IROS 2025 · Paper Collection
IROS 2025 extends pose into complete robotic-state pipelines: category-level representations and occlusion reasoning, visible keypoints and retrieval / efficient distillation, and monocular or event tracking under fast motion and failures.
17
Official proceedings papers
5
Category-level and canonical representations
4
Matching and efficient estimation
4
Dynamic pose tracking
We screened titles of all 1,986 official IROS 2025 proceedings records, checked abstracts, outputs, and experimental subjects for 21 related candidates, and included 17 papers.
View the official IROS 2025 paper directory →Collection in One Sentence
RCGNet, MK-Pose, PLANTPose, and canonical-representation work address category variation through geometry, multimodal keypoints, deformable meshes, and equivariant canonicalization. Occlusion-Aware Diffusion directly incorporates occlusion relations into generative inference.
dGeDi, VAPO, RAG-6DPose, and Uncertainty-Aware KD emphasize selecting, retrieving, and compressing knowledge. DynamicPose, RGBTrack, and event-camera work require recovery under fast motion, missing depth, and tracking failures.
Research Trends
Trends summarize methodological changes and robotic deployment value, rather than paper-count rankings.
01 · Canonical State
Upright & Forward-Facing learns semantic orientation, MK-Pose discovers multimodal keypoints self-supervised, and PLANTPose recovers pose and instance deformation, extending category representations beyond fixed NOCS regression.
02 · Selective Evidence
VAPO selects currently visible important keypoints; RAG-6DPose retrieves relevant CAD knowledge; dGeDi and uncertainty-aware KD avoid distilling unreliable local teacher information.
03 · Dynamic Pose
DynamicPose uses VIO and closed-loop refinement, OmniPose6D defines short-term monocular dynamics, RGBTrack addresses depth-free estimation and recovery, and events supplement frames with high-frequency flow.
04 · Task Coupling
Metal-object work isolates material and lighting variables, JENGA jointly evaluates selection and pose, and two surgical-tool papers address markerless initialization and kinematics for imitation learning.
Data Statistics
Each paper belongs to one group by primary contribution; tags and summaries retain cross-area attributes.
Addresses within-category deformation, symmetry, scale, and occlusion through geometry guidance, learnable keypoints, canonical frames, and conditional diffusion.
Builds robust instance-level pipelines through visible keypoints, local 3D descriptors, 3DGS feature alignment, CAD retrieval, knowledge distillation, and correspondence-free estimation.
Extends static framewise pose prediction to joint camera/object motion, absent depth, high speeds, and recovery across framewise failures.
Connects pose to real robots through metal-object benchmarks, stacked-object selection, differentiable rendering, and monocular tool kinematics.
11 / 17
Category representations and matching efficiency together account for 64.7% of the core collection, highlighting generalizable representations and evidence use in static pose.
4 papers
Directly studies dynamic object-pose tracking across RGB-D, monocular RGB, events, and visual–inertial fusion.
16 papers
Public PDFs, projects, or code were found during metadata verification; author manuscripts remain under review for detail-page preparation.
Recommended Reading
IROS does not use CVPR-style Highlight labels. Recommendations reflect problem representativeness, method completeness, data contributions, and robotic deployment value.
Connects geometric-feature prediction, cross-branch fusion, and scale-aware PnP in an RGB-only category-level pipeline, suitable as an entry point to this direction.
Priority readingModels deformation relative to a base mesh, pose, and synthetic-domain gaps together rather than treating crops only as detection targets.
Representative category-level approachExamines how category coordinate frames are defined through equivariance, symmetry losses, and semantic-orientation supervision.
Priority readingVision, point clouds, and category text jointly support learnable-keypoint discovery and graph relations, including evaluation in real household scenes such as HouseCat6D.
Representative category-level approachIncorporates observations, implicit correspondences, perceptual distance, and occlusion relations directly into diffusion inference.
Priority readingDesigns distillation and weak supervision around pose-specific local geometry requirements beyond generic correspondence recall.
Representative geometric approachAllocates keypoint budgets through continuous visibility and graph importance, balancing speed, keypoint count, and pose stability.
Oral / priority readingImplements retrieval augmentation as explicit multimodal CAD retrieval and pose decoding, illustrating its geometric-vision application independently of language-model analogies.
Representative tracking approachCombines visual–inertial compensation, depth-assisted 2D tracking, multiple hypotheses, and hierarchical refinement in a closed loop for fast-motion failures.
Representative datasetEstablishes large-scale synthetic dynamic data and an uncertainty-aware baseline for monocular short-term tracking under joint motion.
Representative tracking approachUses events for frequent velocity updates and image frames for slower global correction, demonstrating complementary asynchronous sensing for fast tracking.
Representative datasetTreats metallic specular reflection and lighting sensitivity as a distinct data problem, with BOP-compatible data and material-aware baselines.
Representative applicationJointly evaluates object selection and pose, connecting standard pose benchmarks with real depalletizing decisions.
Representative applicationUses instrument geometric primitives and pose-hypothesis spaces to extend differentiable rendering from refinement to single-observation markerless initialization.
