs runs-empty runs-unstable refused fails
reference: the methods as shipped 5 of 5 4 1 — — —
encoder: one scene encoder replaced 235 of 235 127 90 18 — —
stage: one other stage replaced 435 of 435 382 25 — 28 —
stage-set: sets of stages 95 of 9104 86 9 — — —
family: other detector families, as shipped 1 of 1 1 — — — —
family-slot: one component of another detector family replaced 38 of 38 35 2 — 1 —
post: one post-processing step as the variant post 84 of 84 70 14 — — —
loss: one loss of one stage replaced 134 of 134 119 11 — 4 —
recipe: one recipe preset applied (optimizer, schedule, hooks, augmentation) 217 of 217 179 15 11 8 4
cross: options of several slots replaced together 18 of 18 11 7 — — —
all 1262 of 10271 1014 174 29 41 4
  • A cell: grasppanda compose <method config> --set <slot>=<choice> (so an encoder is fitted and adapters are inserted exactly as a user gets them), then grasppanda probe <composed config> --iters 20 --frames 8 --schedule constant: the real trainer with the method's own recipe on the split sample of GraspNet-1Billion (RealSense; scene 0, frames 0, 32, …, 224), then prediction on two of those frames. Both commands of every cell are in the records.
  • The learning rate: the recipe's peak rate at every step (--schedule constant), scaled by the ratio of the batch sizes when the probe's batches are smaller (FineGrasp's 0.008 for 32 frames trains at 0.002 on 8); the optimizer, weight decay, clipping and hooks as the recipe has them. Over the few epochs of a probe the schedules would decay some methods far more than others, and the trends would not compare. A request that sets the schedule (a schedule preset, a schedule-free optimizer, a recipe preset) runs that schedule. Of the 1182 probes on this page, 1103 held the peak rate (schedule: constant in the record); 79 ran the recipe's schedule (schedule: recipe).
  • One exception to "as a user gets them": configs that read the width histogram of Scale-Balanced-Grasp were probed with the reference's own file (scale_statistic.source: legacy_import): the native histogram is a statistic of the whole training split, and a probe reads eight frames.
  • Two cells with one config are one experiment: a cell that composes to the config of another cell (the same config hash) shares its probe; 39 of the 1262 cells do. A cell that composes to its method's shipped config is the method's own component and is not counted as a swap.
  • When and where: 3 cells on 2026-10-15 with GraspPanda at commit 0ddca31, PyTorch 2.7.1+cu118, Python 3.12.7, on one NVIDIA RTX A6000 shared with other jobs, 2 probes at a time. 288 cells on 2026-10-15 with GraspPanda at commit 2a3760a, PyTorch 2.7.1+cu118, Python 3.12.7, on one NVIDIA RTX A6000 shared with other jobs, 2 probes at a time. 538 cells on 2026-10-15 with GraspPanda at commit 96facee, PyTorch 2.7.1+cu118, Python 3.12.7, on one NVIDIA RTX A6000 shared with other jobs, 2 probes at a time. 430 cells on 2026-10-15 with GraspPanda at commit 98e5310, PyTorch 2.7.1+cu118, Python 3.12.7, on one NVIDIA RTX A6000 shared with other jobs, 2 probes at a time. 3 cells on 2026-10-16 with GraspPanda at commit 96facee, PyTorch 2.7.1+cu118, Python 3.12.7, on one NVIDIA RTX A6000 shared with other jobs, 2 probes at a time.

1221 of the 1262 records on this page were judged by judge 2 (judge: 2 in the records), which also records whether the parts of a model learned: the trend of every loss term, the units no gradient reached, the parts that did not change and the spread of the scores. A record of the first judge has none of it, so its learns and predicts are undecided (not checked). The loss of every iteration of every probe of judge 2 is in measured_combinations/losses_per_iteration.json.

The flags of a loss term, named in the tails (active iterations: the term was reported, finite and not exactly 0):

  • flat: the term was active in at least two iterations, and the mean of its last quarter differs from its first active value by less than 5 % of that value.
  • rising: the term was active in at least two iterations, and the mean of the last half of its active iterations exceeds the mean of the first half by more than 10 % of the first half's mean (with an odd count the middle iteration is in neither half).
  • zero_at_end: the term is exactly 0.0, or not reported, at the last completed iteration.
  • constant: the term was reported in at least two iterations with the very same value in every one of them, zeros included.

The schedules of a probe (schedule in a record):

  • recipe: the config's own scheduler, run over the probe. The probe sets train.epochs to ceil(iterations / iterations per epoch), so a schedule that steps per epoch is compressed into those few epochs (GSNet's 0.95 per epoch ends at 0.95^9 = 0.63 after 20 iterations of 2 per epoch, EconomicGrasp's cosine at 2.4 % of its start) and one that steps per iteration runs its first steps only (FineGrasp's warm-up of 800 iterations keeps a probe of 20 at no more than 2.5 % of its rate).
  • constant: no scheduler: every parameter group of every optimizer trains at the recipe's peak rate for the whole probe. The peak is the rate the optimizer gives the group (the configured lr, with the multipliers of ParamGroups), except for a schedule that sets a peak of its own: max_lr of OneCycleLR and CyclicLR, the larger factor of LinearLR. Every other schedule of the toolbox starts at that rate or reaches it at the end of its warm-up. When the probe's batches hold fewer frames than the recipe's, the peak is scaled by their ratio (the linear scaling rule: FineGrasp's 0.008 for 32 frames trains at 0.002 on batches of 8). A batch-norm momentum that the recipe decays per epoch (BNMomentumHook: GraspNet-baseline, Scale-Balanced-Grasp, HGGD) is held at its first epoch's value likewise (hooks.<i>.decay=1): decayed over epochs of a few iterations it can reach its floor of 0.001 within the probe (Scale-Balanced-Grasp's at iteration 73), and the running statistics, which prediction normalises with, then stop following the weights; training itself does not read them. The optimizer (its weight decay and its momentum as configured), clipping, accumulation and every other hook are the recipe's.

Passed the check, failed afterwards.

4 cell(s) passed the config check and failed afterwards: each group is a gap between what the config check knows and what the components need (N stands for a number that differs from cell to cell).

  • at step: RuntimeError: CUDA error: an illegal memory access was encountered CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. 1 cell(s): recipe/HGGD/optimizer/soap.
  • at hook, in the hook SWAHook: EngineError: SWAHook.start_epoch is N, but the run has N epochs: nothing would be averaged. 2 cell(s): recipe/GraspNet-baseline/hooks/swa, recipe/Scale-Balanced-Grasp/hooks/swa.
  • at data, in the transform DepthNormals: ContractError: DepthNormals must run before the cloud of graspnet1b is sampled: put it first in transforms. 1 cell(s): recipe/FineGrasp/augment/dropout.

Where each option works

Each section counts the options a user can --set (244 in all). In its table, a chip counts the methods under which the option reached the level, out of those it was tried under as a swap; trains like shipped is level 2. The list under it says who is left out, and why.

Image encoders

38 options 36 run everywhere 1 undecided 1 as shipped 23 of 36 train like shipped

Image encoders: 38 options, one row each

Compare in the explorer: the same options, cell by cell.

