Pith. sign in

REVIEW 5 major objections 5 minor 31 references

ARGUS: Aligning Robot Scene Geometry Under Shifting Views with Large 3D Vision Models

T0 review · 5 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read ARGUS re-renders robot observations from any camera viewpoint into one fixed canonical view before policy learning, and the authors claim this yields large gains in viewpoint generalization and training efficiency.

desk verdict A credible RGB-only viewpoint-normalization pipeline with strong internal results, but the headline claims about DROID/BridgeV2 and 4-6x efficiency outrun the evidence. read the letter →

arxiv 2608.05579 v1 pith:YDXTTHHB submitted 2026-08-06 cs.RO

Rishik Sathua , Haonan Chen , Katherine Driggs-Campbell This is my paper · ORCID
classification cs.RO
keywords viewpointgeneralizationvisualimitationlearning3Dvisionmodelscanonicaldiffusionpolicycameraextrinsicspointcloudalignmentrobotmanipulation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

ARGUS is a preprocessing pipeline that re-renders robot camera images from a fixed canonical viewpoint before they reach a visuomotor policy, using a large 3D vision model to reconstruct scene geometry rather than relying on depth sensors. The paper claims this removes viewpoint as a nuisance variable, so policies trained on either fixed-view or highly varied multi-view datasets become markedly more successful at novel camera angles and learn far more efficiently. In real-world evaluations across four manipulation tasks, ARGUS's diffusion-policy variant outperformed the KYC and VISTA baselines and classical diffusion policy by a wide margin, and it reached a 90-percent success rate 6x faster than KYC and 4x faster than classical diffusion policy. The deeper point is that canonicalizing geometry before learning may be a practical alternative to collecting massive viewpoint-diverse datasets in the hope that invariance will emerge on its own.

What carries the argument

The load-bearing mechanism is the canonicalization function $h: I \to \tilde{I}$ built from three sub-steps: (i) cross-view point-cloud estimation with VGGT, a pretrained transformer that outputs a point cloud and relative extrinsics from multiple RGB images; (ii) metric scaling and world-frame alignment, where the unknown global scale of VGGT's reconstruction is recovered from the ratio of baselines in the known real-world extrinsics to those predicted by VGGT, $$P^A = \frac{t_A - t_B}{$t^{{\text{VGGT}}$}_A - $t^{{\text{VGGT}}$}_B} $P^{{\text{VGGT}}$}_A,$$ followed by the transform $P_W = {}^W T_A P^A$; and (iii) rendering the aligned point cloud from a fixed canonical viewpoint chosen to maximize workspace visibility. This equation converts a scale-ambiguous, camera-frame reconstruction into a metric, robot-base-frame point cloud, and does so with any pair of calibrated cameras (the pair with the largest baseline for stability). The canonical render is the same at training and evaluation time, so the downstream policy never sees viewpoint variation and never needs to learn it away.

What would settle it

Run ARGUS on a subset of DROID or BridgeV2 where cameras are mounted freely and extrinsics are not pre-calibrated, while removing AprilTag-based calibration, and check whether the baseline-ratio step of Equation 1 can still be computed; if the metric scaling and world-frame alignment silently break without externally supplied extrinsics, the claim that ARGUS enables learning from large-scale viewpoint-diverse robot datasets is unsupported.

Watch

Extended reading notes

Core claim

The paper's central claim is that viewpoint entanglement, not scene content, is a principal obstacle to learning robust visuomotor policies, and that this obstacle can be removed by a deterministic preprocessing function $h$ that maps any observation from any camera viewpoint to a canonical rendered image before policy learning. ARGUS implements $h$ in three steps: estimate a scene point cloud with a large 3D vision model, recover its metric scale from the ratio of the real-world inter-camera baseline to the estimated baseline, align it to the robot base frame using calibrated extrinsics, and render the scene from a fixed virtual viewpoint. The authors show that this decomposition lets a diffusion policy trained on 100 demonstrations generalize to cameras shifted up to roughly $135^\circ$ from its training viewpoints, that convergence to a 90% success rate is 4–6x faster than baselines, and that feeding the scaled point cloud directly to a point-cloud policy (3D Diffuser Actor) matches using raw depth. They frame this as evidence that large-scale 3D vision models can reduce the learning burden on downstream policies and make viewpoint-diverse datasets such as DROID and BridgeV2 usable.

Load-bearing premise

The method requires an imitation-learning dataset with at least two calibrated external cameras that are always pointed at the task space, and the paper's evaluation supplies that calibration with AprilTags on a self-collected dataset rather than testing the method on the large public datasets it names as the motivation.