Paper List
Each card gives task settings, input modalities, a short summary, and the official DOI. Links switch to site details after full-text deep reading is complete.
5 papers
Addresses within-category deformation, symmetry, scale, and occlusion through geometry guidance, learnable keypoints, canonical frames, and conditional diffusion.
Sheng Yu · Di-Hua Zhai · Yuanqing Xia
Setting: Single-RGB 6D pose under within-category shape and scale variation
Input: RGB
Summary: Predicts and fuses image and geometric features through a transformer, then estimates geometry-guided correspondences and solves RANSAC-PnP. Explicitly handles category-instance scale variation without depth dependence.
Marios Glytsos · Panagiotis P. Filntisis · George Retsinas · Petros Maragos
Setting: Joint 6D pose and instance deformation of agricultural objects such as bananas with large appearance and geometry variation
Input: RGB
Summary: Predicts lattice deformation relative to a base mesh and 6D pose, refining synthetic textures with Stable Diffusion. Introduces a banana-pose benchmark and compares with MegaPose and other methods, demonstrating deformation modeling for agricultural manipulation.
Bing Han · Ruitao Pan · Xinyu Zhang · Chenxi Wang · Zhi Zhai · Zhibin Zhao · Xuefeng Chen
Setting: Semantically meaningful upright and forward-facing canonical pose across categories and symmetries
Input: point cloud / 3D shape
Summary: Uses Siamese Vector Neurons for SE(3) equivariance and self-supervised shape–pose disentanglement. Siamese equivariant constraints, automatically generated upright-surface labels, and symmetry-aware losses learn stable upright and forward-facing canonical frames.
Yifan Yang · Peili Song · Enfan Lan · Dong Liu · Jingtai Liu
Setting: Category-level pose and dimensions without instance CAD or category shape priors
Input: RGB + point cloud + category text
Summary: Discovers keypoints self-supervised through attention queries and soft heatmaps, models keypoint relations with graphs, then fuses RGB, point-cloud, and category-text features for pose. Evaluates on CAMERA25, REAL275, and HouseCat6D without category shape priors.
Yanbin Xiong · Buzhen Huang · Hui Ma · Yu Liu · Jun Cheng
Setting: Category-level 6D pose under severe occlusion, evaluated in robotic grasping
Input: RGB-D
Summary: Conditions on 2D/3D observations and constrains reverse diffusion through implicit correspondences, perceptual distance, and object-occlusion relationships. Evaluates occluded scenes on public benchmarks and a UR5 grasping system.
4 papers
Builds robust instance-level pipelines through visible keypoints, local 3D descriptors, 3DGS feature alignment, CAD retrieval, knowledge distillation, and correspondence-free estimation.
Amir Hamza · Andrea Caraffa · Davide Boscaini · Fabio Poiesi
Setting: Compresses accurate but expensive local 3D descriptors into a student for zero-shot 6D pose
Input: point cloud / depth
Summary: Distills efficient student descriptors from GeDi using extensive occlusion and partial observations. A weakly supervised loss suppresses unreliable teacher knowledge in nonsalient local regions. Evaluates zero-shot pose on 5 BOP datasets.
Ruyi Lian · Yuewei Lin · Longin Jan Latecki · Haibin Ling
Setting: Regresses only currently informative visible object keypoints under occlusion and symmetry
Input: RGB
Summary: Generates 3D-keypoint visibility labels from object annotations and symmetry, computes continuous importance with PageRank, and injects both into pose networks. Supports CAD-based and CAD-free representations, evaluated on LM, LM-O, and YCB-V.
Kuanning Wang · Yuqian Fu · Tianyu Wang · Yanwei Fu · Longfei Liang · Yu-Gang Jiang · Xiangyang Xue
Setting: Retrievable CAD multiview appearance and 3D geometry support pose under occlusion
Input: RGB + CAD multiview images + point cloud
Summary: Builds an offline knowledge base of CAD multiview features and 3D points. ReSPC retrieves query-relevant viewpoints and geometry, and a retrieval-augmented decoder predicts 6D pose, emphasizing occlusion and unseen views.
Nassim Ali Ousalah · Anis Kacem · Enjie Ghorbel · Emmanuel Koumandakis · Djamila Aouada
Setting: Compresses large keypoint-pose teachers for limited devices while avoiding unreliable distillation
Input: RGB
Summary: Uses teacher keypoint uncertainty to regulate transfer and aligns intermediate features at key locations, emphasizing reliable pose-relevant knowledge. Compares accuracy, size, and efficiency on LINEMOD and SPEED+.
4 papers
Extends static framewise pose prediction to joint camera/object motion, absent depth, high speeds, and recovery across framewise failures.
Tingbang Liang · Yixin Zeng · Jiatong Xie · Boyu Zhou
Setting: Real-time 6D tracking under simultaneous fast camera/object motion and large ROI shifts
Input: RGB-D / VIO
Summary: VIO compensates camera-induced ROI displacement while depth strengthens 2D tracking. A VIO-guided Kalman filter predicts rotation and multiple pose candidates; hierarchical refinement feeds results to the next frame without new-object retraining.
