ICRA 2026 · Paper Collection
ICRA 2026 treats pose as part of robotic feedback loops. Generative 3D and Gaussian representations reduce onboarding costs, event cameras support fast continuous tracking, and visual–tactile–contact physics connects estimation to assembly, in-hand manipulation, and grasping.
14
Papers meeting the inclusion criteria
13
Method paper
1
Official Oral
1
Datasets / benchmarks
Public availability: The official ICRA 2026 technical program lists titles, authors, keywords, and abstracts, with some author preprints available. Site detail pages are based on verifiable full texts; entries without public full text retain official abstract links rather than substituting abstracts for deep reading.
Counts distinguish formal publication venues: 14 ICRA proceedings papers belong to the main collection, plus 8 RA-L papers listed only in the ICRA presentation appendix without double-counting.
Collection in One Sentence
GFreeDet2, HIPPo, and InstantPose address immediate operation on first encountering an object. YOPO and the submillisecond shape–pose solver shorten inference through end-to-end learning and certifiable optimization, respectively.
Pose becomes a distribution, covariance, or contact-constrained joint state that must update through occlusion, collisions, missing modalities, and grasp decisions.
Research Trends
These trends combine 14 official ICRA papers and 8 RA-L presentations. Directional counts use only proceedings papers and are not official ICRA topic categories.
01 · Instant Onboarding
InstantPose uses one RGB-D image, HIPPo starts from image-to-3D priors and updates from online observations, and GFreeDet2 reconstructs renderable Gaussian objects from multiview RGB. All reduce offline asset preparation through on-site observation.
02 · Dynamic Pose
Two proceedings event-camera approaches explore keypoint–PnP and learning-free prediction–correction, while an RA-L presentation adds synthetic-to-real representations. MGS-Track reconstructs online 3DGS; TwinTrack uses contact dynamics as additional information under occlusion.
03 · Pose as Belief
Proceedings paper SE(3)-PoseFlow directly models multimodal pose distributions. RA-L presentation Aleatoric Pose Uncertainty adds per-frame measurement covariance to existing networks. Both support filtering, active views, or grasp planning beyond point estimates.
04 · Contact-Aware State
M-VTOP, InvariantCloud, and Active Tactile Exploration use particle beliefs, global marker registration, and information gain to preserve observability, recovery, and active improvement after visual occlusion. An RA-L distributed-tactile factor-graph presentation adds contact-physics constraints.
Data Statistics
Each paper belongs to one group by primary contribution; card tags retain cross-area attributes.
Reduces dependence on CAD, depth, target-specific training, and complex postprocessing while advancing real-time category-level joint shape–pose estimation.
Extends framewise estimation to continuous 6DoF tracking under fast motion, monocular depth ambiguity, severe occlusion, and contact impacts.
Incorporates vision, touch, contact constraints, and active exploration into state estimation for assembly, in-hand manipulation, and precision control.
Explicitly models pose distributions and measurement covariance while addressing data gaps for moving cameras, heavy clutter, and industrial assembly.
Interpreting the statistics: Generalization with fewer priors and dynamic tracking have substantial shares. Touch / contact and uncertainty also show evaluation expanding toward recoverability, temporal stability, calibration, and task success. Event-camera papers total 2 .
Recommended Reading
GFreeDet2's Oral label comes from the official ICRA schedule. Other recommendations reflect methodological representativeness, completeness, and robotic deployment value, rather than official rankings.
Encodes category-level 9DoF pose directly into detection-transformer queries and matching, offering a compact detection-as-pose design.
Brings certifiable geometric estimation back into learned pose pipelines, using a 4×4 eigenproblem for iterations on the order of 100 microseconds.
The collection's only main-conference Oral, using Gaussian objects as renderable CAD substitutes within foundation-model pose pipelines.
Uses contact dynamics to correct geometry, physical properties, and pose, making contact informative for tracking.
Combines high-frequency event-flow propagation with low-frequency template correction in an explicit prediction–correction tracker without deep networks.
Uses unified mask observations and belief-based particle filtering for modular fusion, targeting submillimeter closed-loop manipulation accuracy.