Editorial extensions

If this is right

  • A diffusion policy trained on ARGUS-canonicalized observations reaches 90% success on the towel-unfolding task about 6x faster than KYC and 4x faster than classical diffusion policy.
  • With only 40 demonstrations (80 unique viewpoints), ARGUS reaches roughly 70% success, indicating that canonicalization makes the downstream policy substantially more data-efficient.
  • Policies trained on a fixed, two-camera dataset generalize to novel camera configurations spanning the table (up to about 135 degrees of viewpoint change) with relatively consistent success.
  • When the scaled VGGT point cloud is fed directly into 3D Diffuser Actor instead of depth output, the success rate matches the depth-based version, suggesting learned 3D reconstruction can substitute for depth sensors in point-cloud policies.
  • Because ARGUS is a generic preprocessing layer, the same canonicalized observations can be reused with any visuomotor policy architecture, not only diffusion policies.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because ARGUS is a preprocessing step, it should combine with any downstream policy class, including vision-language-action models; the main barrier for closed-loop dynamic tasks is the reported ~0.52 s latency per prediction.
  • The same canonicalization principle could be extended beyond viewpoint to object-centric or scene-invariant frames, a direction the paper notes but does not explore; a testable step is to render from an object-relative camera rather than a fixed world-frame viewpoint.
  • The paper's reliance on calibrated extrinsics suggests a natural extension: replacing AprilTags with VGGT's predicted extrinsics entirely, which would remove the calibration requirement but likely compound the observed translational-shift errors on fine-grained tasks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes ARGUS, a preprocessing pipeline for visuomotor imitation learning that takes two or more RGB images from arbitrary camera viewpoints, reconstructs a 3D point cloud with the pretrained VGGT model, rescales and aligns it to the robot base frame using known real-world extrinsics, and re-renders the scene from a fixed canonical viewpoint before passing the image to a downstream policy. The authors evaluate on four real-world manipulation tasks under both fixed and viewpoint-diverse training datasets, reporting large margins over KYC, VISTA, and classical Diffusion Policy on viewpoint generalization, a 4-6x training efficiency improvement on a selected task, and comparable performance to a depth-based 3D Diffuser Actor when VGGT replaces depth sensing.

Significance. If the reported results are reproducible, ARGUS offers a simple and potentially impactful way to reduce viewpoint entanglement in visuomotor policies without relying on depth sensors. The pipeline contains no fitted parameters: VGGT is used off-the-shelf, metric scale is computed from known extrinsics, and the canonical viewpoint is a hand-selected constant. The real-world evaluation across four tasks and five novel camera configurations is a genuine strength, as is the candid limitations section that identifies precision loss and latency overhead. However, the headline claims are broader than the evidence, especially the applicability to public datasets such as DROID and BridgeV2 and the efficiency numbers, so the contribution is promising but not yet fully established.

