6-DOF GRASPING · LEARNED CANDIDATE RANKING

GraRe: Grasp Candidate Re-Ranking for Frozen 6-DoF Grasp Detectors

Jibao Yuan¹Yuhui Zhao¹Yinzhen Lv²Chao Xu¹Shun Li¹Chenxi Deng¹Shaofei Chen^1,*

¹ College of Intelligence Science and Technology, National University of Defense Technology² Department of Engineering and Information Technology, The University of Melbourne

* Corresponding author

PaperarXiv CodeGitHub DatasetHugging Face ModelsHugging Face CitationBibTeX

Benchmark resultsRobot experiments

Learn to rank existing grasp candidates while keeping the detector and grasp poses frozen.

A NEW ORDER, THE SAME GRASPS

Better ranking. Unchanged candidates.

GraRe combines candidate attributes, local geometry, and object context to bring successful grasps forward.

+13.56max. Average AP gain

5frozen detectors

2RGB-D cameras

UR3 · Robotiq 2F-85 · RealSense D435

01 / MOTIVATION

Why re-rank grasp candidates?

Detector confidence does not always reflect grasp quality. GraRe learns which existing candidates to try first, without changing their poses.

Detector confidenceOriginal ranking

GraReRe-ranked candidates

Abstract

Existing 6-DoF grasp detectors typically rank grasp candidates by detector confidence. However, our analysis on GraspNet-1Billion shows that detector confidence is often poorly aligned with grasp quality, leaving successful grasp candidates at low ranks. Motivated by this observation, we study whether learned re-ranking can improve candidate ordering while keeping detector parameters and grasp candidates unchanged. We propose GraRe, which estimates grasp quality from candidate attributes, shell-stratified local geometry, and object context. Candidate attributes condition the local geometric and object-context representations, and a Transformer fuses all three feature types. The predicted quality is combined with detector confidence to produce the final ranking. Experiments on GraspNet-1Billion with five frozen detectors show consistent improvements, with gains of up to 13.56 points in Average AP. Real-robot experiments further demonstrate robust grasping in cluttered scenes. These results show that improving candidate ranking provides a practical way to enhance frozen 6-DoF grasp detectors.

02 / METHOD

Three views of every grasp.

Candidate attributes condition local geometry and object context. A Transformer brings these complementary representations together to estimate grasp quality.

GraRe architecture: frozen detector, candidate features, shell-wise local geometry, object context, Transformer fusion, and score fusion — Three complementary views predict grasp quality and improve the ordering of existing candidates.

Candidate attributes

Pose, gripper width, and detector confidence guide the representation of each grasp.

Local geometry

Geometry around the gripper captures nearby surfaces and possible finger contacts.

Object context

The visible object shape provides context beyond the immediate gripper neighborhood.

FINAL RANKING SCORE

s = (1 − λ) · z(b) + λ · z(ĝ)

Combine standardized detector confidence b and predicted quality ĝ, then sort.

03 / BENCHMARK RESULT

Evaluated using the official GraspNet-1Billion evaluation protocol.

RealSense · Average AP

DetectorGraRe

35.86

49.42

+13.56

SBG

47.66

52.75

+5.09

52.03

56.14

+4.11

HGGD

46.49

49.53

+3.04

RNGNet

58.10

59.98

+1.88

RealSense results · Detector / **GraRe**
Detector	Seen	Similar	Novel	Average
GN	47.83 / 64.36	42.79 / 58.44	16.94 / 25.45	35.86 / 49.42
SBG	62.27 / 68.47	56.92 / 62.55	23.79 / 27.24	47.66 / 52.75
EG	69.31 / 74.92	61.49 / 65.16	25.28 / 28.35	52.03 / 56.14
HGGD	63.37 / 67.58	52.28 / 55.37	23.83 / 25.66	46.49 / 49.53
RNGNet	75.32 / 77.67	66.95 / 69.64	32.03 / 32.63	58.10 / 59.98

Kinect · Average AP

DetectorGraRe

30.60

38.80

+8.20

45.26

49.75

+4.49

HGGD

42.13

44.94

+2.81

RNGNet

53.05

54.26

+1.21

Kinect results · Detector / **GraRe**
Detector	Seen	Similar	Novel	Average
GN	41.98 / 54.06	37.57 / 46.38	12.24 / 15.97	30.60 / 38.80
EG	63.76 / 69.72	52.42 / 57.47	19.61 / 22.06	45.26 / 49.75
HGGD	60.43 / 64.95	46.74 / 50.03	19.23 / 19.85	42.13 / 44.94
RNGNet	73.50 / 75.34	59.20 / 60.86	26.44 / 26.58	53.05 / 54.26

04 / ABLATIONS

What contributes to better ranking?

Local geometry and auxiliary supervision consistently support better ranking; the value of object context varies by detector.

