Object Pose Estimation · 2025 · ICRA · CCF B

ZeroBP: Learning Position-Aware Correspondence for Zero-Shot 6D Pose Estimation in Bin-Picking

ZeroBP combines local point cloud features with object-centric direction encoding in cross-attention. Alternating superpoint-to-point pose / position refinement reduces mismatches caused by repetitive local geometry in industrial parts.

Paper

Research area

Object Pose Estimation

Task level

Instance-level

Object setting

Unseen objects

Input modalities

RGB-D bin sceneCAD point cloud

Requirements

unseen-object CAD modelcamera intrinsicspicking-box promptinstance segmentation

Full paper metadata Authors, affiliations, research area, datasets, metrics, method tags, and reading status Expand

Authors: Jianqiu Chen, Zikun Zhou, Xin Li, Ye Zheng, Tianpeng Bao, Zhenyu He
Affiliations: Harbin Institute of Technology, Shenzhen / Pengcheng Laboratory / JD.com / SenseTime Research
Conference / journal: Conference · ICRA · CCF B
Research area: Object Pose Estimation
Code availability: Unavailable / not yet recorded
Reading status: Overview:Available Deep Dive:Available
Task setting: Instance-level · Unseen objects
Datasets: Google Scanned Objects synthetic set ROBI
Evaluation metrics: ADD ADD-S ADD(S) / ADD(-S)
Method tags: Novel-object Pose Estimation 2D–3D Correspondences Dense Correspondences & Coordinate Fields Multi-object Pose Estimation Synthetic Data & Foundation Models Absolute Pose Estimation
Requirements: unseen-object CAD modelcamera intrinsicspicking-box promptinstance segmentation

Reading Notes

Overview / Deep Dive

Available src/paper-notes/zerobp-icra2025/coarse.md

Paper overview notes

Quickly assess the paper's scope, problem setting, core method, experimental findings, and whether to read it in depth.

Overview Notes: ZeroBP: Learning Position-Aware Correspondence for Zero-Shot 6D Pose Estimation in Bin-Picking

1. Basic information

Authors: Jianqiu Chen, Zikun Zhou, Xin Li, Ye Zheng, Tianpeng Bao, Zhenyu He
Venue: ICRA 2025
Paper: arXiv:2502.01004
PDF: Authors’ public full text
Reading mode: Overview

2. Positioning in one sentence

ZeroBP encodes directions between points and object centers into cross-attention, going beyond local appearance. At superpoint and point levels, it alternates correspondence, pose, and global-position updates to reduce mismatches caused by repeated industrial-part geometry.

2.1 Guide to the paper’s figures and tables

Figure / table	Content	Reading focus
Figure 1	ZeroBP versus existing zero-shot methods	Position-aware correspondence is the key distinction
Figure 2	PAC and coarse-to-fine pipeline	Coarse superpoint matching precedes fine point matching with alternating updates
Table I / Figure 3	ROBI main results and qualitative alignment	The general model trails object-specific training but clearly exceeds zero-shot baselines

ZeroBP Figure 1: zero-shot bin-picking pose methods compared — Original Figure 1: repeated local geometry confuses correspondences based only on local descriptions; ZeroBP adds each point's global direction within the object. Source: authors' public PDF. Image copyright belongs to the authors and publisher; reproduced here for paper analysis.

3. Task setting

Inputs: bin RGB-D, object CAD cloud, camera intrinsics, and picking-box region.
Output: 6D poses of object instances unseen during training.
Training: 1000 Google Scanned Objects and about 1 million synthetic RGB-D images.
Testing: ROBI, with 7 reflective or weakly textured metal objects, 14 scenes, 1218 images, over 20K instances, and up to about 38 instances per bin.

4. Method pipeline

Picking-box-aware detection / segmentation extracts candidate instances.
KPConv-FPN encodes scene and CAD point clouds separately.
Position-aware superpoint correspondence yields coarse matches.
Current pose estimates object center and point-to-center directions.
Multiplicative positional encoding injects directions into cross-attention.
Point-level matching alternates with pose and global-position refinement.
Weighted SVD solves the rigid transform.