major comments (5)
  1. [Abstract; Section 3.1] The central motivating claim that ARGUS enables efficient learning from large-scale public datasets such as DROID and BridgeV2 is not supported by the experiments. Section 3.1 assumes an imitation learning dataset with at least two calibrated external cameras with known real-world extrinsics that are always pointed toward the task space, and the appendix shows all experiments use AprilTag-calibrated cameras in a lab setting. The paper does not verify that DROID or BridgeV2 satisfy this assumption, nor does it report any experiment on those datasets or a sensitivity analysis to extrinsics noise. Since the metric scaling and world-frame alignment in Eq. (1) depend directly on this assumption, the abstract's claim that ARGUS enables learning from DROID/BridgeV2 is a scope mismatch rather than an established result.
  2. [Section 4.2.1 and Figure 8] The headline efficiency claim that ARGUS converges 4-6x faster is based on a single task (unfold towel) chosen as the highest-average-success task, evaluated under a single camera configuration sampled from region (a). No multiple seeds, confidence intervals, or full convergence curves are reported, so the reader cannot assess whether this factor is robust or an artifact of task and camera selection. This should be reframed as an illustrative case study or supported with repeated runs across tasks and camera configurations.
  3. [Equation (1), Section 3.4] Equation (1) as written divides the translation vector difference (t_A - t_B) by the VGGT translation vector difference, which is not a well-defined vector division. The metric scale should be the ratio of the baseline norms, ||t_A - t_B|| / ||t^VGGT_A - t^VGGT_B||. This is load-bearing because any error in the scale factor propagates directly into the world-aligned point cloud and canonical rendering.
  4. [Section 4.1.1, VISTA comparison] The VISTA comparison may not be apples-to-apples. The text states that ZeroNVS is finetuned on D_diverse to handle the roughly 135-degree viewpoint range, but VISTA is evaluated in the D_fixed setting. If the view-synthesis model is finetuned on the same diverse scenes and viewpoints used for evaluation, VISTA is given access to data not used to train ARGUS or classical Diffusion Policy in that setting; this needs to be clarified and, if true, the comparison should be rerun with ZeroNVS trained only on D_fixed. In addition, ARGUS and VISTA are trained to predict world-frame actions while KYC uses relative actions, which confounds the comparison.
  5. [Section 3.5] The canonical viewpoint rendering step is not specified precisely: the paper does not state how the reconstructed point cloud is rasterized or splatted into an RGB image, what resolution and intrinsics the canonical camera has, how occlusions and holes are handled, or whether the rendering is differentiable or fixed. This omission makes the pipeline difficult to reproduce and should be addressed in a revision.
minor comments (5)
  1. [Throughout] There are typos in the text, including 'highlite' in the contributions list, 'Viepoint' in the appendix title, and 'manipulation polices' in the abstract.
  2. [Table 1] The Unfold Towel row is garbled ('ARGUS9 7 978 8.0±0.447'); the per-camera counts should be formatted consistently so the mean and standard error are verifiable.
  3. [Figure 7 and Tables 1/2] The main text states results are reported in Figure 7 but does not point readers to the underlying numerical results in Tables 1 and 2; please add explicit table references.
  4. [Section 4.1.1 and Appendix A.1] The description of D_diverse says camera viewpoints are sampled 'after collecting each demo' and that 100 demos accumulate 200 unique camera positions per task; please clarify whether both cameras are moved per demo and how the 200 positions arise from 100 demos.
  5. [Appendix A.1] The appendix states that AprilTags are used to calibrate each camera 'at the start of each demo'; please describe the calibration procedure and report its accuracy, since these extrinsics are the only source of world-frame alignment for ARGUS.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ARGUS is a deterministic preprocessing pipeline with no fitted parameters, and its empirical comparisons are grounded in held-out evaluation data.

full rationale

The derivation chain is self-contained and contains no fitted parameters. ARGUS normalizes observations through a fixed function h: images to a canonical rendered image, using VGGT as an external, off-the-shelf 3D vision model, metric scaling from known real-world extrinsics (Eq. 1), and a hand-selected constant canonical viewpoint. The empirical claims in Sec. 4.1.2 and 4.2.2 compare this pipeline against baselines on self-collected AprilTag-calibrated data; no parameter of ARGUS is fitted to evaluation outcomes, and the canonical viewpoint is not optimized against success rates. The self-citations ([13], [18]) appear only in related-work summaries and are not load-bearing for the method. The only notable gap is a scope mismatch: the abstract motivates large-scale datasets like DROID and BridgeV2, but Sec. 3.1 requires at least two calibrated external cameras with known world-frame extrinsics, and all experiments use AprilTag-calibrated self-collected data. That is a correctness/scope risk, not circularity, because the method's internal logic does not equate its outputs with its inputs.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim depends on external 3D reconstruction quality and calibrated multi-camera inputs; no new entities are postulated. The only hand-chosen number is the canonical viewpoint pose. The scale factor in Eq. 1 is computed from known extrinsics, not fitted to outcomes.

free parameters (1)
  • Canonical viewpoint pose
    In Section 3.5 the canonical viewpoint is 'chosen to maximize visibility of the workspace'; it is a fixed hand-selected design choice, not fitted to data, but it determines the observation space the policy learns from.
assumptions (4)
  • domain assumption Dataset contains at least two calibrated external cameras whose views include the workspace
    Stated in Section 3.1; enables cross-view point cloud estimation and baseline-based metric scaling in Eq. 1.
  • domain assumption VGGT returns point clouds and relative extrinsics that are accurate up to a single global scale
    The entire pipeline in Sections 3.3-3.4 depends on this accuracy; Section 5 concedes that VGGT extrinsic errors cause translational shifts in canonical images.
  • domain assumption AprilTag-based extrinsics are accurate enough for the scale computation
    Section 4.1.1 uses AprilTags to provide camera extrinsics; any error propagates through Eq. 1 into the canonical render.
  • domain assumption Scene content (objects, table, robot base) is unchanged between training and evaluation
    Section 3.1 explicitly holds the task scene fixed; results do not address scene-level generalization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ARGUS: Aligning Robot Scene Geometry Under Shifting Views with Large 3D Vision Models." pith.science (2026). https://pith.science/paper/YDXTTHHB

