Pith. sign in

REVIEW 2 major objections 6 minor 24 references

Free-space blade affordances must be inferred as 3D occupancy, not SE(3) poses, to recover multi-modal solutions from unimodal stow data.

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 · grok-4.5

2026-07-12 11:40 UTC pith:SZDBE4IB

load-bearing objection Clean domain argument with diagnostic ablations: spatial reconstruction recovers multi-modal free-space blade affordances from unimodal production stows; cost proxy is the softest link but scoped honestly. the 2 major comments →

arxiv 2607.02549 v1 pith:SZDBE4IB submitted 2026-06-25 cs.CV cs.RO

Learning 3D Affordances for Blade Insertion in Cluttered Stowing

Tianyu Li , Harpreet Sawhney , Minju Jung , Aditya Mehrotra , Kunal Mehrotra , Mudit Agrawal This is my paper
classification cs.CV cs.RO
keywords free-space affordances3D occupancyblade insertionrobotic stowingmasked autoencodermulti-modal predictionSE(3)voxel affordances
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.

The paper argues that free-space affordances for rigid tools—volumes of 3D space a blade can safely occupy inside a cluttered bin—are spatial objects, and that standard methods ask this geometric question in the wrong domain by predicting SE(3) pose distributions. Production stow systems produce abundant but unimodal logs: one executed insertion per bin state, even when many placements are geometrically valid. Generative pose models therefore collapse to a single mode and cannot invent unobserved alternatives. VulcanVoxel instead trains a masked autoencoder to reconstruct blade occupancy from scene geometry in multi-channel voxel space, scoring feasibility locally at each voxel so multiple regions can activate at once. On 1,000 real warehouse cycles it reaches top-5 coverage of 0.89 versus 0.71 for the best pose-based baseline, while a distilled RGB student runs in 30 ms. A sympathetic reader cares because this turns unlabeled production logs into multi-modal free-space reasoning and frames free-space tool use as the complement to surface-contact affordance learning.

Core claim

Blade affordances are subsets of 3D space defined by geometric feasibility and should be inferred in that native domain. A reconstruction objective over multi-channel occupancy fields recovers multi-modal, geometrically valid blade placements from inherently unimodal production data; SE(3) pose models and even spatial flow-matching objectives learn only the observed unimodal distribution and cannot recover unexecuted but feasible alternatives.

What carries the argument

VulcanVoxel: a 3D U-Net masked autoencoder over multi-channel occupancy grids (wall/object/blade plus mask indicators) that reconstructs blade occupancy conditioned on scene geometry, then projects connected components to SE(3) poses by volumetric scoring. Local per-voxel feasibility is what allows multiple independent modes to activate from unimodal labels.

Load-bearing premise

The claim depends on a hand-designed mesh collision-and-reach cost, with fixed safety margins and a threshold set relative to production mean, being a faithful proxy for operational geometric validity despite ignoring force feedback and single-view reconstruction error.

What would settle it

On held-out real bins, if top-5 poses from VulcanVoxel’s occupancy fields meet the geometric cost threshold no more often than the best SE(3) flow-matching baseline, or if those low-cost poses do not succeed at higher rates when actually executed on the stow robot, the central claim fails.

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

If this is right

  • Reconstruction objectives, not merely spatial output representations, are required to extract multi-modal free-space affordances from unimodal logs.
  • Failed production episodes become training signal because free-space volumes can still be identified from scene geometry alone.
  • RGB distillation shows free-space affordances are grounded in visually accessible geometry and can run at real-time rates.
  • Coverage (top-K geometric validity) is the right operational metric when a planner can choose among diverse candidates.
  • The same spatial-domain argument applies to other rigid-tool free-space tasks beyond blade stowing.

Where Pith is reading between the lines

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

  • The domain argument likely generalizes to other unimodal production policies—pick, place, peg-in-hole—where geometry admits alternatives the logs never show.
  • Downstream planners that consume the continuous occupancy field rather than post-hoc top-K poses could exploit the multi-modal structure more directly.
  • Single-view voxel carving and 32³ resolution are practical ceilings; multi-view depth or finer grids are natural tests of how much remaining coverage gap is reconstruction noise.

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

2 major / 6 minor

Summary. The paper argues that free-space blade-insertion affordances are spatial objects and should be inferred as 3D occupancy rather than SE(3) pose distributions. VulcanVoxel is a channel-masked 3D U-Net autoencoder that reconstructs blade occupancy conditioned on multi-channel scene occupancy from single-view RGB-D, then extracts poses by geometric scoring of connected components. Trained on ~10k unannotated production stows, it reports Cov5 = 0.89 vs 0.71 for the best pose baseline (Voxel FM Pose), with a diagnostic ablation (Voxel FM Voxel) showing that a generative objective on the same spatial output collapses multi-modality (Cov100 0.07 vs 0.71). A distilled RGB student runs in 30 ms. The authors release the StowAffordance dataset and evaluate temporal, high-clutter, and production-failure hold-outs.

