ICRA 2025 · Paper Collection
ICRA 2025 moves object pose toward three practical constraints: zero-shot industrial onboarding and complementary views under occlusion; category-level scale, shape, and multiple solutions; and sustained in-hand updates through contact and physics when vision fails.
12
Papers meeting the inclusion criteria
1
Best Paper Finalist
4
Explicitly zero-shot
3
Visual–tactile papers
Official ICRA 2025 statistics report 4,250 submissions and 1,606 accepted papers, plus 503 journal-paper presentations. This collection uses official proceedings to select main-conference papers directly addressing object 6D / 9D pose, shape–pose, articulated-part pose, or in-hand pose.
Collection in One Sentence
ZeroBP, GZS6D-BP, and Feature Exchange Transformer target real bin picking through position-aware correspondence, CAD templates with semantic PPF, and early multiview fusion, respectively, emphasizing interpretable geometry.
MonoDiff9D and OmniShape explicitly represent probabilistic pose / shape solutions. ViTa-Zero, low-resolution tactile factor graphs, and dynamic particle filtering turn contact into physical validation and recovery mechanisms for visual estimators.
Research Trends
Trends summarize task conditions and perception loops across 12 selected core papers, rather than official ICRA topic categories.
01 · Industrial Zero-Shot
ZeroBP addresses missing texture and repeated local shapes; GZS6D-BP addresses textured symmetry. Both target rapid new-part deployment when CAD is available but retraining is undesirable.
02 · Occlusion & Multi-View
The Best Paper Finalist models nonadjacent surfaces through depth-guided graphs. Feature Exchange Transformer exchanges multiview features early through epipolar cross-attention. Both explicitly identify relationships that remain informative under partial visibility.
03 · Pose as Distribution
MonoDiff9D generates 9D poses through diffusion; OmniShape separates normalized-frame projection and shape priors into conditional distributions. Probabilistic generation supplies multiple hypotheses for scale, symmetry, and unobserved shape.
04 · Touch as Constraint
Three visual–tactile papers use factor graphs, particle filters, and test-time optimization. Touch need not recover complete geometry: contact, penetration, slip, or displacement can constrain 6D pose after visual occlusion.
Data Statistics
Each paper belongs to one group by primary contribution; card tags retain overlapping attributes.
Brings zero-shot estimation into real industrial constraints through CAD correspondence, occlusion modeling, multiview feature exchange, and bin picking.
Extends monocular 9D pose and unified rigid/articulated networks to multiple shape–pose hypotheses and semantic inter-object relative poses.
Uses contact, hand proprioception, and particle / factor-graph inference to compensate for visual failures under occlusion, slip, and dynamic interaction.
Interpreting the statistics: No core paper primarily contributes a standard object-pose benchmark dataset; the emphasis is methods and systems. Four industrial and three visual–tactile papers together exceed half the collection, highlighting robotic manipulation relative to contemporary vision conferences.
Recommended Reading
The Finalist label comes from official ICRA 2025 awards. Other recommendations reflect methodological representativeness, problem completeness, and robotic deployment value.
ICRA 2025 Best Conference Paper Finalist. Explicitly models relations between nonadjacent surface points under occlusion with a depth-guided graph.
Addresses repeated local workpiece shapes by incorporating each scene instance's global position directly into correspondence.
Combines fast coarse localization from CAD templates and viewpoint-aligned semantic point-pair refinement in a complete zero-shot grasping system, evaluated on seven BOP core datasets.
Exchanges view features early across multiple backbone layers, allowing complementary visibility and epipolar geometry to shape learned representations.
Directly generates 9D pose through conditional diffusion, using probabilistic modeling instead of fixed shape priors, CAD, or depth sensors.
Unifies category-level and instance-level pipelines by formulating pose as set prediction.
Combines part-instance segmentation and NPCS prediction in a single-stage network, demonstrating 200 Hz visual feedback on a real Kinova robot.
Samples a joint multimodal distribution rather than forcing one shape and pose, representing the nonidentifiability of partial observations.
Extends pose to semantic relationships between two objects, connecting object pose with task-demonstration transfer.
Shows that even one binary contact sensor per finger link can strongly constrain pose when vision, contact, and nonpenetration physics are combined in a factor graph.
Uses contact, penetration, and kinematic constraints at test time to correct arbitrary visual backbones without training a new tactile network.
Paper List
Each card gives task boundaries, input modalities, and a short summary. Primary links point to DOI / IEEE Xplore, with public projects or preprints as supplements.
4 papers
Brings zero-shot estimation into real industrial constraints through CAD correspondence, occlusion modeling, multiview feature exchange, and bin picking.
Jingyang Liu · Zhenyu Lu · Lu Chen · Jing Yang · Chenguang Yang
Setting: Instance-level 6D pose of known objects in heavily occluded grasping scenes
Input: RGB-D + object model
Summary: A depth-guided GNN models complex surface-point relations under occlusion, then adaptively fuses object-mask and binary-code semantics to improve 2D–3D correspondence. Evaluates on LM-O, YCB-V, and real grasping, with particularly strong occlusion gains for some objects.
Jianqiu Chen · Zikun Zhou · Xin Li · Ye Zheng · Tianpeng Bao · Zhenyu He
Setting: 6D pose of randomly stacked, weak-texture workpieces without object-specific training
Input: RGB-D scene + CAD model
Summary: ZeroBP's Position-Aware Correspondence combines local features and global positions to reduce mismatches from missing texture and similar local regions. Reports a 9.1% improvement in mean correct-pose recall over existing zero-shot methods on ROBI.