@misc{pith2026260805579,
  author       = {Pith},
  title        = {Pith review of: ARGUS: Aligning Robot Scene Geometry Under Shifting Views with Large 3D Vision Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YDXTTHHB}},
  note         = {Machine review of arXiv:2608.05579}
}
read the original abstract

Large-scale visuomotor policies have demonstrated impressive performance across a wide range of robot manipulation tasks. However, despite this success, manipulation polices often entangle scene geometry with the corresponding viewpoint, learning where objects lie in an image rather than where it lies in the task space. This entanglement inherently limits the corresponding policy's ability to learn from viewpoint-diverse datasets (ex. DROID, BridgeV2) and generalize beyond the viewpoints captured in their training data. In this work, we present ARGUS, an observation pre-processing pipeline that uses large-scale 3D vision models to align image observations from arbitrary camera viewpoints into a canonical viewpoint before passing it to downstream visuomotor policies. Experiments across training datasets with varying levels of viewpoint diversity, from fixed multi-view camera configurations to highly varied camera placements, show that our method consistently outperforms prior approaches across both limited-view and view-diverse training regimes. In efficiency comparisons, ARGUS demonstrates an ability to learn from view-diverse data, converging to high success rates 4-6x faster than previous methods by leveraging a simplified observation space. Overall, our findings show that leveraging large-scale 3D vision models reduces the learning burden on visuomotor policies, enabling more efficient learning from large-scale, viewpoint-diverse robot datasets.

Figures

Figures reproduced from arXiv: 2608.05579 by the authors.

Figure 1
Figure 1. Overview of ARGUS. Given two RGB images from any two camera viewpoints, we first construct a 3D point cloud of the scene using a Large Vision Model (ex. VGGT [28]). The point cloud is then aligned and scaled to fit the approximate geometry of the real world before it is used to render a 2D RGB image from a canonical viewpoint. This rendered image is then fed into the downstream visuomotor (ex. Diffusion) policy whic… view at source ↗
Figure 2
Figure 2. DDiverse Viewpoints. Visualization of the camera pose diversity present within our di￾verse training datasets. To unify the diverse visual representations from unconstrained camera viewpoints in our datasets, we take inspiration from previous works in depth-based robot manipulation poli￾cies [24, 25, 23] and construct a point cloud of the scene. However, since our goal is to gener￾alize across viewpoints without rel… view at source ↗
Figure 3
Figure 3. Canonical Viewpoint Selection. The canonical viewpoint is chosen to maximize visi￾bility of the workspace and manipulated objects. Given the scaled and world-aligned point cloud Pscaled W , we render the scene from a fixed vir￾tual canonical viewpoint expressed in the robot base frame. This canonical pose is chosen to maximize visibility of the workspace (as seen in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: EvaluationTasks. Real-world evaluation across four manipulation tasks. (a) (b) (c) (d) (e) [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Evaluation-time Camera Viewpoint Positioning. Top-down visualization of camera sampling distributions across five configuration groups. Ellipses indicate the spatial extent of camera positions within each group. 4 Experiments To study ARGUS’s effectiveness as a data pr…
Figure 7
Figure 7. Figure 7: Average success rates across five novel multi-view camera configurations. (Left) Methods are trained on single-viewpoint configuration datasets and evaluated for generalization to unseen viewpoints. (Right) Methods are trained on viewpoint-diverse datasets and evaluate…
Figure 6
Figure 6. Figure 6: Experimental Setup. Labelled experimental setup highliting variable po￾sition cameras and april tags used to cali￾brate them. In both of these setting, we study the success rates of policies across 5 different camera configurations with each configuration sampled from …
Figure 8
Figure 8. Figure 8: Training and sample efficiency across methods. (Left) Training efficiency measured by the [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Success rate comparison between our method and a depth-based policy baseline. Instead, we feed the scaled VGGT-predicted 3D pointcloud directly into 3D Diffuser Actor’s transformer architecture. 4.3.2 Results We report the average success rate of both clas￾sical 3D Dif…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 11 canonical work pages

  1. [1]

    Khazatsky, K

    A. Khazatsky, K. Pertsch, S. Nair, A. Balakrishna, S. Dasari, S. Karamcheti, S. Nasiriany, M. K. Srirama, L. Y . Chen, K. Ellis, P. D. Fagan, J. Hejna, M. Itkina, M. Lepert, Y . J. Ma, P. T. Miller, J. Wu, S. Belkhale, S. Dass, H. Ha, A. Jain, A. Lee, Y . Lee, M. Memmel, S. Park, I. Radosavovic, K. Wang, A. Zhan, K. Black, C. Chi, K. B. Hatch, S. Lin, J. ...