Option Runs Learns Predicts Trains like shipped
convnextv2-atto HGGD ✓ HGGD ? HGGD ✗ HGGD: worse
convnextv2-tiny HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
dinov2-small-reg-depth-stem HGGD ✓ HGGD ✓ HGGD ✓ HGGD: on a par
dinov2-small-reg HGGD ✓ HGGD ? HGGD ✓ HGGD: worse
dinov3-convnext-tiny HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
dinov3-vits16 HGGD ✓ HGGD ? HGGD ✓ HGGD: diverging
efficientnetv2-s HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
efficientvit-b1 HGGD ✓ HGGD ? HGGD ✓ HGGD: worse
eva02-tiny HGGD ✓ HGGD ? HGGD ✓ HGGD: worse
fastvit-sa12 HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
fastvit-t8 HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
hf-convnextv2-tiny HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
hf-depth-anything-v2-small HGGD ✓ HGGD ✓ HGGD ✓ HGGD: worse
hf-dinov2-small-reg HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
hf-dinov2-small HGGD ✓ HGGD ✓ HGGD ✓ HGGD: worse
hf-hiera-tiny HGGD ✓ HGGD ? HGGD ✓ HGGD: worse
hf-radio-v3-base HGGD ? HGGD ? HGGD ? not compared
hf-siglip2-base HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
hf-swinv2-tiny HGGD ✓ HGGD ✓ HGGD ✓ HGGD: on a par
hggd-residual own only own only own only not compared
hiera-small-abswin HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
maxvit-tiny HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
mobilenetv4-conv-small HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
mobilenetv4-hybrid-medium HGGD ✓ HGGD ? HGGD ✗ HGGD: on a par
pe-spatial-tiny HGGD ✓ HGGD ? HGGD ✓ HGGD: worse
repvit-m1 HGGD ✓ HGGD ? HGGD ✓ HGGD: worse
resnet18 HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
resnet34-rgbd HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
resnet34 HGGD ✓ HGGD ✓ HGGD ✓ HGGD: on a par
resnet50 HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
sam2-hiera-tiny HGGD ✓ HGGD ? HGGD ✓ HGGD: worse
siglip2-base HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
swinv2-tiny HGGD ✓ HGGD ? HGGD ✓ HGGD: worse
tv-convnext-tiny HGGD ✓ HGGD ? HGGD ✓ HGGD: worse
tv-efficientnetv2-s HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
tv-mobilenetv3-large HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par
tv-regnet-y-800mf HGGD ✓ HGGD ✓ HGGD ✓ HGGD: on a par
tv-resnet34 HGGD ✓ HGGD ? HGGD ✓ HGGD: on a par

Who is left out, and why:

  • convnextv2-atto: learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: not HGGD (does not predict: no grasp on any frame)
  • convnextv2-tiny: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • dinov2-small-reg: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • dinov3-convnext-tiny: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • dinov3-vits16: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • efficientnetv2-s: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • efficientvit-b1: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • eva02-tiny: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • fastvit-sa12: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • fastvit-t8: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • hf-convnextv2-tiny: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • hf-dinov2-small-reg: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • hf-hiera-tiny: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • hf-radio-v3-base: runs: HGGD undecided (needs your consent (remote code))
  • hf-siglip2-base: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • hggd-residual: runs: own in HGGD
  • hiera-small-abswin: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • maxvit-tiny: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • mobilenetv4-conv-small: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • mobilenetv4-hybrid-medium: learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: not HGGD (does not predict: no grasp on any frame)
  • pe-spatial-tiny: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • repvit-m1: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • resnet18: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • resnet34-rgbd: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • resnet50: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • sam2-hiera-tiny: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • siglip2-base: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • swinv2-tiny: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • tv-convnext-tiny: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • tv-efficientnetv2-s: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • tv-mobilenetv3-large: learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • tv-resnet34: learns: HGGD undecided (the method as shipped does not reach it in a short run either)

Scene encoders

47 options 37 run everywhere 10 under some 9 of 47 train like shipped

Scene encoders: 47 options, one row each

Compare in the explorer: the same options, cell by cell.

Option Runs Learns Predicts Trains like shipped
assanet 5/5 4/5 3/5 5/5 on a par
dgcnn-static 5/5 5/5 2/5 1/5 on a par · 4 worse
dgcnn 5/5 5/5 1/5 0/5 on a par · 1 diverging · 4 worse
dela 5/5 4/5 2/5 4/5 on a par · 1 diverging
litept-structured3d 5/5 5/5 4/5 3/5 on a par · 2 worse
litept 5/5 5/5 4/5 5/5 on a par
minkunet18a-openscene 5/5 3/5 3/5 4/5 on a par · 1 worse
minkunet18a 5/5 4/5 3/5 4/5 on a par · 1 worse
minkunet34c-pointcontrast 5/5 4/5 3/5 5/5 on a par
minkunet34c 5/5 4/5 3/5 4/5 on a par · 1 worse
oacnns 2/5 2/5 1/5 2/2 on a par
octformer 5/5 4/5 3/5 4/5 on a par · 1 worse
ptv3-encoder 2/5 2/5 2/5 2/2 on a par
ptv3-ppt 5/5 3/5 5/5 3/5 on a par · 2 diverging
ptv3-scannet200 4/5 3/5 2/5 4/4 on a par
ptv3 5/5 5/5 5/5 4/5 on a par · 1 worse
ptv3-sonata-small 2/5 2/5 2/5 2/2 on a par
ptv3-sonata 3/5 2/5 1/5 2/3 on a par · 1 diverging
ptv3-utonia 3/5 2/5 1/5 1/3 on a par · 2 diverging
pointmlp 5/5 5/5 3/5 1/5 on a par · 1 diverging · 3 worse
point-mamba 5/5 5/5 3/5 2/5 on a par · 3 worse
pointmetabase-xl 5/5 5/5 1/5 3/5 on a par · 1 diverging · 1 worse
pointmetabase 5/5 5/5 3/5 3/5 on a par · 2 worse
pointnext-b 5/5 5/5 0/5 3/5 on a par · 2 worse
pointnext-l 5/5 5/5 0/5 1/5 on a par · 4 worse
pointnext-xl 5/5 5/5 1/5 2/5 on a par · 3 worse
pointnext 5/5 5/5 1/5 3/5 on a par · 2 worse
pointnet2 4/4 4/4 2/4 0/4 on a par · 4 worse
pointnet2-dense 5/5 5/5 1/5 2/5 on a par · 3 worse
pointnet2-msg 5/5 4/5 4/5 2/5 on a par · 3 worse
pointnet2-msg-dense 5/5 5/5 4/5 3/5 on a par · 2 worse
pointnet2-transformer 4/4 4/4 2/4 1/4 on a par · 3 worse
point-transformer-v1 5/5 3/5 1/5 1/5 on a par · 4 diverging
point-transformer-v2 5/5 5/5 1/5 1/5 on a par · 4 diverging
pointvector-xl 5/5 5/5 4/5 3/5 on a par · 2 worse
pointvector 5/5 5/5 1/5 2/5 on a par · 3 worse
point-vit 5/5 4/5 1/5 0/5 on a par · 5 worse
pyg-edgeconv 4/5 4/5 2/5 1/4 on a par · 3 worse
pyg-point-transformer 4/5 4/5 2/5 0/4 on a par · 4 worse
pyg-pointgnn 4/5 4/5 1/5 0/4 on a par · 4 worse
pyg-pointnet 5/5 5/5 2/5 0/5 on a par · 5 worse
randlanet 5/5 5/5 1/5 1/5 on a par · 3 diverging · 1 worse
spunet-gaussiancross 5/5 2/5 5/5 3/5 on a par · 1 diverging · 1 worse
spunet 4/5 4/5 3/5 4/4 on a par
sparse-unet14 2/2 2/2 1/2 2/2 on a par
uni3d 5/5 5/5 2/5 1/5 on a par · 1 diverging · 3 worse
volt 5/5 2/5 1/5 2/5 on a par · 1 diverging · 2 worse

