Pith. sign in

REVIEW 3 major objections 8 minor 4 cited by

This paper claims that one-shot 6D pose estimation of unseen objects, previously considered nearly impossible from a single image, becomes practical when a single-view 3D generator is combined with joint metric-scale and pose alignment and

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 21:24 UTC pith:R5VDDOA3

load-bearing objection A genuinely novel systems pipeline for one-shot 6D pose, but the claimed SOTA rests on an unverified generated-mesh fidelity assumption; still deserves a careful referee. the 3 major comments →

arxiv 2509.07978 v1 pith:R5VDDOA3 submitted 2025-09-09 cs.CV

One View, Many Worlds: Single-Image to 3D Object Meets Generative Domain Randomization for One-Shot 6D Pose Estimation

Zheng Geng , Nan Wang , Shaocong Xu , Chongjie Ye , Bohan Li , Zhaoxi Chen , Sida Peng , Hao Zhao This is my paper
classification cs.CV
keywords 6D pose estimationone-shot pose estimationsingle-image 3D generationgenerative domain randomizationmetric scale estimationrender-and-compare refinementrobot manipulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper claims that one-shot 6D pose estimation — recovering an unseen object's full 3D position and orientation in a query image from a single reference photo — becomes practical if a 3D model is first generated from that photo and then anchored to real-world metric scale. The proposed pipeline, OnePoseViaGen, turns one RGB-D anchor image into a textured mesh with a modified single-view 3D generator, aligns that mesh to the observed object through a coarse-to-fine module that alternates feature matching, PnP solving, closed-form scale estimation, and render-and-compare refinement, and optionally fine-tunes the pose network on synthetic data built from text-guided, texture-diversified variants of the model. According to the paper, this combination produces state-of-the-art numbers on three challenging benchmarks — for example ADD 81.27 versus 45.6 for Any6D on YCBInEOAT and AR 74.8 versus 28.6 on LM-O — and a dexterous robot hand completes pick-and-place trials at 73.3% versus 6.7–16.7% for the compared baselines. If the claim holds, robots working with long-tail household or factory objects could plan precision manipulation from a single snapshot, with no CAD model and no multi-view capture.

Core claim

On its own terms, the paper's discovery is that a scale-normalized, single-view-generated 3D model can serve as the object model for reliable 6D pose estimation, provided the model is metrically aligned to the anchor view. OnePoseViaGen first segments the anchor image and generates a textured mesh with a normal-bridging single-view generator, then runs coarse-to-fine alignment: SuperPoint and SuperGlue match rendered views against the anchor image, PnP gives an initial pose, a least-squares step recovers the global scale factor, and an iterative render-and-compare network refines pose while re-estimating scale after each update. For training, a text-prompted generator creates structurally co

What carries the argument

The carrying mechanism is the coarse-to-fine alignment module, which jointly recovers the metric scale and the 6D pose of a scale-normalized generated mesh by alternating (1) SuperPoint/SuperGlue matching between rendered views of the mesh and the anchor RGB-D image, (2) a PnP solve for the initial pose, (3) a closed-form least-squares estimate of the global scale from corresponding 3D points, and (4) render-and-compare refinement of pose, re-running the scale step after each update. The second pillar is the text-guided diversification chain: a modified Hi3DGen generates the normalized mesh from the anchor image, and Trellis, prompted with text, produces texture-diverse structural variants t

Load-bearing premise

The entire pipeline rests on the single-view-generated 3D model faithfully matching the true object's geometry, including sides never visible in the anchor image; if the model is wrong where it cannot be seen, feature matching and render-and-compare will converge to a pose of the generated artifact rather than of the real object.

What would settle it

Run the pipeline on objects with known CAD ground truth and correlate per-object pose error with the fidelity of the generated mesh — for example, compute Chamfer distance between each single-view-generated mesh and its ground-truth CAD model, then regress pose error (ADD or AR) on that distance. The paper's own reported failure on the 'ape' category, attributed to mismatch between reconstructed and ground-truth models, predicts a clear positive correlation; a sharper test is to take objects whose hidden sides differ substantially from the generated model (or were deliberately altered) and che

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • One-shot 6D pose estimation for novel objects no longer needs a CAD model, multi-view capture, or a per-object training stage: a single RGB-D anchor image suffices as the only reference.
  • Because scale is recovered jointly with pose, the output is directly usable in metric units for robot grasp planning, insertion, and handoff, rather than in the normalized units of a generated model.
  • The text-guided domain-randomization strategy converts one generated model into a large synthetic training set, raising average recall from 12.6% (no fine-tuning) to 52.4% on an annotated real-world set without any labeled real images.
  • On the three public benchmarks, the reported numbers exceed prior one-shot approaches by wide margins (e.g., LM-O AR 74.8 vs. 28.6), and real-robot pick-and-place succeeds at 73.3% versus 6.7–16.7% for baseline trackers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If pose accuracy is gated by how faithfully the generated mesh matches the object's unseen sides — as the paper's own 'ape' failure suggests — then improving single-view generation fidelity at hidden geometry should translate almost linearly into higher pose accuracy; this is testable by regressing per-object pose error against mesh-to-CAD Chamfer distance.
  • The closed-form scale estimate treats all residuals as zero-mean noise, so objects whose generated model is systematically wrong (e.g., a missing back side) will bias scale, rotation, and translation together; a visibility-weighted or occlusion-aware residual would be a natural extension.
  • The same coarse-to-fine trick of anchoring a scale-normalized generated model with one RGB-D view could extend beyond pose estimation to other metric tasks, such as grasp affordance prediction, object insertion, or teleoperation, that currently rely on expensive per-object scanning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 8 minor