5. Results at a glance

Mean ADD(-S) recall on ROBI:

Method	Mean AR
ZeroPose	29.4
SAM-6D	36.7
ZeroBP	45.8
ST6D (object-specific)	64.1

ZeroBP improves over the contemporary zero-shot baseline by 9.1 percentage points, while remaining well below ROBI-object-trained ST6D.

6. Coarse reading conclusion

Repeated holes, edges, and cylinders make local industrial geometry ambiguous; reliable correspondence needs its location within the whole object. ZeroBP suits frequently changing bin-picking objects with available CAD, provided depth, instance masks, and coarse-center estimates remain stable.

Available src/paper-notes/zerobp-icra2025/deep.md

In-depth paper notes

Support reproduction, comparison, related-work writing, and identifying opportunities for further research.

Deep Reading Notes: ZeroBP: Learning Position-Aware Correspondence for Zero-Shot 6D Pose Estimation in Bin-Picking

1. Basic information

Paper: arXiv:2502.01004
PDF: Authors’ public full text
Task: CAD-conditioned zero-shot RGB-D bin-picking 6D pose

2. Core contribution

ZeroBP identifies missing object-level position context as a bottleneck in zero-shot correspondence. Several flange holes can have nearly identical local geometry; without their directions relative to the object center, matching can be locally plausible but globally wrong.

ZeroBP Figure 2: Position-Aware Correspondence and coarse-to-fine pipeline — Original Figure 2: scene and CAD matching moves from superpoints to points; directional position encoding and alternating pose / position refinement guide correspondence updates. Source: authors' public PDF. Image copyright belongs to the authors and publisher; reproduced here for paper analysis.

3. Position-Aware Correspondence

For model point $x_i$ and object center $c$ , construct a direction:

d_i=\frac{x_i-c}{\lVert x_i-c\rVert_2}.

The scene center is unknown and inferred from current pose / global position. Direction encoding multiplicatively modulates attention features rather than simply concatenating coordinates, producing different matches for similar local descriptions at different object positions.

This suits zero-shot transfer better than raw absolute 3D coordinates:

Directions decouple from absolute object scale and camera distance;
Center-relative directions distinguish repeated structures by global location;
Improved pose makes scene directions more accurate, providing iterative positive feedback.

Center error can corrupt directions, creating a circular dependence. Alternating refinement addresses this instead of fixing positional encoding once.

4. Coarse-to-fine matching and alternating updates

4.1 Superpoint level

KPConv-FPN downsamples clouds into superpoints. Fewer coarse candidates with larger receptive fields first build high-recall matches, avoiding costly all-point attention.

4.2 Point level

Recover original points within matched superpoint groups and refine correspondence. Position-aware cross-attention combines local features with center-relative directions.

The system alternates:

Update correspondence using current positional encoding;
Solve $R,t$ from correspondences using weighted SVD;
Update scene object center and directions with the new pose;
Repeat correspondence estimation.

This is iterative optimization combining learned matching with an interpretable geometric solver.

5. Data and evaluation

Synthetic training: 1000 GSO objects, about 1M RGB-D images.
ROBI testing: 7 metal parts, 14 scenes, 1218 images, over 20K instances.
Density: up to about 38 instances per bin.
Main metric: ADD(-S) recall under an object-diameter threshold, aggregated across objects.

ZeroBP Table I and Figure 3: quantitative and qualitative ROBI results — Original Table I / Figure 3: ZeroBP leads the zero-shot ROBI comparison while object-specific ST6D remains stronger; qualitative results show alignment for dense, reflective parts. Source: authors' public PDF. Image copyright belongs to the authors and publisher; reproduced here for paper analysis.