Yunzhi Lin · Yipu Zhao · Fu-Jen Chu · Xingyu Chen · Wei-Yao Wang · Hao Tang · Patricio A. Vela · Matt Feiszli · Kevin J. Liang
Setting: Short-term monocular pose tracking with potentially moving cameras and objects
Input: monocular RGB video
Summary: Releases synthetic data and protocols for short-term monocular dynamic-object tracking, with an uncertainty-aware keypoint-refinement baseline. Real-data evaluation tests transfer from synthetic dynamics to real video.
Teng Guo · Jingjin Yu
Setting: Initializes and tracks from RGB video using metric CAD, without deployment-time depth input
Input: RGB video + CAD model
Summary: Adds binary-search object-depth recovery to FoundationPose-style render-and-compare, using XMem, Kalman filtering, and a state machine for masks, motion prediction, and recovery. Unknown CAD scale can also be inferred from an initial depth estimate.
Zhichao Li · Arren Glover · Chiara Bartolozzi · Lorenzo Natale
Setting: High-speed 6DoF tracking when conventional frames suffer blur or insufficient update rates
Input: events + RGB-D / low-rate RGB pose
Summary: Estimates object 6DoF velocity from event optical flow, using synchronized depth and interaction matrices for 2D–3D motion constraints. A 5 Hz RGB global-pose estimator suppresses integration drift. High-speed gains have quantitative support on four synthetic objects; real sequences are qualitative only.
4 papers
Connects pose to real robots through metal-object benchmarks, stacked-object selection, differentiable rendering, and monocular tool kinematics.
Thomas Pöllabauer · Michael Gasser · Tristan Wirth · Sarah Berkei · Volker Knauthe · Arjan Kuijper
Setting: Instance-level pose of highly reflective metal household and industrial objects under lighting and background changes
Input: RGB
Summary: Releases IGD, a BOP-compatible synthetic dataset with 60 metal objects and 90K training images. Separately adds curvature-keypoint heatmaps and ideal-material image reconstruction to GDRNPP. Both approaches are evaluated only on synthetic test data, without real metal scenes or combined-model results.
Sai Srinivas Jeevanandam · Sandeep Inuganti · Shreedhar Govil · Didier Stricker · Jason Rambach
Setting: Selects safely graspable top-layer instances from dense stacks and estimates their 6DoF poses
Input: RGB + IMU + CAD
Summary: YOLOv8 and RGB-only ZebraPose generate candidates; IMU gravity ranks brick heights, and modal / rendered-amodal mask-area ratios reject occluded objects. Custom top-1 precision-mAP is 0.79 on 12K synthetic samples; real construction-site results are video demonstrations only.
Zekai Liang · Zih-Yun Chiu · Florian Richter · Michael C. Yip
Setting: Markerless surgical-instrument calibration and tracking initialization from one surgical-view image
Input: monocular RGB + robot geometry
Summary: Starting with a 640×480 endoscopic RGB image, SAM2 mask, dVRK CAD / kinematics, and inaccurate joint angles, optimizes 500 candidates in an RCM-compatible four-dimensional space, then jointly refines camera–robot extrinsics and three distal joints. Real experiments evaluate RCM consistency and open-loop grasping, without absolute extrinsic ground truth or runtime. Online tracking remains future work.
Juo-Tung Chen · Xinhao Chen · Ji Woong Kim · Paul Maria Scheikl · Richard Jaepyeong Cha · Axel Krieger
Setting: Recovers surgical-tool trajectories, base poses, and joint angles from monocular surgical video for robot learning
Input: monocular RGB video + articulated tool model
Summary: Initializes the first frame with 324 candidates from a 3×3 mask-center grid and 36 viewing-axis angles. Joint-conditioned 3D Gaussian Splatting refines 6D pose and three joints, followed by framewise warm starts. ADE on two tasks is 9.7 / 12.0 mm; policies supervised by estimated trajectories achieve 70% / 60% success, below ground-truth-supervised 100% / 80%.
Scope, Sources & Deduplication
Inclusion: official IROS 2025 proceedings papers primarily outputting instance-level / category-level object 6D pose, object tracking, pose-specific local descriptors and geometric solvers, pose benchmarks, or explicit rigid / articulated tool states such as surgical instruments.
Excluded: human / hand / robot pose, camera localization, grasp pose alone, 3D-box tracking, and systems consuming existing poses for control. VISO-Grasp centers on active views and grasping; AlignCAPE evaluates category-agnostic 2D keypoints on MP-100; SplatPose estimates camera-to-3DGS-scene pose; Correspondence-Free Patterns evaluates camera / robot motion relative to wall patterns. All four are excluded from core 6D object-pose counts.
Deduplication: RA-L, T-RO, and T-ASE journal papers remain in journal collections even when presented at IROS. Formal publication follows IEEE DOI and IROS proceedings; arXiv, manuscripts, projects, and code repositories provide open reading links.
Primary sources: Complete DBLP IROS 2025 directory, IEEE proceedings DOI, paper DOIs, and public author materials. Summaries, task classifications, and trends are editorial work.