    2024

  2. [2]

    O. X.-E. Collaboration, A. O’Neill, A. Rehman, A. Gupta, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jain, A. Tung, A. Bewley, A. Her- zog, A. Irpan, A. Khazatsky, A. Rai, A. Gupta, A. Wang, A. Kolobov, A. Singh, A. Garg, A. Kembhavi, A. Xie, A. Brohan, A. Raffin, A. Sharma, A. Yavary, A. Jain, A. Balakr- ishna, A. W...

    arXiv

  3. [3]

    Walke, K

    H. Walke, K. Black, A. Lee, M. J. Kim, M. Du, C. Zheng, T. Zhao, P. Hansen-Estruch, Q. Vuong, A. He, V . Myers, K. Fang, C. Finn, and S. Levine. Bridgedata v2: A dataset for robot learning at scale. InConference on Robot Learning (CoRL), 2023

    2023

  4. [4]

    E. Jang, A. Irpan, M. Khansari, D. Kappler, F. Ebert, C. Lynch, S. Levine, and C. Finn. BC-z: Zero-shot task generalization with robotic imitation learning. In5th Annual Conference on Robot Learning, 2021. URLhttps://openreview.net/forum?id=8kbp23tSGYv

    work page 2021

  5. [5]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Haus- man, B. Ichter, S. Jakubczak, T. Jones, L. Ke, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, L. X. Shi, J. Tanner, Q. Vuong, A. Walling, H. Wang, and U. Zhilinsky.π 0: A vision-language-action flow model for general robot control, 2026. URLhttps://arxiv. o...

    arXiv 2026

  6. [6]

    Ghosh, H

    Octo Model Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, C. Xu, J. Luo, T. Kreiman, Y . Tan, L. Y . Chen, P. Sanketi, Q. Vuong, T. Xiao, D. Sadigh, C. Finn, and S. Levine. Octo: An open-source generalist robot policy. InProceedings of Robotics: Science and Systems, Delft, Netherlands, 2024

    2024

  7. [7]

    A. Xie, L. Lee, T. Xiao, and C. Finn. Decomposing the generalization gap in imitation learning for visual robotic manipulation, 2023. URLhttps://arxiv.org/abs/2307.03659

    arXiv 2023

  8. [8]

    Pumacay, I

    W. Pumacay, I. Singh, J. Duan, R. Krishna, J. Thomason, and D. Fox. The colosseum: A benchmark for evaluating generalization for robotic manipulation, 2024. URLhttps: //arxiv.org/abs/2402.08191

    arXiv 2024

Show all 31 references
  1. [9]

    J. Gao, A. Xie, T. Xiao, C. Finn, and D. Sadigh. Efficient data collection for robotic manipula- tion via compositional generalization, 2024. URLhttps://arxiv.org/abs/2403.05110

    2024 arXiv

  2. [10]

    Y . Seo, J. Kim, S. James, K. Lee, J. Shin, and P. Abbeel. Multi-view masked world models for visual robotic manipulation, 2023. URLhttps://arxiv.org/abs/2302.02408

    2023 arXiv

  3. [11]

    Sadeghi, A

    F. Sadeghi, A. Toshev, E. Jang, and S. Levine. Sim2real view invariant visual servoing by recurrent control, 2017. URLhttps://arxiv.org/abs/1712.07642

    2017 arXiv

  4. [12]

    S. Tian, B. Wulfe, K. Sargent, K. Liu, S. Zakharov, V . Guizilini, and J. Wu. View-invariant policy learning via zero-shot novel view synthesis.arXiv, 2024

    2024

  5. [13]

    H. Chen, C. Zhu, S. Liu, Y . Li, and K. R. Driggs-Campbell. Tool-as-interface: Learning robot policies from observing human tool use. InProceedings of Robotics: Conference on Robot Learning (CoRL), 2025. 10

    2025

  6. [14]

    R. Gao*, A. Holynski*, P. Henzler, A. Brussee, R. Martin-Brualla, P. P. Srinivasan, J. T. Barron, and B. Poole*. Cat3d: Create anything in 3d with multi-view diffusion models.Advances in Neural Information Processing Systems, 2024