Method	Mean AR	Training relationship
ZeroPose	29.4	zero-shot
SAM-6D	36.7	zero-shot
ZeroBP	45.8	zero-shot
ST6D	64.1	ROBI object-specific

The 9.1 gain is the percentage-point difference between 45.8 and 36.7, not a relative percentage. ST6D’s 64.1 shows substantial remaining value in target-specific training.

6. Ablation studies

Under the GT-mask protocol:

Position encoding / refinement	AR
No positional encoding	36.4
Direct 3D coordinates	47.3
3D coordinates + alternate refinement	49.3
directional vector + alternate refinement	55.7

Directional encoding adds 19.3 over no positional encoding, supporting the value of global context. Alternating refinement also helps ordinary coordinates, indicating a separate benefit from iteration itself.

PAC at both superpoint and point levels reaches 55.7 AR with about 10.4M parameters and 0.076 s correspondence-module inference. Using only one level lowers accuracy; both global disambiguation and local refinement benefit from position awareness.

7. Assumptions and limitations

Metric CAD and reliable depth are required; missing points from transparent / specular surfaces damage KPConv features.
Picking-box regions / instance masks remain necessary; the pose backend is not an open-scene detector.
Direction encoding depends on object center; poor initialization may converge to an incorrect fixed point.
Directions cannot recover unobservable information for fully symmetric objects; symmetry-aware evaluation or grasp objectives remain necessary.
Synthetic GSO differs from industrial metal parts; reported gains do not eliminate sim-to-real gaps.

8. Deep reading conclusion

Zero-shot correspondence needs global object location as well as transferable local features. Directional encoding, coarse-to-fine matching, and alternating geometric updates form a feedback loop, with the 36.4 to 55.7 ablation gain offering compelling evidence. ZeroBP merits reproduction for ICRA 2025 industrial zero-shot pose research.

Related papers

2026 · CVPR

3PT

3PT unifies detection, segmentation, and 6-DoF pose with two Transformers for 2D / 3D object-centric understanding. Continuous refinement of object representations and rendered-feature alignment support zero-shot cross-object perception on BOP-H3, BOP-Industrial, and BOP-Classic.

2026 · AAAI

6DAttack

Studies backdoor attacks on PnP-based and end-to-end 6-DoF pose pipelines. Structurally 3D triggers and modified pose labels are injected into training data so that models behave normally on clean inputs but produce controlled incorrect poses when triggered.

2026 · ICRA

Active Tactile Exploration

A violation-implicit loss incorporates physical-constraint violations into joint estimation of freely moving unknown convex rigid-body trajectories and cuboid / 20-vertex polytope shapes. Observed / Fisher information at a local MLE estimates EIG to select the next touch. In simulation, the EIG strategy yields stable errors below 4 mm after 5 cuboid actions; real Trifinger results at step 5 substantially outperform random selection. Convex geometry, local optima, multiple initializations, and tactile / proprioceptive errors limit the method.

ZeroBP: Learning Position-Aware Correspondence for Zero-Shot 6D Pose Estimation in Bin-Picking

Overview / Deep Dive

Paper overview notes

Overview Notes: ZeroBP: Learning Position-Aware Correspondence for Zero-Shot 6D Pose Estimation in Bin-Picking

1. Basic information

2. Positioning in one sentence

2.1 Guide to the paper’s figures and tables

3. Task setting

4. Method pipeline

5. Results at a glance

6. Coarse reading conclusion

In-depth paper notes

Deep Reading Notes: ZeroBP: Learning Position-Aware Correspondence for Zero-Shot 6D Pose Estimation in Bin-Picking

1. Basic information

2. Core contribution

3. Position-Aware Correspondence

4. Coarse-to-fine matching and alternating updates

4.1 Superpoint level

4.2 Point level

4.3 Pose / position refinement

5. Data and evaluation

6. Ablation studies

7. Assumptions and limitations

8. Deep reading conclusion

Related papers

3PT

6DAttack

Active Tactile Exploration