Summary. OnePoseViaGen targets one-shot 6D pose estimation for unseen objects from a single RGB-D anchor plus query images. It first generates a scale-normalized, textured 3D mesh from the anchor image using a modified Hi3DGen, then coarsely aligns the mesh to the anchor via multi-view SuperPoint/SuperGlue matching and PnP, recovers metric scale by closed-form least squares, and refines pose/scale iteratively with a FoundationPose-style render-and-compare module. An optional per-object text-guided Trellis diversification generates synthetic training data for fine-tuning the pose network. The authors report large improvements over Oryon, LoFTR, Gedi, Any6D on YCBInEOAT, TOYL, and LM-O (e.g., ADD 81.27 vs 45.6; AR 55.7 vs 43.3; AR 74.8 vs 28.6), an ablation showing fine alignment is critical, and real-world robotic grasping success rates.

Significance. If the results are reproducible, the paper makes a strong practical contribution: it turns single-view generative 3D reconstruction into a usable model source for one-shot metric pose estimation, with a clean scale-recovery formulation and a domain-randomization fine-tuning procedure. The benchmark gains are substantial, and the authors disclose that public-benchmark numbers exclude the optional fine-tuning, document failure modes, and report compute costs. The main unmet burden is verification of the load-bearing assumption that the generated mesh is faithful enough on unseen sides of the object, since no quantitative geometry-fidelity or oracle-model comparison is provided.

major comments (3)
  1. [§3.3, Appendix D] The central claim is that a single-view generated textured mesh suffices for reliable pose estimation. The manuscript never quantifies the fidelity of O_N to the true object. The only evidence is qualitative (Figs. 11, 14), and the authors state that on 'ape' the low accuracy is due to 'a mismatch between the reconstructed and ground-truth models'. Because Tables 1–3 report end-to-end metrics, they cannot separate geometry error from pose-estimation error. Please add (i) a quantitative geometry-fidelity measure between generated and GT meshes, including the back side (e.g., Chamfer distance or volumetric IoU over the full surface), and (ii) an oracle ablation that runs the identical alignment pipeline with the GT mesh instead of O_N. This directly tests whether high-fidelity generation is necessary or whether the alignment module is robust enough to accommodate shape error.
  2. [§3.4, scale optimization] The least-squares objective min_alpha sum ||alpha * P_hat_i - P_i||^2 treats all residual discrepancies as zero-mean noise. A systematic error in the generated shape—missing or hallucinated back geometry, wrong proportions—biases alpha as well as R and t because the matched visible points are not a random sample of the model error. The 'model imperfections' sentence in the same section acknowledges the risk, but no experiment probes it. Add a sensitivity analysis with controlled synthetic deformations on the unseen side, or report per-category scale error where GT dimensions are known, to demonstrate that recovered scale and pose do not drift under systematic shape bias.
  3. [§3.4, coarse alignment PnP] The PnP description as written is not reproducible. The text defines matched pairs (p'_i from rendered view o_i, p_i from I_A), lifts them to P'_i and P_i, then says 'apply the PnP algorithm on P'_i and their corresponding 2D projections p_i in image o_i'. Since p_i is in I_A, it cannot be a projection 'in image o_i'; if the intended 2D points are p'_i, PnP is degenerate. Please state exactly which 3D-2D correspondences are used—presumably P'_i in object coordinates with p_i in the anchor image. The coarse stage carries the rest of the pipeline, so this needs to be unambiguous.
minor comments (8)
  1. [Throughout] The dataset name is written both TOYL and TYOL (Tables 2/4, Figs. 12/13/14); use one abbreviation consistently.
  2. [Appendix C] 'The detailed information of all modules in One-2-3-Pose' appears to be a leftover from a different paper; this should be OnePoseViaGen.
  3. [References] References [84] and [95] are the same arXiv paper (TRELLIS), listed twice; merge them.
  4. [Appendix opening] 'In section D introduces a complete pipeline...' is ungrammatical, and the ground-truth pipeline is actually in Section E. Please fix the cross-reference.
  5. [Table 4] The table uses 'naive finetuning' while the text says 'Native fine-tuning'; align the terminology.
  6. [Abstract / §4.1] The abstract describes the public-benchmark results as state-of-the-art without noting that those numbers exclude the optional fine-tuning module; the disclosure appears only in Section 4.1. Add a parenthetical in the abstract or results summary.
  7. [Figure 2] The figure and its caption contain unresolved '??' placeholders in labels (e.g., 'Query Image(??) 6D Object Pose'); replace them with actual variables.
  8. [§3.4 notation] The superscript N in P_i^N is introduced inconsistently in the scale optimization paragraph; clarify the notation for metric camera-space points versus normalized model points.