Who is left out, and why:

  • assanet: learns: not FineGrasp (does not learn: score rising) · predicts: not EconomicGrasp (does not predict: no grasp on any frame); FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • dgcnn-static: predicts: not EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp
  • dgcnn: predicts: not GSNet, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, EconomicGrasp, FineGrasp; diverging under Scale-Balanced-Grasp
  • dela: learns: not FineGrasp (does not learn: score rising) · predicts: not GSNet, EconomicGrasp (does not predict: no grasp on any frame); FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under FineGrasp
  • litept-structured3d: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp, FineGrasp
  • litept: predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • minkunet18a-openscene: learns: not GSNet (does not learn: objectness flat); not FineGrasp (does not learn: score rising) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp
  • minkunet18a: learns: not EconomicGrasp (does not learn: width flat) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp
  • minkunet34c-pointcontrast: learns: not EconomicGrasp (does not learn: width flat) · predicts: not EconomicGrasp (does not predict: no grasp on any frame); FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • minkunet34c: learns: not GSNet (does not learn: objectness flat) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp
  • oacnns: runs: not GSNet, EconomicGrasp, FineGrasp (unstable twice in the short probe) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: not compared under GSNet, EconomicGrasp, FineGrasp: the swap did not run at level 1
  • octformer: learns: not FineGrasp (does not learn: score rising) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp
  • ptv3-encoder: runs: not GSNet, EconomicGrasp, FineGrasp (unstable twice in the short probe) · trains like shipped: not compared under GSNet, EconomicGrasp, FineGrasp: the swap did not run at level 1
  • ptv3-ppt: learns: not GSNet, EconomicGrasp (does not learn: objectness flat) · trains like shipped: diverging under Scale-Balanced-Grasp, FineGrasp
  • ptv3-scannet200: runs: not FineGrasp (unstable twice in the short probe) · learns: not EconomicGrasp (does not learn: width rose 10 %) · predicts: not EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: not compared under FineGrasp: the swap did not run at level 1
  • ptv3: trains like shipped: worse under FineGrasp
  • ptv3-sonata-small: runs: not GSNet, EconomicGrasp, FineGrasp (unstable twice in the short probe) · trains like shipped: not compared under GSNet, EconomicGrasp, FineGrasp: the swap did not run at level 1
  • ptv3-sonata: runs: not GSNet, EconomicGrasp (unstable twice in the short probe) · learns: not FineGrasp (does not learn: objectness rising; graspness rose 1230 %) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under FineGrasp; not compared under GSNet, EconomicGrasp: the swap did not run at level 1
  • ptv3-utonia: runs: not GSNet, EconomicGrasp (unstable twice in the short probe) · learns: not FineGrasp (does not learn: objectness rising; graspness rose 1023 %) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under Scale-Balanced-Grasp, FineGrasp; not compared under GSNet, EconomicGrasp: the swap did not run at level 1
  • pointmlp: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp, FineGrasp; diverging under Scale-Balanced-Grasp
  • point-mamba: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp, FineGrasp
  • pointmetabase-xl: predicts: not GSNet, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under Scale-Balanced-Grasp; worse under FineGrasp
  • pointmetabase: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp, FineGrasp
  • pointnext-b: predicts: not GraspNet-baseline (does not predict: no grasp on frame 2); not GSNet, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp, FineGrasp
  • pointnext-l: predicts: not GraspNet-baseline, GSNet, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp
  • pointnext-xl: predicts: not GSNet, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp, FineGrasp
  • pointnext: predicts: not GSNet, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp, FineGrasp
  • pointnet2: runs: own in GraspNet-baseline · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp
  • pointnet2-dense: predicts: not GSNet, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp, FineGrasp
  • pointnet2-msg: learns: not EconomicGrasp (does not learn: width rose 17 %) · predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp, FineGrasp
  • pointnet2-msg-dense: predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp, FineGrasp
  • pointnet2-transformer: runs: own in Scale-Balanced-Grasp · predicts: not GSNet (does not predict: no grasp on any frame); FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp, FineGrasp
  • point-transformer-v1: learns: not Scale-Balanced-Grasp (does not learn: score rose 16 %); not FineGrasp (does not learn: angle rising; depth rising) · predicts: not GraspNet-baseline, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp
  • point-transformer-v2: predicts: not GraspNet-baseline, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp
  • pointvector-xl: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp, FineGrasp
  • pointvector: predicts: not GSNet, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp, FineGrasp
  • point-vit: learns: not EconomicGrasp (does not learn: width flat) · predicts: not GraspNet-baseline, GSNet (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp
  • pyg-edgeconv: runs: not FineGrasp (unstable twice in the short probe) · predicts: not GSNet (does not predict: no grasp on any frame); Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, Scale-Balanced-Grasp, EconomicGrasp; not compared under FineGrasp: the swap did not run at level 1
  • pyg-point-transformer: runs: FineGrasp undecided (unstable once, not repeated) · predicts: not GSNet (does not predict: no grasp on any frame); Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp; not compared under FineGrasp: the swap did not run at level 1
  • pyg-pointgnn: runs: FineGrasp undecided (unstable once, not repeated) · predicts: not GSNet (does not predict: no grasp on frame 1); not EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp; not compared under FineGrasp: the swap did not run at level 1
  • pyg-pointnet: predicts: not GSNet, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp
  • randlanet: predicts: not GSNet, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under GSNet, EconomicGrasp, FineGrasp; worse under Scale-Balanced-Grasp
  • spunet-gaussiancross: learns: not GSNet, EconomicGrasp (does not learn: objectness flat); not FineGrasp (does not learn: view rising; angle rising) · trains like shipped: worse under GSNet; diverging under Scale-Balanced-Grasp
  • spunet: runs: not FineGrasp (unstable twice in the short probe) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: not compared under FineGrasp: the swap did not run at level 1
  • sparse-unet14: runs: own in GSNet, EconomicGrasp, FineGrasp · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • uni3d: predicts: not EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp, FineGrasp; diverging under Scale-Balanced-Grasp
  • volt: learns: not GSNet, EconomicGrasp, FineGrasp (does not learn: objectness flat) · predicts: not GSNet, EconomicGrasp (does not predict: no grasp on any frame); Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under Scale-Balanced-Grasp, FineGrasp; diverging under EconomicGrasp

Not options, so not counted:

  • feature-propagation: an adapter: compose inserts it where the next stage needs it, so it is measured inside the cells it was inserted in; inserted in 65 cells (43 runs, 18 runs-empty, 4 refused).
  • furthest-point-subsample: an adapter: compose inserts it where the next stage needs it, so it is measured inside the cells it was inserted in; inserted in 90 cells (54 runs, 36 runs-empty).
  • image-painted-subsampled: point painting: a wrapper written into a config by hand (compose does not write it); no probe record.
  • image-painted: point painting: a wrapper written into a config by hand (compose does not write it); no probe record.

Seed proposals

24 options 22 run everywhere 2 under some 1 of 24 train like shipped

Seed proposals: 24 options, one row each

Compare in the explorer: the same options, cell by cell.