Models sampleable pose distributions on SE(3) and demonstrates how uncertainty changes viewpoint selection and grasp planning.
Paper List
Each card gives its ICRA paper code, task boundaries, input modality, and a short summary. Cards with completed deep notes open site details; others retain official PaperCept abstracts without claiming full-text review.
4 papers
Reduces dependence on CAD, depth, target-specific training, and complex postprocessing while advancing real-time category-level joint shape–pose estimation.
Jiahong Chen · JingHao Wang · Zi Wang · Ziwen Wang · Banglei Guan · Qifeng Yu
Setting: Instance-level 6D pose of textureless industrial objects from multiview RGB without reliable depth
Input: multi-view RGB
Summary: DKPMV strengthens keypoint prediction with attention aggregation and symmetry-aware training, then fuses dense multiview keypoint geometry through three-stage progressive optimization. The official abstract reports performance above multiview RGB and RGB-D comparisons on ROBI.
Hakjin Lee · Junghoon Seo · Jaehoon Sim
Setting: Joint detection, rotation, translation, and scale of multiple unseen category instances from one RGB image
Input: single monocular RGB
Summary: YOPO adds a lightweight pose head, box-conditioned translation, and a 6D-aware Hungarian cost to a detection Transformer for end-to-end detection and 9DoF pose. Requires no pseudo-depth, CAD, or additional data. The abstract reports 79.6% IoU50 and 54.1% at 10°10cm on REAL275.
Lorenzo Shaikewitz · Tim Nguyen · Luca Carlone
Setting: Joint shape, rotation, and translation from RGB-D semantic keypoints, with global-optimality assessment
Input: RGB-D + category-level semantic keypoints
Summary: Represents unknown shape through a linear active shape model, expresses MAP shape–pose estimation as a nonlinear eigenvalue condition over unit quaternions, and solves it with self-consistent field iteration. Lagrange multipliers provide a compact global-optimality certificate for fast outlier rejection.
Gu Wang · Xingyu Liu · Jingyi Tang · Chengxi Li · Yingyue Li · Ziqin Huang · Xiangyang Ji
Setting: Multiview RGB onboarding replaces CAD for unseen-object 2D / 6D detection in conventional or fisheye RGB queries
Input: multi-view reference RGB + query RGB
Summary: GFreeDet2 reconstructs 3D Gaussian objects from reference RGB using visual-hull initialization and projection-aware perspective cropping, unifying pinhole and fisheye cameras. Converts existing CAD-based unseen-object detectors to model-free estimation with limited changes. The abstract reports new state of the art on three BOP-H3 datasets.
5 papers
Extends framewise estimation to continuous 6DoF tracking under fast motion, monocular depth ambiguity, severe occlusion, and contact impacts.
Xingjian Yang · Ashis Banerjee
Setting: Novel-object initialization and real-time edge tracking under clutter, lighting changes, and abrupt motion
Input: RGB-D + object mesh
Summary: A unified lighting-invariant color-pair representation supports both live RGB-D-to-mesh initialization and lightweight temporal correspondence validation, avoiding incompatible features between accurate initialization and low-latency tracking.
Zhe Wang · Qijin Song · Zihao Li · Jingyu Xiao · Weibang Bai
Setting: Continuous known-object 6DoF tracking under motion blur, low light, and high-speed dynamics
Input: event stream + 3D model keypoints
Summary: Detects keypoints on event time surfaces, then continuously propagates them using event polarity, coordinates, and local density. 2D–3D keypoint hash correspondences and EPnP recover pose through an explicit event-geometry pipeline.
Wen Yang · Zhixian Xie · Yiting Wang · Vamsi Sai Abhijit Tadepalli · Heni Ben Amor · Shan Lin · Wanxin Jin
Setting: Continuous novel-object tracking under severe occlusion and fast motion in falling collisions and dexterous in-hand manipulation
Input: vision + contact dynamics
Summary: TwinTrack's Real2Sim stage initializes geometry visually, then jointly updates geometry residuals, mass, inertia, and friction through contact-dynamics consistency. Its Sim2Real stage adaptively fuses visual tracking and contact-dynamics predictions, exceeding 20 Hz on a customized MJX engine.