Circularity Check

0 steps flagged

No significant circularity: external benchmarks and anchor-depth calibration carry the claims; only minor component-level self-citation.

full rationale

The headline SOTA numbers (Tables 1-3) are scored against external benchmarks (YCBInEOAT, TOYL, LM-O) with ground-truth poses and are reported without the optional fine-tuning (Sec. 4.1). The metric scale alpha is estimated from anchor depth by least squares (Sec. 3.4: min_alpha sum ||alpha P_hat_i - P_i^N||^2); this is calibration from available input, not a held-out prediction, and the relative pose T_A->Q in Eq. (1) uses a query-side FoundationPose estimate, so no reported number reduces to a fitted parameter by construction. The pipeline does rely on the authors' own Hi3DGen [82] as the 3D generator; this is component-level self-citation, but the central claim that generated geometry supports one-shot pose is validated end-to-end against external benchmarks, and the paper's own Appendix D exposes the generator-fidelity bottleneck (ape category: 'mismatch between the reconstructed and ground-truth models'), which is a limitation/correctness risk, not a hidden circular reduction. No uniqueness theorem is imported and no known result is renamed as a derivation. Hence score 2 for minor self-citation, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities. It rests on a chain of pre-trained models (Hi3DGen, SuperGlue, FoundationPose, Trellis), each carrying hidden training-data biases, and on unstated engineering hyperparameters that gate the reported gains (n views, iteration budgets, LoRA configuration). The headline benchmark results exclude fine-tuning, so the free parameters above mostly affect the base pipeline's alignment quality rather than the SOTA table.

free parameters (4)
  • number of coarse-alignment rendered views n
    Sec 3.4 renders the generated model from n spherical viewpoints; n is never stated, yet it controls coarse-stage runtime and match quality.
  • fine-alignment iteration budget
    Sec 3.4 refines 'until convergence or a maximum number of iterations'; neither the convergence criterion nor the maximum is specified.
  • LoRA hyperparameters for pose fine-tuning
    Appendix C reports '51 minutes to achieve 90% of the highest AR' for a single LoRA module; rank, learning rate, and data size are unspecified, so the 52.4% AR result cannot be reproduced from the text.
  • text prompts for Trellis diversification
    Diversity of the generated variants depends on VLM-produced prompts (Appendix A); the mapping from anchor image to prompt is described only by example, leaving the actual data distribution uncontrolled.
axioms (5)
  • domain assumption The Hi3DGen-generated mesh is geometrically consistent with the true object, including unobserved sides
    Sec 3.3 and Appendix D: the whole alignment and pose pipeline inherits generation quality; the authors' own failure analysis attributes the 'ape' failure to reconstructed versus ground-truth model mismatch.
  • domain assumption Anchor and query depth maps are metric and accurate enough for lifting matched pixels to 3D
    Sec 3.4 lifts 2D matches using 1/d K^-1 p; depth noise and missing depth at boundaries directly bias the scale alpha and the PnP solution.
  • domain assumption SuperPoint/SuperGlue correspondences between rendered synthetic views and real images are correct and numerous enough for PnP
    Sec 3.4 coarse stage: the view 'with the most matches' is selected; on textureless or blurred objects this fails, as documented in Figures 15 and 16.
  • domain assumption FoundationPose render-and-compare refinement converges from the coarse pose given a faithfully generated model
    Sec 3.4: the fine stage is FoundationPose's network applied to a generated mesh; the ablation (AR drops from 55.7 to 32.9 without it) shows this assumption carries most of the performance.
  • domain assumption Scale residuals are zero-mean, so the least-squares alpha is unbiased
    Sec 3.4, L2 = sum ||alpha P_hat - P^N||^2, treats generated-model error versus observed depth as noise; systematic model error biases alpha, R, and t jointly.

pith-pipeline@v1.3.0-alltime-deepseek · 24197 in / 19994 out tokens · 216502 ms · 2026-08-04T21:24:31.474558+00:00 · methodology

0 comments
read the original abstract

Estimating the 6D pose of arbitrary unseen objects from a single reference image is critical for robotics operating in the long-tail of real-world instances. However, this setting is notoriously challenging: 3D models are rarely available, single-view reconstructions lack metric scale, and domain gaps between generated models and real-world images undermine robustness. We propose OnePoseViaGen, a pipeline that tackles these challenges through two key components. First, a coarse-to-fine alignment module jointly refines scale and pose by combining multi-view feature matching with render-and-compare refinement. Second, a text-guided generative domain randomization strategy diversifies textures, enabling effective fine-tuning of pose estimators with synthetic data. Together, these steps allow high-fidelity single-view 3D generation to support reliable one-shot 6D pose estimation. On challenging benchmarks (YCBInEOAT, Toyota-Light, LM-O), OnePoseViaGen achieves state-of-the-art performance far surpassing prior approaches. We further demonstrate robust dexterous grasping with a real robot hand, validating the practicality of our method in real-world manipulation. Project page: https://gzwsama.github.io/OnePoseviaGen.github.io/

Figures

Figures reproduced from arXiv: 2509.07978 by Bohan Li, Chongjie Ye, Hao Zhao, Nan Wang, Shaocong Xu, Sida Peng, Zhaoxi Chen, Zheng Geng.