Average AP · RealSense / Kinect
Variant	GN	SBG	EG	HGGD	RNGNet
Detector	35.86 / 30.60	47.66 / —	52.03 / 45.26	46.49 / 42.13	58.10 / 53.05
Full GraRe	49.42 / 38.80	52.75 / —	56.14 / 49.75	49.53 / 44.94	59.98 / 54.26
w/o Candidate	47.63 / 37.21	51.74 / —	54.19 / 48.15	47.99 / 43.73	59.98 / 53.79
w/o Local	44.99 / 36.48	50.60 / —	54.45 / 48.43	47.77 / 43.15	58.76 / 53.60
w/o Object	48.71 / 38.04	52.59 / —	55.98 / 49.73	49.92 / 45.05	59.83 / 54.44
w/o Shell-wise FPS	48.27 / 38.12	52.64 / —	55.96 / 49.71	47.90 / 43.33	58.69 / 53.66
w/ Concat Fusion	48.72 / 38.12	52.73 / —	56.04 / 49.92	49.88 / 45.24	59.85 / 54.42
w/o Aux. Losses	48.92 / 38.50	52.27 / —	55.70 / 49.40	49.47 / 44.55	59.77 / 54.10

05 / RANKING ANALYSIS

How does the ordering change?

GraRe brings successful candidates forward while preserving the detector’s candidate set.

Ranking quality: original figure from the GraRe manuscript

Better ordering, with headroom remaining.

Successful candidates move closer to the front. DS@K rewards successful grasps at earlier ranks; FS@K measures how often a successful grasp appears in the first K candidates.

Curves average across detectors for each camera. The oracle uses ground-truth quality to show the remaining ranking headroom.

Outcome changes: original figure from the GraRe manuscript

Both collision and force-closure errors matter.

Collision and force-closure failures account for 47.86% and 49.7% of failure-to-success changes, respectively. Improvements go beyond rejecting collisions.

Regressions remain: force-closure failures account for 63.19% of success-to-failure changes. Percentages describe the changed-outcome groups, not the overall success rate.

Fusion sensitivity: original figure from the GraRe manuscript

Detector confidence retains useful information.

GN benefits most from learned quality; RNGNet benefits from combining it with the detector’s original confidence.

The paper uses λ = 1.0 for GN, 0.8 for SBG / EG / HGGD, and 0.5 for RNGNet across the available cameras.

Qualitative cases: original figure from the GraRe manuscript

Successful corrections—and a ranking regression.

The first example demotes a colliding grasp. In the second, GraRe promotes a collision-free GN candidate from rank 17 to rank 1, improving the required friction coefficient from 0.6 to 0.2.

The RNGNet example reverses a correct detector preference: a candidate requiring μ = 0.6 displaces one requiring μ = 0.2. Relative quality errors can therefore overturn a correct first choice.

06 / REAL ROBOT

From ranking to grasping.

A UR3 arm, a Robotiq 2F-85 gripper, and a RealSense D435 camera. Ten cluttered scene configurations, three detectors, and two ranking policies.

Higher success, more cleared scenes

10 scenes per detector and ranking policy
Detector	Ranking	Successful grasps	Cleared scenes
GN	Detector	73.4%58 / 79	10%1 / 10
GN	GraRe	88.9%72 / 81	100%10 / 10
SBG	Detector	70.4%57 / 81	30%3 / 10
SBG	GraRe	87.2%68 / 78	90%9 / 10
EG	Detector	68.2%60 / 88	70%7 / 10
EG	GraRe	91.0%71 / 78	100%10 / 10

A successful grasp removes at least one object. A cleared scene has no objects remaining.

Across 60 trials, GraRe raises grasp success for all three detectors and clears more cluttered scenes.

Extra processing time · RTX 2060
GN	SBG	EG
0.84 s	0.52 s	0.50 s

Most of the added time is spent preparing local geometry.

GraRe improves ordering; performance remains bounded by the candidates the frozen detector can generate.

From selected candidate to execution

Promoted candidates can remove objects from clutter, while difficult contacts can still fail.

GN Scene 4 carton and SBG Scene 6 banana: candidate selection, execution, and post-grasp keyframes — GN promotes a carton grasp from rank 60 to 9; SBG promotes a banana grasp from rank 12 to 1. Both executions succeed.

EG Scene 8 successful carton grasp and GN Scene 8 failed screwdriver grasp with selected candidates and execution frames — EG promotes a red-carton grasp from rank 20 to 2 and succeeds. GN promotes a screwdriver grasp from rank 56 to 1, but it does not lift the object.

07 / RESOURCES

Paper, code, and released assets

ARXIV

CITATION

BibTeX

@article{yuan2026grare,
  title   = {GraRe: Grasp Candidate Re-Ranking for Frozen 6-DoF Grasp Detectors},
  author  = {Yuan, Jibao and Zhao, Yuhui and Lv, Yinzhen and Xu, Chao and Li, Shun and Deng, Chenxi and Chen, Shaofei},
  journal = {arXiv preprint arXiv:2608.00946},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.00946}
}