Zijiang Zhang · Huimin Lu · Jintong Cai · Tohru Kamiya · Seiichi Serikawa
Setting: Industrial bin picking with new objects and textured symmetric workpieces, without retraining
Input: RGB-D + CAD model
Summary: Renders CAD templates online for coarse pose, then refines with camera-view-aligned semantic point-pair features. Retains training-free, deployable geometric estimation while reducing ambiguity on textured symmetric objects.
Momen Khalil · Vincent Dietrich · Slobodan Ilic
Setting: Multicamera complementary views for 6D pose in reflective, textureless, heavily occluded bin picking
Input: multi-view RGB-D
Summary: Feature Exchange Transformer combines self-attention and epipolar cross-attention for coarse-to-fine exchange across network layers. As a multiview enhancement to EpiSurfEmb, outperforms the original baseline and other multiview pose methods on ROBI.
5 papers
Extends monocular 9D pose and unified rigid/articulated networks to multiple shape–pose hypotheses and semantic inter-object relative poses.
Jian Liu · Wei Sun · Hui Yang · Jin Zheng · Zichen Geng · Hossein Rahmani · Ajmal Mian
Setting: Joint rotation, translation, and scale of unseen within-category instances from one RGB image
Input: monocular RGB
Summary: MonoDiff9D first constructs a point cloud from DINOv2 zero-shot coarse depth, fuses global image and point-cloud features, then denoises Gaussian noise into 9D pose with a timestep-conditioned Transformer. Requires no within-category CAD or shape priors.
Jiale Ren · Hong Liu · Jinfu Liu · Peifeng Jiang
Setting: One end-to-end network for category-level and instance-level RGB object pose
Input: single RGB + category prior features
Summary: CIPE directly regresses pose from RGB, using a learnable rotation representation to place rotation learning in Euclidean space. Prior-query fusion combines pretrained point-cloud features and bounding-box queries. Outperforms previous RGB-only methods on datasets for both tasks.
Jingshun Huang · Haitao Lin · Tianyu Wang · Yanwei Fu · Yu-Gang Jiang · Xiangyang Xue
Setting: Joint part-instance segmentation, part pose, and scale for unseen articulated objects
Input: point cloud
Summary: YOEO predicts pointwise semantic labels, centroid offsets, and NPCS, clusters part instances through center voting, then aligns separated NPCS regions with real point clouds for pose / size. Targets latency and error propagation in multistage articulated-pose pipelines.
Katherine Liu · Sergey Zakharov · Dian Chen · Takuya Ikeda · Greg Shakhnarovich · Adrien Gaidon · Rares Ambrus
Setting: Multiple complete-shape and pose hypotheses from one observation, without known 3D models or categories
Input: single partial observation
Summary: OmniShape decomposes completion into two conditional distributions: projection of observations into normalized object frames, and geometric priors on triplanar neural fields. Two diffusion models combine to generate multiple consistent pose–shape hypotheses.
Denis Hadjivelichkov · Sicelukwanda Zwane · Marc Peter Deisenroth · Lourdes Agapito · Dimitrios Kanoulas
Setting: Transfers semantically meaningful relative poses to new object pairs from one interaction demonstration
Input: single-view RGB-D example
Summary: Automatically detects interacting parts and establishes one-shot part correspondences through geometry and visual semantics. Given a demonstrated relative pose, predicts semantically equivalent cross-poses for new instances, such as tool–target bottleneck poses.
3 papers
Uses contact, hand proprioception, and particle / factor-graph inference to compensate for visual failures under occlusion, slip, and dynamic interaction.
Lukas Mack · Felix Grüninger · Benjamin A. Richardson · Regine Lendway · Katherine J. Kuchenbecker · Joerg Stueckler
Setting: Low-resolution binary contact supplements visual pose under severe multifinger-hand self-occlusion
Input: RGB-D + hand proprioception + binary tactile contacts
Summary: Initializes from visual estimates such as FoundationPose, then combines vision, hand proprioception, binary-contact distances, and nonpenetration constraints in a factor graph with robust costs. Simulation with 17 YCB objects and real multifinger-hand experiments show gains under high occlusion / visual noise.
Zicai Peng · Te Cui · Guangyan Chen · Haoyang Lu · Yi Yang · Yufeng Yue
Setting: Continuously corrects two objects' dynamic poses during held-object slip or pushing of a contacted object
Input: vision + tactile displacement
Summary: Fuses vision and touch through particle filtering: vision tracks the contacted object, while tactile displacement estimates held-object pose changes. Experiments with 13 geometrically different objects cover slip and forced motion, aiming to sustain precise manipulation as interaction states change.
Hongyu Li · James Akl · Srinath Sridhar · Tye Brady · Taskin Padir
Setting: New objects become occluded or suffer catastrophic visual tracking failure during grasping, transport, and bimanual handover
Input: visual pose backbone + tactile + proprioception
Summary: ViTa-Zero models gripper–object interaction as a spring–mass system: tactile contacts produce attractive forces, while proprioceptive / penetration constraints produce repulsion. Test-time optimization runs only when visual poses are physically infeasible. Works with FoundationPose or MegaPose without dedicated tactile training data.
Scope & Sources
Inclusion: official ICRA 2025 proceedings papers directly estimating object 6D / 9D pose, articulated-part pose, joint shape–pose, in-hand object pose, or inter-object relative pose as transferable manipulation goals.
Excluded: grasp pose, human / hand pose, camera and robot localization, SLAM, single-fruit / flower agricultural applications, surgical-instrument pose, policies taking object pose as input, and target-trajectory generation without current-state estimation from observations. Semantic Cross-Pose is grouped as a relational-pose extension to distinguish it from BOP tasks.
Primary sources: Official ICRA 2025 Program, Official proceedings contents, IEEE DOI metadata, official awards, and public author project pages. Classifications, summaries, and trends are editorial work and do not replace the original papers.