Figure 1
Figure 1. Figure 1: OnePoseViaGen. (Left) From a single anchor image, we generate a textured 3D model that lacks real-world scale and pose. Our coarse-to-fine alignment resolves this challenge, and optional domain-randomized fine-tuning further boosts robustness. (Right) While prior methods (Oryon, LoFTR, Gedi) largely fail in this one-shot setting, our approach achieves dramatic gains (ADD 81.3, ADD-S 93.1), enabling reliabl… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of OnePoseViaGen. Given an anchor RGB-D image IA, we generate a tex￾tured 3D model ON via our normal-guided generative pipeline. To ground ON in real-world metrics, we align it with IA via multi-view feature matching and render&compare framework. For a query image IQ, we apply a render&compare&selection pipeline [74]. Finally, the relative transformation TA→Q is computed. To enhance generalization… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of Coarse-to-fine Alignment Process. Motivation. Although Sec. 3.3 provides a 3D shape ON , it is in a normalized space, not matching the real object’s scale and pose in IA. For accurate metric-space 6D pose estimation, determining the correct scale and aligning the model with the observed object in the 3D scene is essential. This alignment is challenging due to discrepancies between the idealized… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results of robot manipulation. The left column is the anchor images, the middle col￾umn shows the generated models, and the right part shows the pose estimation results during robot manip￾ulation. To validate the practical applicability and robustness of our method, we conducted two real-world robotic manipulation tasks using 6D pose estimates as the percep￾tion input (See Appendix.D for detail… view at source ↗
Figure 5
Figure 5. Figure 5: Diversified Models. This figure showcases the results of two types of diversified texture generation. For each object, the original model is displayed on the right, while the model with diversified textures is shown on the left. It can be observed that our proposed method, which utilizes text prompts to promote texture diversity, is capable of generating rich and varied surface textures. This contributes t… view at source ↗
Figure 6
Figure 6. Figure 6: Generated training dataset. This figure illustrates the dataset generated using our diversified texture models. It is evident that our generated dataset encompasses a rich variety of backgrounds, object poses, occlusion relationships, and lighting conditions. This contributes to narrowing the domain gap between the training dataset and real-world scenarios. BlenderProc [118] rendering pipeline. For each sy… view at source ↗
Figure 7
Figure 7. Figure 7: Distribution Analysis of Training Dataset. Top-left: Distribution of Azimuth angles for objects within the generated dataset; Top-right: Distribution of Elevation angles for objects; Bottom￾left: Distribution of object distances from the camera; Bottom-right: Distribution of the proportion of object area visibility; These plots demonstrate that the constructed training dataset encompasses a rich diversity … view at source ↗
Figure 8
Figure 8. Figure 8: Finetune Performance with Time [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of Metrics on the LM-O Dataset. Left: Illustration of the MSPD metric variation with respect to the visible size of objects in pixels. Middle: Demonstration of the MSSD metric variation according to the object size in meters. Right: Presentation of the VSD metric variation with respect to τ (denoting the misalignment tolerance) and object size. These graphs effectively capture the metrics’ de… view at source ↗
Figure 10
Figure 10. Figure 10: Performance in LM-O dataset. In each image, the red, green, and blue lines represent the x, y, and z axes of the model, respectively, while the pink line shows the rendered contour under the estimated pose. By comparing the rendered contour with the ground-truth outline of the object, it is evident that our method is highly robust, performing well across various objects and under different occlusion scena… view at source ↗
Figure 11
Figure 11. Figure 11: Comparison Between Original and Generated Models on LM-O dataset. The first row displays the original object models, while the second row shows the generated models under the same pose. The third row presents a bottom-view comparison of the generated models. As can be seen, the generated models exhibit high quality and closely resemble the original objects in terms of texture and structure, demonstrating … view at source ↗
Figure 12
Figure 12. Figure 12: Visualization of Metrics on the TYOL Dataset. Left: Illustration of the MSPD metric variation with respect to the visible size of objects in pixels. Middle: Demonstration of the MSSD metric variation according to the object size in meters. Right: Presentation of the VSD metric variation with respect to τ (denoting the misalignment tolerance) and object size. These graphs effectively capture the metrics’ d… view at source ↗
Figure 13
Figure 13. Figure 13: Performance in TYOL dataset. In each image, the red, green, and blue lines represent the x, y, and z axes of the model, respectively, while the pink line shows the rendered contour under the estimated pose. By comparing the rendered contour with the ground-truth outline of the object, it is evident that our method is highly robust, performing well across various objects and under different occlusion scena… view at source ↗
Figure 14
Figure 14. Figure 14: Comparison Between Original and Generated Models on TOYL dataset. The first row displays the original object models, while the second row shows the generated models under the same pose. The third row presents a bottom-view comparison of the generated models. As can be seen, the generated models exhibit high quality and closely resemble the original objects in terms of texture and structure, demonstrating … view at source ↗
Figure 15
Figure 15. Figure 15: Failure modes in real-world test. Left: The original image. Middle: Detection results showing the bounding box and model’s XYZ axes. Right: Rendered model image based on the detected pose. This visualization demonstrates that our method struggles in scenarios characterized by high symmetry, significant occlusion, and severe motion blur. Despite these challenges, the method shows promise in more favorable … view at source ↗
Figure 16
Figure 16. Figure 16: Failure modes in LMO and TOYL dataset. In each image, the red, green, and blue lines represent the x, y, and z axes of the model, respectively, while the pink line shows the rendered contour under the estimated pose. The top-right corner of each image displays an enlarged view of the corresponding failure region. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Real-world Robot Manipulation Examples. We recommend that readers visit the webpage to watch a dynamic demonstration video. tags serve as known geometric reference points, significantly improving the robustness and accuracy of the reconstruction process. A monocular camera is then used to record a video sequence at a fixed frame rate, ensuring full coverage of the most of AprilTags from multi￾ple viewpoin… view at source ↗
Figure 18
Figure 18. Figure 18: Testing dataset examples. The dataset we have constructed is showcased from various views. It can be seen that our testing dataset includes a range of object poses, occlusion relationships, and distances from the camera, demonstrating its richness and diversity. Finally, we constructed the following test dataset. The dataset visualization is shown in [PITH_FULL_IMAGE:figures/full_fig_p029_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Distribution Analysis of Testing Dataset. Top-left: Distribution of Azimuth angles for objects within the generated dataset; Top-right: Distribution of Elevation angles for objects; Bottom￾left: Distribution of object distances from the camera; Bottom-right: Distribution of the proportion of object area visibility; These plots demonstrate that the constructed testing dataset encompasses a rich diversity o… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Video Models as Native 4D Renderers: World-Grounded Conditioning from Animated Mesh

    cs.CV 2026-07 conditional novelty 6.0

    Conditioning a video diffusion model on animated-mesh G-buffer maps (tracking + world position + normals) improves camera-and-object control over depth conditioning in the authors' 68-video benchmark.