    2024

  7. [15]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. URLhttps://arxiv. org/abs/2003.08934

    2020 arXiv

  8. [16]

    J. Y . Zhang, A. Lin, M. Kumar, T.-H. Yang, D. Ramanan, and S. Tulsiani. Cameras as rays: Pose estimation via ray diffusion, 2024. URLhttps://arxiv.org/abs/2402.14817

    2024 arXiv

  9. [17]

    Jiang, J

    T. Jiang, J. Ji, X. Tan, J. Fang, A. Bhattad, V . Guizilini, and M. R. Walter. Do you know where your camera is? view-invariant policy learning with camera conditioning, 2025. URL https://arxiv.org/abs/2510.02268

    2025

  10. [18]

    H. Chen, Y . Niu, K. Hong, S. Liu, Y . Wang, Y . Li, and K. R. Driggs-Campbell. Predicting object interactions with behavior primitives: An application in stowing tasks. In7th An- nual Conference on Robot Learning, 2023. URLhttps://openreview.net/forum?id= VH6WIPF4Sj

    2023

  11. [19]

    Shridhar, L

    M. Shridhar, L. Manuelli, and D. Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. InProceedings of the 6th Conference on Robot Learning (CoRL), 2022

    2022

  12. [20]

    Grotz, M

    M. Grotz, M. Shridhar, T. Asfour, and D. Fox. Peract2: Benchmarking and learning for robotic bimanual manipulation tasks, 2024. URLhttps://arxiv.org/abs/2407.00278

    2024 arXiv

  13. [21]

    Gervet, Z

    T. Gervet, Z. Xian, N. Gkanatsios, and K. Fragkiadaki. Act3d: 3d feature field transformers for multi-task robotic manipulation, 2023. URLhttps://arxiv.org/abs/2306.17817

    2023 arXiv

  14. [22]

    Y . Ze, G. Yan, Y .-H. Wu, A. Macaluso, Y . Ge, J. Ye, N. Hansen, L. E. Li, and X. Wang. Gnfactor: Multi-task real robot learning with generalizable neural feature fields, 2024. URL https://arxiv.org/abs/2308.16891

    2024 arXiv

  15. [23]

    Gkanatsios, J

    N. Gkanatsios, J. Xu, M. Bronars, A. Mousavian, T.-W. Ke, and K. Fragkiadaki. 3d flowmatch actor: Unified 3d policy for single- and dual-arm manipulation, 2025. URLhttps://arxiv. org/abs/2508.11002

    2025 arXiv

  16. [24]

    T.-W. Ke, N. Gkanatsios, and K. Fragkiadaki. 3d diffuser actor: Policy diffusion with 3d scene representations.Arxiv, 2024

    2024

  17. [25]

    Goyal, J

    A. Goyal, J. Xu, Y . Guo, V . Blukis, Y .-W. Chao, and D. Fox. Rvt: Robotic view transformer for 3d object manipulation.arXiv:2306.14896, 2023

    2023 arXiv

  18. [26]

    Goyal, V

    A. Goyal, V . Blukis, J. Xu, Y . Guo, Y .-W. Chao, and D. Fox. Rvt-2: Learning precise manipu- lation from few demonstrations, 2024. URLhttps://arxiv.org/abs/2406.08545

    2024 arXiv

  19. [27]

    Rustler, V

    L. Rustler, V . V olprecht, and M. Hoffmann. Empirical comparison of four stereoscopic depth sensing cameras for robotics applications.IEEE Access, 13:67564–67577, 2025. ISSN 2169-

    2025

  20. [28]

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny. Vggt: Visual geometry grounded transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

    2025

  21. [29]

    Wang and E

    J. Wang and E. Olson. Apriltag 2: Efficient and robust fiducial detection. pages 4193–4198, 10 2016. doi:10.1109/IROS.2016.7759617

    2016

  22. [30]

    Sargent, Z

    K. Sargent, Z. Li, T. Shah, C. Herrmann, H.-X. Yu, Y . Zhang, E. R. Chan, D. Lagun, L. Fei-Fei, D. Sun, and J. Wu. ZeroNVS: Zero-shot 360-degree view synthesis from a single real image. CVPR, 2024, 2023. 11 Appendix A Additional Experimental Details and Results A.1 Hardware De...

    2024

  23. [3536]

    URLhttp://dx.doi.org/10.1109/ACCESS

    doi:10.1109/access.2025.3560810. URLhttp://dx.doi.org/10.1109/ACCESS. 2025.3560810

    2025

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.