Zhiyuan Chen · Fan Lu · Guo Yu · Sanqing Qu · Ya Wu · Huang Yuan · Alois Knoll · Guang Chen
Setting: Online tracking and reconstruction of previously unseen objects from monocular video without depth sensors
Input: monocular RGB video + object masks
Summary: MGS-Track uses mask-augmented DUSt3R correspondences and point maps as geometric priors for online object-centric 3DGS, then jointly optimizes Gaussians and 6DoF pose coarse to fine. Capacity-aware densification and mask-consistent pruning control temporal model growth.
Zhichao Li · Chiara Bartolozzi · Lorenzo Natale · Arren Glover
Setting: Low-latency 6D tracking when fast motion exceeds RGB-D frame rates or causes blur
Input: event camera + object templates
Summary: Recovers 6D object velocity from event-based optical flow, propagates pose, and suppresses drift through template-based local correction. The official abstract reports that this learning-free method outperforms some existing methods on fast-moving objects.
3 papers
Incorporates vision, touch, contact constraints, and active exploration into state estimation for assembly, in-hand manipulation, and precision control.
Miquel Oller · Qiyang Qian · Radu Corcodel · Siddarth Jain
Setting: Closed-loop in-hand pose of small precision objects under occlusion, noise, or missing modalities
Input: vision + tactile + contact + CAD
Summary: M-VTOP unifies visual, tactile, and contact observations as geometry-centered masks. Particle filtering maintains and refines a multimodal belief. Requires object CAD but no task-specific retraining, with graceful degradation under noisy or missing modalities.
Pengfei Ye · Yuxiang Ma · Yi Zhou · Wei Chen · Wenzhen Dong · Molong Duan
Setting: Stable tactile-marker-only 6DoF tracking over long in-hand sequences, especially reducing yaw drift
Input: vision-based tactile marker point clouds
Summary: InvariantCloud uses global invariance and unique indexing of tactile-surface marker constellations for one-shot globally invariant point-cloud registration. Targets improved Z-axis yaw, long-sequence drift, and repeatable relocalization after tracking loss compared with local accumulation.
Ethan Kroll Gordon · Bruke Baraki · Hien Bui · Michael Posa
Setting: Tactile-only joint shape and position recovery with minimal robot motion when contacts move unknown objects
Input: tactile contact sequence
Summary: Identifies cuboid / convex-polyhedral geometry online through a constraint-violation loss avoiding rigid-contact numerical stiffness, then plans contacts by Expected Information Gain. Random data from less than 10 seconds after first contact yields usable models; active exploration accelerates convergence further.
2 papers
Explicitly models pose distributions and measurement covariance while addressing data gaps for moving cameras, heavy clutter, and industrial assembly.
Yufeng Jin · Niklas Wilhelm Funk · Vignesh Prasad · Zechu Li · Mathias Franzius · Jan Peters · Georgia Chalvatzaki
Setting: Full multimodal pose distributions under symmetry, occlusion, or partial visibility
Input: RGB-D
Summary: SE(3)-PoseFlow performs flow matching on the Lie-group manifold, representing multimodal pose posteriors under symmetry and severe occlusion with samples. Alongside results on Real275, YCB-V, and LM-O, applies distributions to active-view disambiguation and uncertainty-aware grasp synthesis.
Panagiotis Sapoutzoglou · Orestis Vaggelis · Athina Zacharia · Evangelos Sartinas · Maria Pateraki
Setting: Industrial tools and components in real assembly, covering instance-level and novel-object 6D pose
Input: RGB-D + static onboarding sequences
Summary: IndustryShapes' classic set includes approximately 4.6K images and 6K poses across single-object, multi-object, and same-category multi-instance scenes. Its extended set adds static RGB-D onboarding sequences for model-free / sequence-based methods, addressing the household-to-industry domain gap.
Journal Presentation Appendix
The following 8 papers appear in the ICRA 2026 program but are formally published in IEEE RA-L. Presentation and detail links are retained, while papers are excluded from ICRA proceedings totals and directional counts.