  2. Pose Anything Anywhere:Model-free Object Poses from Arbitrary References

    cs.CV 2026-06 unverdicted novelty 6.0

    PANY is a multi-view transformer framework for model-free 6D object pose estimation from arbitrary sparse references that reports SOTA gains of +12% on YCB-V and +20% on LM-O.

  3. Reconstruction by Generation: 3D Multi-Object Scene Reconstruction from Sparse Observations

    cs.CV 2026-04 unverdicted novelty 6.0

    RecGen achieves state-of-the-art 3D multi-object scene reconstruction from sparse RGB-D views by combining compositional synthetic scene generation with strong 3D shape priors, outperforming SAM3D by 30%+ in shape qua...

  4. SAM 3D: 3Dfy Anything in Images

    cs.CV 2025-11 unverdicted novelty 6.0

    SAM 3D reconstructs 3D objects from single images with geometry, texture, and pose using human-model annotated data at scale and synthetic-to-real training, achieving 5:1 human preference wins.

Reference graph

Works this paper leans on

139 extracted references · 47 canonical work pages · cited by 4 Pith papers · 3 internal anchors

  1. [1]

    Z. Chen, A. Walsman, M. Memmel, K. Mo, A. Fang, K. Vemuri, A. Wu, D. Fox, and A. Gupta. Urdformer: A pipeline for constructing articulated simulation environments from real-world images.arXiv preprint arXiv:2405.11656, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    Srinivasan, J

    K. Srinivasan, J. Collins, E. Heiden, I. Ng, J. Bohg, and A. Garg. Dexmots: Dexterous manipulation with differentiable simulation. 2024

    2024

  3. [3]

    S. K. Dwivedi, D. Anti ´c, S. Tripathi, O. Taheri, C. Schmid, M. J. Black, and D. Tzionas. Interactvlm: 3d interaction reasoning from 2d foundational models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 22605–22615, 2025

    2025

  4. [4]

    Zhong, Y

    L. Zhong, Y . Zhang, H. Zhao, A. Chang, W. Xiang, S. Zhang, and L. Zhang. Seeing through the occluders: Robust monocular 6-dof object pose tracking via model-guided video object segmentation.IEEE Robotics and Automation Letters, 5(4):5159–5166, 2020

    2020

  5. [5]

    H. Xia, E. Su, M. Memmel, A. Jain, R. Yu, N. Mbiziwo-Tiapo, A. Farhadi, A. Gupta, S. Wang, and W.-C. Ma. Drawer: Digital reconstruction and articulation with environment realism. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 21771– 21782, 2025

    2025

  6. [6]

    J. Yang, B. Ivanovic, O. Litany, X. Weng, S. W. Kim, B. Li, T. Che, D. Xu, S. Fidler, M. Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self- supervision.arXiv preprint arXiv:2311.02077, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    S. Xue, S. Cheng, P. Kachana, and D. Xu. Neural field dynamics model for granular object piles manipulation. InConference on Robot Learning, pages 2821–2837. PMLR, 2023

    2023

  8. [8]

    Z. Chen, R. A. Potamias, S. Chen, and C. Schmid. Hort: Monocular hand-held objects reconstruction with transformers.arXiv preprint arXiv:2503.21313, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    Rashid, S

    A. Rashid, S. Sharma, C. M. Kim, J. Kerr, L. Y . Chen, A. Kanazawa, and K. Goldberg. Language embedded radiance fields for zero-shot task-oriented grasping. In7th Annual Con- ference on Robot Learning, 2023

    2023

  10. [10]

    Q. Wang, C. Deng, T. G. W. Lum, Y . Chen, Y . Yang, J. Bohg, Y . Zhu, and L. Guibas. Neural attention field: Emerging point relevance in 3d scenes for one-shot dexterous grasping.arXiv preprint arXiv:2410.23039, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    J. Wu, T. Xue, J. J. Lim, Y . Tian, J. B. Tenenbaum, A. Torralba, and W. T. Freeman. Single image 3d interpreter network. InEuropean Conference on Computer Vision, pages 365–382. Springer, 2016

    2016

  12. [12]

    Kulhanek, S

    J. Kulhanek, S. Peng, Z. Kukelova, M. Pollefeys, and T. Sattler. Wildgaussians: 3d gaussian splatting in the wild.arXiv preprint arXiv:2407.08447, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Matsuki, R

    H. Matsuki, R. Murai, P. H. Kelly, and A. J. Davison. Gaussian splatting slam. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039– 18048, 2024. 9

    2024

  14. [14]

    Dellaert, D

    F. Dellaert, D. M. Rosen, J. Wu, R. Mahony, and L. Carlone. Shonan rotation averaging: Global optimality by surfing so (p) n. InEuropean Conference on Computer Vision, pages 292–308. Springer, 2020

    2020

  15. [15]

