Industrial bin picking requires pose estimates that are accurate under clutter, partial occlusion, and the reflective, texture-less nature of manufactured parts. Most recent approaches to 6D pose estimation follow multi-stage pipelines. These achieve high accuracy, but incur memory and runtime footprints that do not cope with the operational constraints of many real-world industrial deployments. Holistic pipelines estimate poses in a single pass, with cost controlled through indiscriminate subsampling or lowered resolution. Continuing this line, we present SCCV-6D, a fully sparse multi-view RGB-D estimator. We reconstruct the 3D space at millimeter precision, prune background via a staged heatmap mechanism, and predict class, translation offset and rotation at highest resolution for an arbitrary number of objects simultaneously. Rotation is formulated as canonical coordinate voting. Points vote for their canonical object coordinates with an associated confidence, enabling closed-form pose recovery with a symmetry-aware weighted Procrustes solve. Along an in-cluster point recovery induced by offset votes, this removes time-consuming post-processing stages. We reuse the task-supervised attention priors from the heatmap formulation for a novel attention-gated distillation objective that reweights teacher-student feature alignment. On the IPD test split we report competitive 0.826 AP at 1.13 s per scene, significantly faster than the leading entries. On MV-YCB our approach reaches 96.1 ADD-S, and remains competitive even on the densely packed bins of ROBI.
Figures & tables
Figure 1: Fast and accurate. SCCV-6D estimates every pose in a single forward pass and a batched closed-form solve. It reaches 0.826 AP on the IPD test split in 1.13 s per scene, 30× faster than the leading entry at 89% of its accuracy. SCCV-6D leads the only competitor in the same latency band by 8.0 AP. Runtimes of competing entries are self-reported on heterogeneous hardware and indicate a trend rather than a matched comparison.
Figure 2: SCCV-6D. n RGB-D views are fused into a sparse TSDF, yielding P . Two sparse, voxel-based heatmap stages focus task-relevant foreground regions I′ , before recovering to point-level resolution Q . A shared PTv3 trunk feeds all three estimation heads. The translation votes uj naturally form clusters. Along class predictions, these gate a FiLM-conditioned canonical-coordinate head, and one single batched symmetry-aware weighted Procrustes solve returns 6D poses [(R^k,t^k)] .
AP
AP
AP mm
AP
Time
Method
AP
MSPD
MSSD
MSSD
25
[s]
FRTPose-WAPRv2 (MV) [ 48 ]
.932
.968
.897
.661
.897
33.9
3PT-Pose-Ind. [ 26 ]
.923
.933
.914
.822
.914
24.52
SEVD-0504 [ 1 ]
.873
.913
.833
.644
.833
13.4
AlignPose [ 38 ]
.798
.852
.743
.582
.743
–
MatchU [ 22 ]
.520
.566
.475
.340
.475
–
Table 1: IPD test split under the BOP protocol [ 40 ] . All competing entries are multi-stage systems relying on upstream detections and, in several cases, on ensembles of detectors and refiners. Competing runtimes are self-reported [ 40 ] on heterogeneous hardware with differing modalities and view counts.
CosyPose
CosyPose
MV-3D-KP
MEC-Net
Ours
+PVNet
+LINE2D
[ 33 ]
[ 54 ]
Modality
RGB
RGB-D
RGB-D
RGB
RGB-D
Tube Fitting
39.4
74.8
94.0
89.4
90.9
Chrome Screw
17.4
73.0
90.8
86.7
83.2
Eye Bolt
21.6
85.1
93.2
93.2
92.7
Gear
50.6
80.2
85.2
91.4
88.3
Table 2: ROBI (Ensenso, four views) under ADD ∗ . Baselines as reported in [ 54 ] .
Method
ADD-S
ADD-(S)
CosyPose [ 27 ]
92.3
87.7
MV6D [ 10 ]
91.2
85.6
SyMFM6D w/o sym. [ 11 ]
95.2
91.5
SyMFM6D [ 11 ]
95.4
91.7
SDT-6D [ 32 ]
90.6
89.1
SCCV-6D (ours)
96.1
92.8
Table 3: MV-YCB-SymMovCam, three views, area under the ADD-S and ADD-(S) curves.
Configuration
AP
AP MSSDmm
Per-voxel rotation, batched ICP
0.712
0.468
w/o color (depth-only heads)
0.769
0.483
w/o point recovery (per-cell)
0.812
0.522
Single five-stage encoder per head
0.814
0.517
unweighted Procrustes solve
0.816
0.542
SCCV-6D (full)
0.826
0.548
Table 4: Component study on the IPD test split. Each row replaces or removes a single component of the full model. Effects are measured independently and are not additive.