Francesco Di Felice · Alberto Remus · Stefano Gasperini · Benjamin Busam · Lionel Ott · Stefan Thalhammer · Federico Tombari · Carlo Alberto Avizzano
InstantPose generates object models from very few views using a Large Reconstruction Model, then estimates unseen-object instance-level 6D pose from a single RGB-D query. Compares on YCB-V with methods requiring high-quality geometry and evaluates pose utility through real grasping.
Yibo Liu · Zhaodong Jiang · Binbin Xu · Guile Wu · Yuan Ren · Tongtong Cao · Bingbing Liu · Rui Heng Yang · Amir Rasouli · Jinjun Shan
HIPPo Dreamer combines multiview diffusion and a 3D reconstruction foundation model to generate an initial mesh from one observation within seconds. Measurement-guided optimization subsequently updates geometry, appearance, and pose together, progressively replacing generative priors with online observations.
Ryo Okumura · Tadahiro Taniguchi
PartPose selects a pose-estimation rigid ROI through Bayesian optimization before render-and-compare, avoiding a whole-object rigidity assumption. Across nine object categories, the abstract reports substantially better translation, rotation, and grasp success than direct FoundationPose use.
Oussama Abdul Hay · Xiaoqian Huang · Muhammad Ahmed Humais · Abdulla Ayyad · Randa Almadhoun · Yahya Zweiri
Trains an Augmented Event Encoder on domain-randomized synthetic event templates, mapping real queries and templates into a matchable latent space. The official abstract claims the first learned event-based 6D object-pose method and evaluates on E-POSE.
Mark Van der Merwe · Kei Ota · Dmitry Berenson · Nima Fazeli · Devesh Jha
Combines local tactile geometry with kinematic and force constraints from quasi-static rigid interactions in a factor graph. Pose and extrinsic contacts jointly satisfy observations and physics, resolving ambiguity from local tactile sensing more effectively than geometry-only or simple contact-informed pipelines.
Thomas Jantos · Stephan Weiss · Jan Steinbrener
Freezes the original pose predictor and adds detached MLPs for rotation and translation aleatoric uncertainty, feeding pose and predicted covariance into an EKF. This low-cost modular design suits object-relative navigation and edge deployment better than fixed covariance.
Xiangting Meng · Jiaqi Yang · Mingshu Chen · Chenxin Yan · Yujiao Shi · Wenchao Ding · Laurent Kneip
DynOPETs generates pseudo-labels through pose estimation and tracking, then globally refines them with pose-graph optimization, reducing annotation difficulty with simultaneous camera and object motion. The official abstract reports systematic evaluation of 18 existing methods.
Seunghyeok Back · Joosoon Lee · Kangmin Kim · Heeseon Rho · Geonhyup Lee · Raeyoung Kang · Sangbeom Lee · Sangjun Noh · Youngjin Lee · Taeyeop Lee · Kyoobin Lee
Contains 1,000 highly cluttered scenes, 200 objects, and 75 environment configurations, averaging 14.1 objects and 62.6% occlusion per scene. Provides 736K 6D poses and 9.3B feasible grasps on 52K RGB-D images, with grasp-network training gains in simulation and reality.
Scope & Sources
Inclusion: ICRA 2026 technical-program papers directly outputting object 6D / 9D pose, pose distributions, object tracking, in-hand pose, or datasets and evaluations for these tasks. Multipart objects are included only when graspable rigid-part pose is explicitly estimated.
Excluded: human / hand pose, robot localization, camera pose / SLAM, satellite and drone ego-attitude, grasp pose, 3D detection / tracking alone, and manipulation policies taking pose as given. Active Tactile Exploration is grouped as a tactile extension because it outputs contact-dependent rigid shape and location state rather than standard BOP pose.
RA-L, T-RO, and T-ASE papers presented at ICRA enter only the journal-presentation appendix, avoiding duplicate proceedings counts. Cards use canonical slugs for formal journal detail pages.
Primary sources: Official ICRA 2026 Program, PaperCept Online Program and project / code pages disclosed in abstracts. Classifications, summaries, and trends are editorial work and do not replace formal papers.