    Q.-Y . Zhou, J. Park, and V . Koltun. Fast global registration. InEuropean conference on computer vision, pages 766–782. Springer, 2016

    2016

  16. [16]

    C. Choy, J. Park, and V . Koltun. Fully convolutional geometric features. InProceedings of the IEEE/CVF international conference on computer vision, pages 8958–8966, 2019

    2019

  17. [17]

    Simeonov, Y

    A. Simeonov, Y . Du, Y .-C. Lin, A. R. Garcia, L. P. Kaelbling, T. Lozano-P ´erez, and P. Agrawal. Se (3)-equivariant relational rearrangement with neural descriptor fields. In Conference on Robot Learning, pages 835–846. PMLR, 2023

    2023

  18. [18]

    A. Zeng, S. Song, M. Nießner, M. Fisher, J. Xiao, and T. Funkhouser. 3dmatch: Learning lo- cal geometric descriptors from rgb-d reconstructions. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1802–1811, 2017

    2017

  19. [19]

    X. Chen, H. Zhao, G. Zhou, and Y .-Q. Zhang. Pq-transformer: Jointly parsing 3d objects and layouts from point clouds.IEEE Robotics and Automation Letters, 7(2):2519–2526, 2022

    2022

  20. [20]

    Y . Wang, Z. Wang, M. Nakura, P. Bhowal, C.-L. Kuo, Y .-T. Chen, Z. Erickson, and D. Held. Articubot: Learning universal articulated object manipulation policy via large scale simula- tion.arXiv preprint arXiv:2503.03045, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Zeng, K.-T

    A. Zeng, K.-T. Yu, S. Song, D. Suo, E. Walker, A. Rodriguez, and J. Xiao. Multi-view self- supervised deep learning for 6d pose estimation in the amazon picking challenge. In2017 IEEE International Conference on Robotics and Automation (ICRA), pages 1386–1383, 2017. doi:10.1109/ICRA.2017.7989165

    work page arXiv 2017

  22. [22]

    Kappler, F

    D. Kappler, F. Meier, J. Issac, J. Mainprice, C. G. Cifuentes, M. W ¨uthrich, V . Berenz, S. Schaal, N. Ratliff, and J. Bohg. Real-time perception meets reactive motion generation. IEEE Robotics and Automation Letters, 3(3):1864–1871, 2018

    2018

  23. [23]

    B. Wen, W. Lian, K. Bekris, and S. Schaal. CatGrasp: Learning category-level task-relevant grasping in clutter from simulation. InInternational Conference on Robotics and Automation (ICRA), pages 6401–6408, 2022

    2022

  24. [24]

    Blukis, T

    V . Blukis, T. Lee, J. Tremblay, B. Wen, I. S. Kweon, K.-J. Yoon, D. Fox, and S. Birchfield. One-shot neural fields for 3D object understanding. 2023

    2023

  25. [25]

    Zhuang, S

    C. Zhuang, S. Li, and H. Ding. Instance segmentation based 6d pose estimation of indus- trial objects using point clouds for robotic bin-picking.Robotics and Computer-Integrated Manufacturing, 82:102541, 2023

    2023

  26. [26]

    Eisner, Y

    B. Eisner, Y . Yang, T. Davchev, M. Vecerik, J. Scholz, and D. Held. Deep se (3)-equivariant geometric reasoning for precise placement tasks.arXiv preprint arXiv:2404.13478, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Y . Zhao, M. Bogdanovic, C. Luo, S. Tohme, K. Darvish, A. Aspuru-Guzik, F. Shkurti, and A. Garg. Anyplace: Learning generalized object placement for robot manipulation.arXiv preprint arXiv:2502.04531, 2025

    arXiv 2025

  28. [28]

    Mahler, J

    J. Mahler, J. Liang, S. Niyaz, M. Laskey, R. Doan, X. Liu, J. A. Ojea, and K. Goldberg. Dex- net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics.arXiv preprint arXiv:1703.09312, 2017

    Pith/arXiv arXiv 2017

  29. [29]

    W. Shen, G. Yang, A. Yu, J. Wong, L. P. Kaelbling, and P. Isola. Distilled feature fields enable few-shot language-guided manipulation.arXiv preprint arXiv:2308.07931, 2023. 10

    Pith/arXiv arXiv 2023

  30. [30]

    J. Kerr, C. M. Kim, M. Wu, B. Yi, Q. Wang, K. Goldberg, and A. Kanazawa. Robot see robot do: Imitating articulated object manipulation with monocular 4d reconstruction.arXiv preprint arXiv:2409.18121, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Caggiano, S

    V . Caggiano, S. Dasari, and V . Kumar. Myodex: Generalizable representations for dexterous physiological manipulation

  32. [32]

    T. Chen, J. Xu, and P. Agrawal. A system for general in-hand object re-orientation. In Conference on Robot Learning, pages 297–307. PMLR, 2022

    2022

  33. [33]

    Romero, H.-S

    B. Romero, H.-S. Fang, P. Agrawal, and E. Adelson. Eyesight hand: Design of a fully- actuated dexterous robot hand with integrated vision-based tactile sensors and compliant actuation. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1853–1860. IEEE, 2024

    2024

  34. [34]

    T. Chen, M. Tippur, S. Wu, V . Kumar, E. Adelson, and P. Agrawal. Visual dexterity: In-hand reorientation of novel and complex object shapes.Science Robotics, 8(84):eadc9244, 2023

    2023

  35. [35]