Option Runs Learns Predicts Trains like shipped
dense-ffps 5/5 5/5 4/5 4/5 on a par · 1 worse
dense-fps 5/5 5/5 5/5 4/5 on a par · 1 worse
dense-fpsample 5/5 5/5 5/5 4/5 on a par · 1 worse
dense-fusion-fps 5/5 5/5 3/5 2/5 on a par · 3 worse
dense-random 5/5 5/5 4/5 4/5 on a par · 1 worse
dense-torch-cluster 5/5 5/5 4/5 4/5 on a par · 1 worse
dense-voxel 5/5 5/5 3/5 3/5 on a par · 2 worse
graspness-ffps 5/5 4/5 3/5 3/5 on a par · 1 diverging · 1 worse
graspness-fps-fill 5/5 4/5 3/5 3/5 on a par · 2 worse
graspness-fps-topk 5/5 3/5 4/5 2/5 on a par · 3 worse
graspness-fps 2/2 1/2 1/2 0/2 on a par · 2 worse
graspness-fpsample 5/5 5/5 4/5 3/5 on a par · 2 worse
graspness-fusion-fps 5/5 3/5 5/5 2/5 on a par · 3 worse
graspness-nms 5/5 4/5 3/5 0/5 on a par · 5 worse
graspness-random 5/5 5/5 4/5 3/5 on a par · 2 worse
graspness-sfps 5/5 5/5 3/5 3/5 on a par · 2 worse
graspness-top-fraction 5/5 4/5 3/5 1/5 on a par · 4 worse
graspness-topk 5/5 2/5 3/5 0/5 on a par · 5 worse
graspness-torch-cluster 5/5 5/5 4/5 3/5 on a par · 2 worse
graspness-voxel 5/5 5/5 3/5 0/5 on a par · 1 diverging · 4 worse
objectness-fps 3/5 3/5 3/5 2/3 on a par · 1 worse
product-fps 3/5 3/5 2/5 3/3 on a par
scored-fps 5/5 5/5 3/5 3/5 on a par · 2 worse
SubsampledSeeds (defaults) 3/3 3/3 2/3 1/3 on a par · 2 worse

Who is left out, and why:

  • dense-ffps: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp
  • dense-fps: trains like shipped: worse under FineGrasp
  • dense-fpsample: trains like shipped: worse under FineGrasp
  • dense-fusion-fps: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, Scale-Balanced-Grasp, FineGrasp
  • dense-random: predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp
  • dense-torch-cluster: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp
  • dense-voxel: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under Scale-Balanced-Grasp, FineGrasp
  • graspness-ffps: learns: not GraspNet-baseline (does not learn: objectness flat) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under GraspNet-baseline; worse under Scale-Balanced-Grasp
  • graspness-fps-fill: learns: not GraspNet-baseline (does not learn: objectness flat) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, Scale-Balanced-Grasp
  • graspness-fps-topk: learns: not GraspNet-baseline (does not learn: objectness flat); not FineGrasp (does not learn: angle rising) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, Scale-Balanced-Grasp
  • graspness-fps: runs: own in GSNet, EconomicGrasp, FineGrasp · learns: not GraspNet-baseline (does not learn: objectness flat) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, Scale-Balanced-Grasp
  • graspness-fpsample: predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, Scale-Balanced-Grasp
  • graspness-fusion-fps: learns: not GraspNet-baseline (does not learn: objectness flat); not FineGrasp (does not learn: angle rising) · trains like shipped: worse under GraspNet-baseline, Scale-Balanced-Grasp, FineGrasp
  • graspness-nms: learns: not GraspNet-baseline (does not learn: objectness flat; angle flat) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp
  • graspness-random: predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, Scale-Balanced-Grasp
  • graspness-sfps: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, Scale-Balanced-Grasp
  • graspness-top-fraction: learns: not Scale-Balanced-Grasp (does not learn: score rising; tolerance flat) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp
  • graspness-topk: learns: not GraspNet-baseline (does not learn: objectness rose 9 %); not Scale-Balanced-Grasp (does not learn: score rose 12 %; tolerance rose 53 %); not FineGrasp (does not learn: angle rising) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp
  • graspness-torch-cluster: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, Scale-Balanced-Grasp
  • graspness-voxel: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, EconomicGrasp, FineGrasp; diverging under Scale-Balanced-Grasp
  • objectness-fps: runs: not GraspNet-baseline, Scale-Balanced-Grasp (incompatible by design) · trains like shipped: worse under FineGrasp
  • product-fps: runs: not GraspNet-baseline, Scale-Balanced-Grasp (incompatible by design) · predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • scored-fps: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, Scale-Balanced-Grasp
  • SubsampledSeeds (defaults): runs: own in GraspNet-baseline, Scale-Balanced-Grasp · predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp, FineGrasp

View stages

21 options 19 run everywhere 2 under some 3 of 21 train like shipped

View stages: 21 options, one row each

Compare in the explorer: the same options, cell by cell.

Option Runs Learns Predicts Trains like shipped
direction-regression 5/5 4/5 4/5 1/5 on a par · 1 diverging · 3 worse
direction-vmf 5/5 5/5 5/5 1/5 on a par · 4 worse
healpix-192 5/5 5/5 4/5 3/5 on a par · 2 worse
healpix-768 5/5 5/5 4/5 5/5 on a par
icosphere-162 5/5 5/5 5/5 3/5 on a par · 2 worse
icosphere-642 5/5 4/5 4/5 5/5 on a par
lattice-angular-cost 5/5 4/5 5/5 1/5 on a par · 4 worse
lattice-best-view 5/5 4/5 5/5 1/5 on a par · 4 worse
lattice-fibonacci-120 5/5 5/5 4/5 5/5 on a par
lattice-hemisphere-60 5/5 5/5 4/5 4/5 on a par · 1 worse
lattice-soft-argmax 5/5 5/5 5/5 4/5 on a par · 1 worse
lattice-topk 5/5 4/5 5/5 2/5 on a par · 3 worse
lattice-view 5/5 5/5 4/5 4/5 on a par · 1 worse
objectness-view-graspable 2/4 2/4 0/4 0/2 on a par · 2 worse
objectness-view 2/4 1/4 2/4 0/2 on a par · 1 diverging · 1 worse
rotation-6d 5/5 2/5 5/5 1/5 on a par · 1 diverging · 3 worse
rotation-9d 5/5 2/5 5/5 1/5 on a par · 1 diverging · 3 worse
rotation-quaternion 5/5 1/5 5/5 1/5 on a par · 4 worse
template-view 2/2 1/2 2/2 0/2 on a par · 2 worse
two-vector-baseline 5/5 2/5 5/5 1/5 on a par · 4 worse
two-vector 5/5 1/5 5/5 1/5 on a par · 4 worse

Who is left out, and why:

  • direction-regression: learns: not GraspNet-baseline (does not learn: score rose 11 %) · predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp, FineGrasp; diverging under Scale-Balanced-Grasp
  • direction-vmf: trains like shipped: worse under GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp
  • healpix-192: predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, Scale-Balanced-Grasp
  • healpix-768: predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • icosphere-162: trains like shipped: worse under GraspNet-baseline, Scale-Balanced-Grasp
  • icosphere-642: learns: not EconomicGrasp (does not learn: width flat) · predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • lattice-angular-cost: learns: not FineGrasp (does not learn: angle rose 42 %) · trains like shipped: worse under GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp
  • lattice-best-view: learns: not FineGrasp (does not learn: angle rose 40 %) · trains like shipped: worse under GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp
  • lattice-fibonacci-120: predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • lattice-hemisphere-60: predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp
  • lattice-soft-argmax: trains like shipped: worse under Scale-Balanced-Grasp
  • lattice-topk: learns: not FineGrasp (does not learn: angle rising; depth rising) · trains like shipped: worse under GSNet, Scale-Balanced-Grasp, FineGrasp
  • lattice-view: predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under Scale-Balanced-Grasp
  • objectness-view-graspable: runs: own in Scale-Balanced-Grasp; not EconomicGrasp, FineGrasp (incompatible by design) · predicts: not GraspNet-baseline (does not predict: no grasp on frame 1); not GSNet (does not predict: no grasp on any frame) · trains like shipped: worse under GraspNet-baseline, GSNet
  • objectness-view: runs: own in GraspNet-baseline; not EconomicGrasp, FineGrasp (incompatible by design) · learns: not Scale-Balanced-Grasp (does not learn: angle flat) · trains like shipped: diverging under GSNet; worse under Scale-Balanced-Grasp
  • rotation-6d: learns: not GraspNet-baseline (does not learn: view_frame rose 6 %); not GSNet, Scale-Balanced-Grasp (does not learn: view_frame rising) · trains like shipped: worse under GSNet, EconomicGrasp, FineGrasp; diverging under Scale-Balanced-Grasp
  • rotation-9d: learns: not GraspNet-baseline (does not learn: view_frame rose 35 %); not GSNet (does not learn: view_frame flat); not FineGrasp (does not learn: view_frame rising) · trains like shipped: worse under GSNet, EconomicGrasp, FineGrasp; diverging under Scale-Balanced-Grasp
  • rotation-quaternion: learns: not GraspNet-baseline (does not learn: view_frame rose 14 %); not GSNet (does not learn: view_frame rose 6 %); not Scale-Balanced-Grasp (does not learn: view_frame rose 9 %); not EconomicGrasp (does not learn: view_frame flat) · trains like shipped: worse under GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp
  • template-view: runs: own in GSNet, EconomicGrasp, FineGrasp · learns: not GraspNet-baseline (does not learn: angle flat) · trains like shipped: worse under GraspNet-baseline, Scale-Balanced-Grasp
  • two-vector-baseline: learns: not GraspNet-baseline (does not learn: view_frame rising; score rose 31 %); not Scale-Balanced-Grasp (does not learn: view_frame flat); not EconomicGrasp (does not learn: score flat) · trains like shipped: worse under GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp
  • two-vector: learns: not GraspNet-baseline (does not learn: view_frame rose 30 %; score rose 91 %); not GSNet (does not learn: view_frame rising); not Scale-Balanced-Grasp (does not learn: view rising; view_frame flat); not EconomicGrasp (does not learn: view_frame rising; score flat) · trains like shipped: worse under GraspNet-baseline, GSNet, EconomicGrasp, FineGrasp

Assigners

4 options 0 run everywhere 3 under some 1 under none 1 of 3 train like shipped

Assigners: 4 options, one row each

Compare in the explorer: the same options, cell by cell.

Option Runs Learns Predicts Trains like shipped
compact-label 0/3 0/3 0/3 not compared
nearest-label-object-seeds 2/4 2/4 1/4 1/1 on a par
scale-balanced 2/4 2/4 2/4 0/1 on a par · 1 worse
NearestLabelAssigner (defaults) 2/4 2/4 1/4 0/1 on a par · 1 diverging

Who is left out, and why:

  • compact-label: runs: own in EconomicGrasp, FineGrasp; not GraspNet-baseline, GSNet, Scale-Balanced-Grasp (incompatible by design)
  • nearest-label-object-seeds: runs: own in GraspNet-baseline; not EconomicGrasp, FineGrasp (incompatible by design) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: not comparable under Scale-Balanced-Grasp
  • scale-balanced: runs: own in Scale-Balanced-Grasp; not EconomicGrasp, FineGrasp (incompatible by design) · trains like shipped: not comparable under GraspNet-baseline; worse under GSNet
  • NearestLabelAssigner (defaults): runs: own in GSNet; not EconomicGrasp, FineGrasp (incompatible by design) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: not comparable under GraspNet-baseline; diverging under Scale-Balanced-Grasp

Groupers

28 options 25 run everywhere 3 under some 6 of 28 train like shipped

Groupers: 28 options, one row each

Compare in the explorer: the same options, cell by cell.

Option Runs Learns Predicts Trains like shipped
ball 5/5 5/5 4/5 4/5 on a par · 1 worse
box 5/5 5/5 5/5 4/5 on a par · 1 worse
cylinder-attentive 5/5 5/5 4/5 4/5 on a par · 1 worse
cylinder-edgeconv 5/5 5/5 5/5 4/5 on a par · 1 diverging
cylinder-invres 5/5 5/5 4/5 3/5 on a par · 1 diverging · 1 worse
cylinder-vector-attention 5/5 5/5 4/5 4/5 on a par · 1 diverging
depth-ball 5/5 5/5 5/5 3/5 on a par · 2 worse
depth-box 5/5 5/5 4/5 5/5 on a par
depth-cylinder-attentive 5/5 5/5 4/5 5/5 on a par
depth-cylinder-multi-scale 4/4 3/4 4/4 4/4 on a par
depth-cylinder-projection 5/5 5/5 4/5 5/5 on a par
depth-embed 2/5 2/5 2/5 1/2 on a par · 1 worse
depth-knn 5/5 5/5 5/5 4/5 on a par · 1 worse
depth-pool 3/5 3/5 3/5 2/3 on a par · 1 worse
knn 5/5 4/5 4/5 3/5 on a par · 2 worse
multi-scale-linear 5/5 5/5 4/5 4/5 on a par · 1 worse
multi-scale 4/4 4/4 4/4 2/4 on a par · 1 diverging · 1 worse
pyg-edge-conv 5/5 5/5 5/5 4/5 on a par · 1 worse
pyg-gatv2-conv 5/5 5/5 4/5 4/5 on a par · 1 worse
pyg-point-transformer-conv 5/5 5/5 5/5 4/5 on a par · 1 worse
pyg-pointnet-conv 5/5 5/5 4/5 2/5 on a par · 3 worse
pyg-transformer-conv 5/5 5/5 4/5 4/5 on a par · 1 diverging
roi-aware 5/5 5/5 4/5 4/5 on a par · 1 diverging
roi-grid 5/5 5/5 4/5 5/5 on a par
voxel-set-abstraction 3/5 3/5 3/5 3/3 on a par
CylinderGroup (defaults) 4/4 4/4 4/4 2/4 on a par · 2 worse
CylinderInteractionGroup (defaults) 4/4 4/4 4/4 2/4 on a par · 2 worse
DepthCylinderGroup (defaults) 4/4 4/4 4/4 2/4 on a par · 2 worse

Who is left out, and why:

  • ball: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp
  • box: trains like shipped: worse under FineGrasp
  • cylinder-attentive: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp
  • cylinder-edgeconv: trains like shipped: diverging under FineGrasp
  • cylinder-invres: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline; diverging under GSNet
  • cylinder-vector-attention: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under GSNet
  • depth-ball: trains like shipped: worse under Scale-Balanced-Grasp, FineGrasp
  • depth-box: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • depth-cylinder-attentive: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • depth-cylinder-multi-scale: runs: own in Scale-Balanced-Grasp · learns: not EconomicGrasp (does not learn: width rose 9 %)
  • depth-cylinder-projection: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • depth-embed: runs: not GSNet, EconomicGrasp, FineGrasp (incompatible by design) · trains like shipped: worse under GraspNet-baseline
  • depth-knn: trains like shipped: worse under FineGrasp
  • depth-pool: runs: not GraspNet-baseline, Scale-Balanced-Grasp (incompatible by design) · trains like shipped: worse under FineGrasp
  • knn: learns: not GraspNet-baseline (does not learn: angle flat) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, FineGrasp
  • multi-scale-linear: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp
  • multi-scale: runs: own in FineGrasp · trains like shipped: worse under GraspNet-baseline; diverging under Scale-Balanced-Grasp
  • pyg-edge-conv: trains like shipped: worse under FineGrasp
  • pyg-gatv2-conv: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp
  • pyg-point-transformer-conv: trains like shipped: worse under FineGrasp
  • pyg-pointnet-conv: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, EconomicGrasp, FineGrasp
  • pyg-transformer-conv: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under GSNet
  • roi-aware: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under GSNet
  • roi-grid: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • voxel-set-abstraction: runs: not GraspNet-baseline, Scale-Balanced-Grasp (not supported yet)
  • CylinderGroup (defaults): runs: own in GSNet · trains like shipped: worse under GraspNet-baseline, FineGrasp
  • CylinderInteractionGroup (defaults): runs: own in EconomicGrasp · trains like shipped: worse under GraspNet-baseline, FineGrasp
  • DepthCylinderGroup (defaults): runs: own in GraspNet-baseline · trains like shipped: worse under Scale-Balanced-Grasp, FineGrasp

No row for DepthEmbedGroup, DepthPoolGroup, MultiRadiusGroup, MultiScaleGroup named by its type: it needs an option only a preset supplies, so the bare request is refused before any contract is compared. Its presets have rows.

Head bundles

10 options 8 run everywhere 2 under some 3 of 10 train like shipped

Head bundles: 10 options, one row each

Compare in the explorer: the same options, cell by cell.

Option Runs Learns Predicts Trains like shipped
decoupled-grid-bins 5/5 3/5 5/5 4/5 on a par · 1 diverging
decoupled-grid-class 5/5 2/5 5/5 3/5 on a par · 1 diverging · 1 worse
decoupled-grid-regression 5/5 2/5 5/5 2/5 on a par · 3 diverging
decoupled-grid 5/5 3/5 5/5 3/5 on a par · 1 diverging · 1 worse
economic-focal 3/5 2/5 3/5 3/3 on a par
economic 1/3 1/3 1/3 GSNet: on a par
operation-objectness 4/4 4/4 4/4 1/4 on a par · 1 diverging · 2 worse
operation 4/4 4/4 3/4 2/4 on a par · 2 worse
swad-focal 5/5 5/5 5/5 5/5 on a par
swad 4/4 4/4 3/4 3/4 on a par · 1 diverging

Who is left out, and why:

  • decoupled-grid-bins: learns: not GSNet, FineGrasp (does not learn: angle flat) · trains like shipped: diverging under EconomicGrasp
  • decoupled-grid-class: learns: not GSNet, EconomicGrasp, FineGrasp (does not learn: angle flat) · trains like shipped: worse under GSNet; diverging under EconomicGrasp
  • decoupled-grid-regression: learns: not GSNet, EconomicGrasp, FineGrasp (does not learn: angle flat) · trains like shipped: diverging under GSNet, EconomicGrasp, FineGrasp
  • decoupled-grid: learns: not GSNet, EconomicGrasp (does not learn: angle flat) · trains like shipped: worse under GSNet; diverging under EconomicGrasp
  • economic-focal: runs: not GraspNet-baseline, Scale-Balanced-Grasp (incompatible by design) · learns: not FineGrasp (does not learn: angle rising)
  • economic: runs: own in EconomicGrasp, FineGrasp; not GraspNet-baseline, Scale-Balanced-Grasp (incompatible by design)
  • operation-objectness: runs: own in Scale-Balanced-Grasp · trains like shipped: worse under GSNet, EconomicGrasp; diverging under FineGrasp
  • operation: runs: own in GraspNet-baseline · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp
  • swad: runs: own in GSNet · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under EconomicGrasp

Losses

20 options 17 run everywhere 3 under some 10 of 19 train like shipped

Losses: 20 options, one row each

A loss cell replaces one term per kind and method: the last term of that kind in the pipeline (proposal, view, head), the one nearest the grasp. Those terms:

Method class term unit term value term
GraspNet-baseline head.angle_loss — head.tolerance_loss
GSNet proposal.objectness_loss head.score_loss head.width_loss
Scale-Balanced-Grasp head.angle_loss — head.tolerance_loss
EconomicGrasp head.score_loss view.loss head.width_loss
FineGrasp head.score_loss view.loss head.width_loss

Compare in the explorer: the same options, cell by cell.

Option Runs Learns Predicts Trains like shipped
focal 5/5 5/5 3/5 5/5 on a par
quality-focal 3/4 3/4 2/4 3/3 on a par
balanced-l1 5/5 5/5 3/5 3/5 on a par · 2 worse
softmax-focal 5/5 5/5 5/5 5/5 on a par
varifocal 3/3 3/3 2/3 2/3 on a par · 1 diverging
huber 5/5 5/5 4/5 5/5 on a par
label-smoothing 5/5 5/5 4/5 5/5 on a par
bce 3/3 3/3 2/3 2/3 on a par · 1 diverging
ghm-r 5/5 4/5 4/5 2/5 on a par · 1 diverging · 2 worse
poly1 5/5 5/5 3/5 4/5 on a par · 1 worse
ghm-c 5/5 4/5 4/5 0/5 on a par · 1 diverging · 4 worse
charbonnier 5/5 5/5 4/5 5/5 on a par
ohem 5/5 5/5 4/5 4/5 on a par · 1 worse
ohem-smooth-l1 5/5 5/5 3/5 4/5 on a par · 1 worse
cauchy 5/5 5/5 3/5 5/5 on a par
class-balanced 5/5 3/5 4/5 5/5 on a par
dice 4/5 3/5 3/5 4/4 on a par
lovasz-softmax 4/5 3/5 4/5 3/4 on a par · 1 worse
uncertainty-weighting 5/5 5/5 4/5 5/5 on a par
distribution-focal 5/5 3/5 5/5 not compared

Who is left out, and why:

  • focal: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • quality-focal: runs: not GraspNet-baseline (incompatible by design) · predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • balanced-l1: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp, FineGrasp
  • varifocal: predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under GSNet
  • huber: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • label-smoothing: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • bce: predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under FineGrasp
  • ghm-r: learns: not FineGrasp (does not learn: view rising) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp; diverging under FineGrasp
  • poly1: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline
  • ghm-c: learns: not Scale-Balanced-Grasp (does not learn: angle rising) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp; diverging under FineGrasp
  • charbonnier: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • ohem: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet
  • ohem-smooth-l1: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp
  • cauchy: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • class-balanced: learns: not GraspNet-baseline, Scale-Balanced-Grasp (does not learn: angle flat) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • dice: runs: not Scale-Balanced-Grasp (incompatible by design) · learns: not GraspNet-baseline (does not learn: angle flat) · predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • lovasz-softmax: runs: not Scale-Balanced-Grasp (incompatible by design) · learns: not GraspNet-baseline (does not learn: angle flat) · trains like shipped: worse under GSNet
  • uncertainty-weighting: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • distribution-focal: fits only the width term of the head bundle decoupled-grid-bins; measured with it · learns: not GSNet, FineGrasp (does not learn: angle flat)

Post-processing

15 options 14 run everywhere 1 without a record

Post-processing: 15 options, one row each

Compare in the explorer: the same options, cell by cell.