AP
AP
AP mm
Time
Model
AP
MSPD
MSSD
MSSD
[s]
Teacher
0.841
0.884
0.798
0.552
1.11
−25% , w/o weighting
0.723
0.781
0.665
0.485
1.02
−25% , occupancy [ 55 ]
0.768
0.829
0.707
0.505
1.03
−25% , heatmap (ours)
0.827
0.866
0.789
0.542
1.06
−35% , occupancy [ 55 ]
0.747
0.821
0.673
0.504
1.03
Table 5: Distillation Results on the IPD validation set. Percentages denote the parameter reduction with respect to the teacher. The teacher is the full model of Table 1 .
Bin picking in real industrial environments remains challenging due to severe clutter, occlusions, and the high cost of traditional 3D sensing setups. We present Pickalo, a modular 6D pose-based bin-picking pipeline built entirely on low-cost hardware. A wrist-mounted RGB-D camera actively explores the scene from multiple viewpoints, while raw stereo streams are processed with BridgeDepth to obtain refined depth maps suitable for accurate collision reasoning. Object instances are segmented with a Mask-RCNN model trained purely on photorealistic synthetic data and localized using the zero-shot SAM-6D pose estimator. A pose buffer module fuses multi-view observations over time, handling object symmetries and significantly reducing pose noise. Offline, we generate and curate large sets of antipodal grasp candidates per object; online, a utility-based ranking and fast collision checking are queried for the grasp planning. Deployed on a UR5e with a parallel-jaw gripper and an Intel RealSense D435i, Pickalo achieves up to 600 mean picks per hour with 96-99% grasp success and robust performance over 30-minute runs on densely filled euroboxes. Ablation studies demonstrate the benefits of enhanced depth estimation and of the pose buffer for long-term stability and throughput in realistic industrial conditions. Videos are available at https://mesh-iit.github.io/project-jl2-camozzi/
MESH Facility (formerly iCub Tech), Istituto Italiano di Tecnologia, Genova 16163, Italy · Institut des Systèmes Intelligents et de Robotique (ISIR), Paris 75005, France · Generative Bionics, Genova 16152, Italy +1
While current 6D pose estimation benchmarks have reached near-saturation on household objects, they often fail to capture the stochastic and optical complexities of industrial environments. We introduce XYZ-IBD, a high-precision benchmark for object detection and 6D pose estimation specifically designed for industrial bin-picking. XYZ-IBD addresses the domain gap by providing 75 multi-view real-world scenes containing approximately 273k annotated instances of metallic, symmetrical, and specular objects. Unlike existing datasets, our benchmark features high-density stochastic stacking and multi-instance ambiguity, reflecting authentic robotic manipulation challenges. We employ a rigorous multi-stage and semi-automatic annotation pipeline, ensuring sub-millimeter annotation accuracy. The annotations are validated through our designed error quantification scheme, securing the reliability of the annotation quality. In addition to real-world evaluation data, we provide a large-scale complementary synthetic training set that is rendered under a realistic bin-picking simulation. Benchmarking state-of-the-art (SOTA) methods for 2D detection and 6D pose estimation reveals a significant performance degradation compared to standard household benchmarks, highlighting the unsolved challenges of industrial vision. XYZ-IBD establishes a new frontier for robust pose estimation in complex, high-occlusion, and reflective scenarios. The dataset and benchmark are publicly available at https://xyz-ibd.github.io.
Junwen Huang, Jiaqi Hu, Peter KT Yu +3
Technical University of Munich · Munich Center for Machine Learning (MCML) · XYZ Robotics +2
6D object pose estimation is fundamental to robotic manipulation and automation. Recent zero-shot methods have significantly improved generalization to unseen objects, but most still rely on large-scale pretrained models with substantial GPU computation and memory demands. These requirements complicate deployment on robotic platforms where perception, planning, and control share limited computational resources, while learned intermediate representations offer limited geometric interpretability for task-specific adaptation. To address these limitations, we propose G6D, a learning-free, geometry-driven RGB-D 6D pose solver. Given an RGB-D observation, an object instance mask, camera intrinsics, and a CAD model, G6D generates pose hypotheses through template-based geometric matching and refines them using silhouette and depth consistency, forming a purely geometry-driven pose estimation paradigm. This paradigm requires neither pretrained visual models nor target-specific training and preserves interpretable geometric representations throughout pose estimation. Moreover, adjustable hypothesis counts provide flexible accuracy-computation trade-offs, while a CPU-only configuration supports deployment without GPU resources. Experiments on LineMOD and five BOP19 datasets demonstrate advanced performance. Real-world pick-and-place experiments further demonstrate G6D's applicability to robotic manipulation. The complete project is publicly available at https://ai4control.github.io/G6D-Project-Page .