    J. Bohg, A. Morales, T. Asfour, and D. Kragic. Data-driven grasp synthesis—a survey.IEEE Transactions on robotics, 30(2):289–309, 2013

    2013

  36. [36]

    Driess, F

    D. Driess, F. Xia, M. S. Sajjadi, C. Lynch, A. Chowdhery, A. Wahid, J. Tompson, Q. Vuong, T. Yu, W. Huang, et al. Palm-e: An embodied multimodal language model. 2023

    2023

  37. [37]

    James, Z

    S. James, Z. Ma, D. R. Arrojo, and A. J. Davison. Rlbench: The robot learning benchmark & learning environment.IEEE Robotics and Automation Letters, 5(2):3019–3026, 2020

    2020

  38. [38]

    Z. Wei, Z. Xu, J. Guo, Y . Hou, C. Gao, Z. Cai, J. Luo, and L. Shao. D (r, o) grasp: A unified representation of robot and object interaction for cross-embodiment dexterous grasping.arXiv preprint arXiv:2410.01702, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Murali, B

    A. Murali, B. Sundaralingam, Y .-W. Chao, W. Yuan, J. Yamada, M. Carlson, F. Ramos, S. Birchfield, D. Fox, and C. Eppner. Graspgen: A diffusion-based framework for 6-dof grasping with on-generator training.arXiv preprint arXiv:2507.13097, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    P. R. Florence, L. Manuelli, and R. Tedrake. Dense object nets: Learning dense visual object descriptors by and for robotic manipulation.arXiv preprint arXiv:1806.08756, 2018

    Pith/arXiv arXiv 2018

  41. [41]

    Billard and D

    A. Billard and D. Kragic. Trends and challenges in robot manipulation.Science, 364(6446): eaat8414, 2019

    2019

  42. [42]

    D. Guo, F. Sun, H. Liu, T. Kong, B. Fang, and N. Xi. A hybrid deep architecture for robotic grasp detection. In2017 IEEE international conference on robotics and automation (ICRA), pages 1609–1614. IEEE, 2017

    2017

  43. [43]

    Y . Deng, X. Guo, Y . Wei, K. Lu, B. Fang, D. Guo, H. Liu, and F. Sun. Deep reinforcement learning for robotic pushing and picking in cluttered environment. In2019 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS), pages 619–626. Ieee, 2019

    2019

  44. [44]

    Sajjan, M

    S. Sajjan, M. Moore, M. Pan, G. Nagaraja, J. Lee, A. Zeng, and S. Song. Clear grasp: 3d shape estimation of transparent objects for manipulation. In2020 IEEE international conference on robotics and automation (ICRA), pages 3634–3642. IEEE, 2020

    2020

  45. [45]

    K. Ding, B. Chen, R. Wu, Y . Li, Z. Zhang, H.-a. Gao, S. Li, G. Zhou, Y . Zhu, H. Dong, et al. Preafford: Universal affordance-based pre-grasping for diverse objects and environments. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7278–7285. IEEE, 2024

    2024

  46. [46]

    Gervet, S

    T. Gervet, S. Chintala, D. Batra, J. Malik, and D. S. Chaplot. Navigating to objects in the real world.Science Robotics, 8(79):eadf6991, 2023. 11

    2023

  47. [47]

    Forster, M

    C. Forster, M. Pizzoli, and D. Scaramuzza. Svo: Fast semi-direct monocular visual odometry. In2014 IEEE international conference on robotics and automation (ICRA), pages 15–22. IEEE, 2014

    2014

  48. [48]

    Kaufmann, A

    E. Kaufmann, A. Loquercio, R. Ranftl, A. Dosovitskiy, V . Koltun, and D. Scaramuzza. Deep drone racing: Learning agile flight in dynamic environments. InConference on Robot Learn- ing, pages 133–145. PMLR, 2018

    2018

  49. [49]

    Sabbah, K

    M. Sabbah, K. Wojciechowski, H. Soh, D. Hsu, L. Righetti, N. Mansard, B. Watier, and V . Bonnet. Optimal motion prediction for human-to-robot handovers. 2025

    2025

  50. [50]

    M. Runz, M. Buffier, and L. Agapito. MaskFusion: Real-time recognition, tracking and reconstruction of multiple moving objects. 2018

    2018

  51. [51]

    Marchand, H

    E. Marchand, H. Uchiyama, and F. Spindler. Pose estimation for augmented reality: A hands- on survey.IEEE Transactions on Visualization and Computer Graphics (TVCG), 22(12): 2633–2651, 2015

    2015

  52. [52]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  53. [53]

    O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, et al. Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024

    Pith/arXiv arXiv 2024

  54. [54]

    Venkataraman, Y

    S. Venkataraman, Y . Wang, Z. Wang, Z. Erickson, and D. Held. Real-world offline rein- forcement learning from vision language model feedback.arXiv preprint arXiv:2411.05273, 2024

    Pith/arXiv arXiv 2024

  55. [55]

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakr- ishnan, K. Hausman, et al. Do as i can, not as i say: Grounding language in robotic affor- dances.arXiv preprint arXiv:2204.01691, 2022

    Pith/arXiv arXiv 2022

  56. [56]

    Zitkovich, T

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In Conference on Robot Learning, pages 2165–2183. PMLR, 2023

    2023

  57. [57]

    Y . J. Ma, W. Liang, G. Wang, D.-A. Huang, O. Bastani, D. Jayaraman, Y . Zhu, L. Fan, and A. Anandkumar. Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:2310.12931, 2023