Option Runs Learns Predicts Trains like shipped
approach-cone 6/6 5/6 4/6 —
center-gripper 6/6 5/6 5/6 —
collision-strict 6/6 5/6 4/6 —
collision 6/6 6/6 4/6 —
furthest-sampling 6/6 6/6 5/6 —
gripper-limits 6/6 6/6 5/6 —
mesh-collision 6/6 6/6 4/6 —
nms 6/6 6/6 4/6 —
score-threshold 6/6 5/6 4/6 —
table-plane 6/6 5/6 4/6 —
to-base-frame 6/6 6/6 4/6 —
top-k 6/6 6/6 4/6 —
width-from-cloud 6/6 6/6 5/6 —
workspace 6/6 5/6 4/6 —
target-mask no record no record no record —

Who is left out, and why:

  • approach-cone: learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • center-gripper: learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • collision-strict: learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • collision: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • furthest-sampling: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • gripper-limits: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • mesh-collision: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • nms: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • score-threshold: learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • table-plane: learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • to-base-frame: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • top-k: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • width-from-cloud: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • workspace: learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • target-mask: reads observation.masks, which no shipped dataset fills

Training recipe

37 options 22 run everywhere 14 under some 1 under none 6 of 23 train like shipped

Training recipe: 37 options, one row each

Compare in the explorer: the same options, cell by cell.

Option Runs Learns Predicts Trains like shipped
optimizer=adamw-no-decay-1d 6/6 6/6 5/6 4/6 on a par · 2 worse
optimizer=adamw 6/6 6/6 5/6 3/6 on a par · 3 worse
optimizer=ademamix 6/6 6/6 5/6 4/6 on a par · 2 worse
optimizer=encoder-lr-0.1 5/5 5/5 5/5 1/5 on a par · 1 diverging · 3 worse
optimizer=layer-decay 5/5 5/5 4/5 2/5 on a par · 1 diverging · 2 worse
optimizer=prodigy 6/6 2/6 6/6 4/6 on a par · 2 worse
optimizer=schedulefree-adamw 6/6 2/6 6/6 not compared
optimizer=schedulefree-radam 6/6 2/6 6/6 not compared
optimizer=schedulefree-sgd 3/6 2/6 2/6 not compared
optimizer=soap 5/6 4/6 5/6 4/5 on a par · 1 worse
optimizer=stableadamw 6/6 6/6 5/6 3/6 on a par · 3 worse
optimizer=timm-adan 6/6 6/6 5/6 2/6 on a par · 1 diverging · 3 worse
optimizer=timm-adopt 6/6 5/6 5/6 3/6 on a par · 3 worse
optimizer=timm-lamb 6/6 3/6 6/6 5/6 on a par · 1 diverging
optimizer=timm-lion 6/6 5/6 5/6 0/6 on a par · 1 diverging · 5 worse
optimizer=timm-madgrad 6/6 6/6 5/6 3/6 on a par · 1 diverging · 2 worse
optimizer=timm-mars 6/6 5/6 5/6 4/6 on a par · 2 worse
schedule=constant 5/6 5/6 5/6 —
schedule=cosine-restarts 5/6 5/6 5/6 —
schedule=cosine 5/6 5/6 5/6 —
schedule=multistep-warmup 5/6 4/6 4/6 —
schedule=onecycle 6/6 6/6 5/6 —
schedule=poly 5/6 5/6 5/6 —
schedule=warmup-cosine 5/6 4/6 5/6 —
hooks=accumulate-2 5/5 5/5 4/5 2/5 on a par · 3 worse
hooks=agc 6/6 5/6 5/6 6/6 on a par
hooks=amp 0/6 0/6 0/6 not compared
hooks=ema 6/6 5/6 4/6 6/6 on a par
hooks=freeze-encoder 4/5 2/5 4/5 4/4 on a par
hooks=grad-clip 6/6 6/6 4/6 5/6 on a par · 1 worse
hooks=swa 4/6 4/6 3/6 4/4 on a par
augment=dropout 5/6 4/6 4/6 5/5 on a par
augment=jitter 5/6 5/6 4/6 5/5 on a par
augment=shift 5/6 5/6 3/6 3/5 on a par · 2 worse
recipe=adamw-onecycle 6/6 6/6 5/6 not compared
recipe=finetune-encoder 4/5 2/5 3/5 not compared
recipe=ptv3-adamw-onecycle 5/5 5/5 4/5 not compared

Who is left out, and why:

  • optimizer=adamw-no-decay-1d: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp, HGGD
  • optimizer=adamw: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under Scale-Balanced-Grasp, FineGrasp, HGGD
  • optimizer=ademamix: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp, HGGD
  • optimizer=encoder-lr-0.1: trains like shipped: worse under GraspNet-baseline, GSNet, EconomicGrasp; diverging under FineGrasp
  • optimizer=layer-decay: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp; diverging under FineGrasp
  • optimizer=prodigy: learns: not GraspNet-baseline (does not learn: angle flat); not Scale-Balanced-Grasp (does not learn: objectness flat; view flat); not EconomicGrasp, FineGrasp (does not learn: objectness flat) · trains like shipped: worse under Scale-Balanced-Grasp, FineGrasp
  • optimizer=schedulefree-adamw: learns: not GraspNet-baseline (does not learn: score rose 19 %; angle flat); not Scale-Balanced-Grasp (does not learn: angle flat); not EconomicGrasp (does not learn: objectness flat); HGGD undecided (the method as shipped does not reach it in a short run either) · trains like shipped: not compared under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp, HGGD: it trained on a schedule of its own, not on that of the method as shipped
  • optimizer=schedulefree-radam: learns: not GraspNet-baseline (does not learn: score flat; angle flat); not Scale-Balanced-Grasp (does not learn: angle flat); not EconomicGrasp (does not learn: objectness flat); HGGD undecided (the method as shipped does not reach it in a short run either) · trains like shipped: not compared under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp, HGGD: it trained on a schedule of its own, not on that of the method as shipped
  • optimizer=schedulefree-sgd: runs: not GSNet, EconomicGrasp, FineGrasp (unstable twice in the short probe) · learns: not Scale-Balanced-Grasp (does not learn: angle flat) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: not compared under GraspNet-baseline, Scale-Balanced-Grasp, HGGD: it trained on a schedule of its own, not on that of the method as shipped; not compared under GSNet, EconomicGrasp, FineGrasp: the swap did not run at level 1
  • optimizer=soap: runs: not HGGD (not verified: one CUDA illegal memory access on a shared GPU) · learns: not GraspNet-baseline (does not learn: score flat; angle flat) · trains like shipped: worse under FineGrasp
  • optimizer=stableadamw: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, FineGrasp, HGGD
  • optimizer=timm-adan: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GSNet, EconomicGrasp, HGGD; diverging under FineGrasp
  • optimizer=timm-adopt: learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp, FineGrasp, HGGD
  • optimizer=timm-lamb: learns: not GraspNet-baseline (does not learn: score rising; angle flat); not Scale-Balanced-Grasp (does not learn: angle flat); not FineGrasp (does not learn: objectness flat) · trains like shipped: diverging under FineGrasp
  • optimizer=timm-lion: learns: not GraspNet-baseline (does not learn: score flat; angle flat) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, EconomicGrasp, FineGrasp, HGGD; diverging under Scale-Balanced-Grasp
  • optimizer=timm-madgrad: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: diverging under GraspNet-baseline; worse under Scale-Balanced-Grasp, FineGrasp
  • optimizer=timm-mars: learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp, HGGD
  • schedule=constant: runs: not FineGrasp (unstable twice in the short probe)
  • schedule=cosine-restarts: runs: not FineGrasp (unstable twice in the short probe)
  • schedule=cosine: runs: not FineGrasp (unstable twice in the short probe)
  • schedule=multistep-warmup: runs: not FineGrasp (unstable twice in the short probe) · learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • schedule=onecycle: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • schedule=poly: runs: not FineGrasp (unstable twice in the short probe)
  • schedule=warmup-cosine: runs: not FineGrasp (unstable twice in the short probe) · learns: HGGD undecided (the method as shipped does not reach it in a short run either)
  • hooks=accumulate-2: runs: own in HGGD · predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under GraspNet-baseline, GSNet, EconomicGrasp
  • hooks=agc: learns: not FineGrasp (does not learn: angle rising) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • hooks=amp: runs: not GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp, HGGD (not supported yet)
  • hooks=ema: learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • hooks=freeze-encoder: runs: not FineGrasp (unstable twice in the short probe) · learns: not GSNet (does not learn: objectness flat; graspness rose 8 %); not EconomicGrasp (does not learn: objectness flat; graspness flat) · trains like shipped: not compared under FineGrasp: the swap did not run at level 1
  • hooks=grad-clip: predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under FineGrasp
  • hooks=swa: runs: not GraspNet-baseline, Scale-Balanced-Grasp (not verified: the probe is shorter than the epoch at which averaging starts) · predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • augment=dropout: runs: not FineGrasp (fails: the transform order of the composed config is wrong (DepthNormals must run first)) · learns: HGGD undecided (the method as shipped does not reach it in a short run either) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either)
  • augment=jitter: runs: not HGGD (not supported yet) · predicts: FineGrasp undecided (the method as shipped does not reach it in a short run either)
  • augment=shift: runs: not HGGD (not supported yet) · predicts: Scale-Balanced-Grasp, FineGrasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: worse under EconomicGrasp, FineGrasp
  • recipe=adamw-onecycle: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: not compared under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp, HGGD: it trained on a schedule of its own, not on that of the method as shipped
  • recipe=finetune-encoder: runs: not FineGrasp (unstable twice in the short probe) · learns: not GSNet (does not learn: objectness flat; graspness rose 152 %); not EconomicGrasp (does not learn: objectness flat; graspness rose 165 %) · predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: not compared under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp: it trained on a schedule of its own, not on that of the method as shipped; not compared under FineGrasp: the swap did not run at level 1
  • recipe=ptv3-adamw-onecycle: predicts: Scale-Balanced-Grasp undecided (the method as shipped does not reach it in a short run either) · trains like shipped: not compared under GraspNet-baseline, GSNet, Scale-Balanced-Grasp, EconomicGrasp, FineGrasp: it trained on a schedule of its own, not on that of the method as shipped

A preset that names the scene encoder (optimizer=encoder-lr-0.1, optimizer=layer-decay, hooks=freeze-encoder, recipe=finetune-encoder, recipe=ptv3-adamw-onecycle) applies to the seed-based methods only: HGGD has no scene encoder.

The methods as shipped

Each method's own config, probed the same way: the base every swap is held against. A level the method as shipped does not reach in a short run is undecided for its swaps.

Method Verdict Grasps Learns Predicts
GraspNet-baseline runs 690 · 551 ✓ ✓
GSNet runs 206 · 229 ✓ ✓
Scale-Balanced-Grasp runs-empty 0 · 0 ✓ ✗ no grasp on any frame
EconomicGrasp runs 1024 · 1024 ✓ ✓
FineGrasp runs 1024 · 1024 ✓ ✗ every score of frame 1 is -1.687; every score of frame 2 is -1.687
HGGD runs 336 · 329 ✗ anchor_offset flat ✓

Sets of stages

A set replaces the encoder and every stage at once. The rules accept 9104 sets of stage options (per kind of encoder output); 95 were probed (86 runs, 9 runs-empty), covering 118 of the 118 pairs of a stage variant and a kind of encoder. A set that runs says nothing about another set with the same parts.

The probed sets, by the method whose data recipe serves their labels
Data recipe of Sets probed runs runs-empty
GraspNet-baseline 16 16 —
GSNet 35 30 5
Scale-Balanced-Grasp 21 17 4
EconomicGrasp 23 23 —

Every probed set, with its request and its verdict, is in the records (group stage-set). Swap rules counts the same rules over registered classes: 3014 sets.

Several slots at once

These cells replace an encoder and options of other slots in one request: a hand-picked sample of 18, at least two per method (11 runs, 7 runs-empty). A combination not listed was not probed; its verdict does not follow from its parts.

The 18 requests, method by method
Method Requests probed runs runs-empty
GraspNet-baseline 2 2 —
GSNet 6 3 3
Scale-Balanced-Grasp 2 — 2
EconomicGrasp 3 2 1
FineGrasp 3 2 1
HGGD 2 2 —
Method Request Verdict
GraspNet-baseline --set backbone=pointnet2-msg --set grouper=depth-ball --set post=nms runs
GraspNet-baseline --set backbone=assanet --set view=two-vector-baseline --set grouper=depth-knn --set optimizer=timm-lamb runs
GSNet --set backbone=point-transformer-v2 --set grouper=cylinder-vector-attention runs
GSNet --set backbone=pointvector --set proposal=graspness-nms --set grouper=roi-grid runs-empty
GSNet --set backbone=minkunet34c --set view=healpix-192 --set grouper=multi-scale-linear runs
GSNet --set backbone=volt --set proposal=graspness-fpsample --set view=icosphere-162 --set grouper=knn runs-empty
GSNet --set backbone=ptv3 --set proposal=dense-fps --set view=direction-vmf --set grouper=ball runs
GSNet --set backbone=uni3d --set view=rotation-6d --set grouper=multi-scale --set optimizer=schedulefree-adamw runs-empty
Scale-Balanced-Grasp --set backbone=pointmetabase --set grouper=depth-cylinder-attentive runs-empty
Scale-Balanced-Grasp --set backbone=dgcnn --set grouper=depth-box --set post=gripper-limits runs-empty
EconomicGrasp --set backbone=dela --set grouper=pyg-edge-conv --set optimizer=adamw runs-empty
EconomicGrasp --set backbone=spunet-gaussiancross --set proposal=graspness-voxel --set grouper=box --set schedule=warmup-cosine runs
EconomicGrasp --set backbone=pointmlp --set view=lattice-topk --set grouper=multi-scale runs
FineGrasp --set backbone=randlanet --set view=lattice-soft-argmax --set hooks=ema runs-empty
FineGrasp --set backbone=point-transformer-v1 --set grouper=cylinder-invres runs
FineGrasp --set backbone=point-transformer-v1 --set view=lattice-best-view --set grouper=depth-ball runs
HGGD --set image_encoder=convnextv2-atto --set post=nms --set optimizer=adamw --set schedule=cosine runs
HGGD --set image_encoder=tv-mobilenetv3-large --set post=top-k --set hooks=ema runs

The records

measured_combinations.json holds every cell: its two commands, its verdict, why it was refused or failed, and what the probe saw; losses_per_iteration.json every loss term at every iteration. The explorer answers any selection from them, with the command.

Add a component to the grid

A new encoder or stage enters the grid by being registered. Give it a preset if its options cannot be guessed (Compose and probe), probe its cells and regenerate. tests/unit/test_combine_grid.py fails until every cell of a registered component has a record.

python -m tools.probe_grid run --only 'encoder/*/MyEncoder' 'stage/*/*/my-preset'
python -m tools.probe_grid export
python -m tools.make_combination_docs