    Pith/arXiv arXiv 2023

  58. [58]

    O’Neill, A

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024

    2024

  59. [59]

    Khazatsky, K

    A. Khazatsky, K. Pertsch, S. Nair, A. Balakrishna, S. Dasari, S. Karamcheti, S. Nasiriany, M. K. Srirama, L. Y . Chen, K. Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset.arXiv preprint arXiv:2403.12945, 2024

    Pith/arXiv arXiv 2024

  60. [60]

    Kumar, R

    V . Kumar, R. Shah, G. Zhou, V . Moens, V . Caggiano, A. Gupta, and A. Rajeswaran. Robo- hive: A unified framework for robot learning.Advances in Neural Information Processing Systems, 36:44323–44340, 2023

    2023

  61. [61]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Dif- fusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, page 02783649241273668, 2023. 12

    2023

  62. [62]

    W. Kehl, F. Manhardt, F. Tombari, S. Ilic, and N. Navab. SSD-6D: Making RGB-Based 3D Detection and 6D Pose Estimation Great Again . In2017 IEEE International Conference on Computer Vision (ICCV), pages 1530–1538, Los Alamitos, CA, USA, Oct. 2017. IEEE Com- puter Society. doi:10.1109/ICCV .2017.169. URLhttps://doi.ieeecomputersociety. org/10.1109/ICCV.2017.169

    doi:10.1109/iccv 2017

  63. [63]

    S. Peng, Y . Liu, Q. Huang, X. Zhou, and H. Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. InCVPR, 2019

    2019

  64. [64]

    B. Wen, C. Mitash, S. Soorian, A. Kimmel, A. Sintov, and K. E. Bekris. Robust, occlusion- aware pose estimation for objects grasped by adaptive hands. In2020 IEEE Interna- tional Conference on Robotics and Automation (ICRA), pages 6210–6217, 2020. doi: 10.1109/ICRA40945.2020.9197350

    arXiv 2020

  65. [65]

    J. J. Lim, H. Pirsiavash, and A. Torralba. Parsing ikea objects: Fine pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2992–2999, 2013

    2013

  66. [66]

    C. Wang, D. Xu, Y . Zhu, R. Mart ´ın-Mart´ın, C. Lu, L. Fei-Fei, and S. Savarese. Densefu- sion: 6d object pose estimation by iterative dense fusion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3343–3352, 2019

    2019

  67. [67]

    Chi and S

    C. Chi and S. Song. Garmentnets: Category-level pose estimation for garments via canon- ical space shape completion. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3324–3333, October 2021

    2021

  68. [68]

    X. Li, H. Wang, L. Yi, L. J. Guibas, A. L. Abbott, and S. Song. Category-level articulated object pose estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

    2020

  69. [69]

    Goodwin, S

    W. Goodwin, S. Vaze, I. Havoutis, and I. Posner. Zero-shot category-level object pose es- timation. Number 13699 in Lecture Notes in Computer Science, pages 516–532. Springer, 2022

    2022

  70. [70]

    G. Li, D. Zhu, G. Zhang, W. Shi, T. Zhang, X. Zhang, and J. Li. Sd-pose: structural discrep- ancy aware category-level 6d object pose estimation. InProceedings of the IEEE/CVF Winter Conference on applications of computer vision, pages 5685–5694, 2023

    2023

  71. [71]

    J. Lin, Z. Wei, Z. Li, S. Xu, K. Jia, and Y . Li. Dualposenet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3560– 3569, 2021

    2021

  72. [72]

    V . N. Nguyen, T. Groueix, M. Salzmann, and V . Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9903–9913, 2024

    2024

  73. [73]

    E. P. ¨Ornek, Y . Labb´e, B. Tekin, L. Ma, C. Keskin, C. Forster, and T. Hodan. Foundpose: Un- seen object pose estimation with foundation features. InEuropean Conference on Computer Vision, pages 163–182. Springer, 2024

    2024

  74. [74]

    B. Wen, W. Yang, J. Kautz, and S. Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17868–17879, 2024

    2024

  75. [75]

    Y . He, Y . Wang, H. Fan, J. Sun, and Q. Chen. Fs6d: Few-shot 6d pose estimation of novel objects. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pages 6814–6824, 2022. 13

    2022

  76. [76]

    Labb´e, L

    Y . Labb´e, L. Manuelli, A. Mousavian, S. Tyree, S. Birchfield, J. Tremblay, J. Carpentier, M. Aubry, D. Fox, and J. Sivic. Megapose: 6d pose estimation of novel objects via render & compare.arXiv preprint arXiv:2212.06870, 2022

    Pith/arXiv arXiv 2022

  77. [77]

    S. Moon, H. Son, D. Hur, and S. Kim. Co-op: Correspondence-based novel object pose estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

    2025

  78. [78]

    J. Sun, Z. Shen, Y . Wang, H. Bao, and X. Zhou. Loftr: Detector-free local feature match- ing with transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922–8931, 2021

    2021

  79. [79]

    Corsetti, D

    J. Corsetti, D. Boscaini, C. Oh, A. Cavallaro, and F. Poiesi. Open-vocabulary object 6d pose estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18071–18080, 2024

    2024

  80. [80]

    X. Liu, G. Wang, R. Zhang, C. Zhang, F. Tombari, and X. Ji. Unopose: Unseen object pose estimation with an unposed rgb-d reference image.arXiv preprint arXiv:2411.16106, 2024

    Pith/arXiv arXiv 2024

Showing first 80 references.