Significance. If the domain argument and ablations hold, this is a clear, practically relevant contribution to free-space affordance learning for rigid tools, complementary to surface-contact affordance work. Strengths include: (i) controlled isolation of representation vs objective (Tables 3–4), especially Voxel FM Voxel vs VulcanVoxel; (ii) real warehouse scale without human labels; (iii) a released RGB-D + trajectory dataset; (iv) an efficient distilled student with partial multi-modality retained. The framing that generative pose models cannot recover unobserved geometric alternatives from unimodal execution logs is well motivated and of broader interest beyond stowing.

major comments (2)
  1. Sec. 3.5, Eqs. (9)–(10) and CovK definition (Eq. 11): Headline Cov5 results rest on the hand-designed proxy C_total < 1.0 (one std above production mean 0.63), with fixed margins m_e = 0.01 m and Wall wall margin 0.05 m. The paper correctly scopes this as geometric validity of the perceptual decision, not full task success, but does not report sensitivity of Cov5/C_avg to the threshold, margins, or reconstruction error from single-view TSDF + voxel carving. Because cost differences vs weak baselines are large, the ranking is likely robust; still, a short sensitivity table (e.g., thresholds 0.8/1.0/1.2 and margin sweeps) is needed so readers can judge how load-bearing the specific cutoff is for the 0.89 vs 0.71 claim.
  2. Sec. 3.1 and evaluation protocol: Both training voxel labels and the evaluation mesh cost depend on the same single-view reconstruction pipeline (MaskDINO + TSDF + voxel carving against deformed fabric walls). This is not pure label circularity—poses come from production and multi-modality is measured by tip spread/DBSCAN independent of the executed pose—but reconstruction bias could systematically favor methods that stay near reconstructed free space. Please quantify reconstruction fidelity (e.g., held-out multi-view or known CAD where available) or show that relative method rankings are stable under controlled mesh noise/perturbation, so the domain argument is not confounded by shared reconstruction artifacts.
minor comments (6)
  1. Table 2: VoxFMVox is listed as “Vox.→Pose FM” but Sec. 3 and the text describe Voxel FM Voxel as voxelized input and output. Align the I/O column with the prose.
  2. Abstract states voxel→voxel inference as 1.4 s; Sec. 3.4 and Table 2 report ~2.3 s / 2300 ms. Reconcile timing numbers.
  3. Typos and formatting: “never exexuted” (p. 3); missing spaces in several compound phrases (“Manymanipulation”, “free-spaceaffordanceproblem”); URL line break in abstract; inconsistent “ms” vs “ms.”
  4. Fig. 4 caption and body: green-dot occupancy visuals are useful; adding a failure/ambiguous case and a side-by-side with Voxel FM Voxel would make the multi-modality claim more readable.
  5. Sec. 3.3: N_θ = 18 at 20° about the insertion axis and τ = 0.5 are free parameters; a one-line ablation or justification would help reproducibility.
  6. Related work: CabiNet is well positioned; a sentence distinguishing free-space tool occupancy from general collision checking could be slightly sharper for readers outside warehouse robotics.

Circularity Check

1 steps flagged

No load-bearing circularity in the domain argument or multi-modality recovery; only mild residual from calibrating the Cov5 validity threshold to production C_total statistics.

specific steps
  1. fitted input called prediction [Sec. 3.5 Evaluation Protocol (and Table 3 Cov5 definition)]
    "Production executions achieve C_total = 0.63 on our test set. We use C_total < 1.0 as the validity threshold for coverage evaluation, set at one standard deviation above the production mean of 0.63. This accommodates geometrically valid predictions that differ in execution style from production while remaining within a safe operating envelope."

    The primary multi-modality metric Cov5 counts a prediction as valid only if its C_total falls below a threshold that is explicitly fitted to the production system’s own mean + 1σ on the identical cost function and test set. Absolute coverage numbers are therefore partially relative to production performance by construction; the circularity is mild because the same fixed threshold is applied uniformly to all methods, C_total is not the training objective, and the paper’s strongest evidence is the comparative ablation rather than the absolute threshold value.

full rationale

This is an empirical robotics/ML paper whose central claim (spatial masked reconstruction of blade occupancy recovers multi-modal geometrically feasible placements from unimodal production executions) is supported by held-out temporal test sets, an ablation isolating the reconstruction objective (Voxel FM Voxel vs. VulcanVoxel: identical spatial output, Cov100 0.07 o0.71), and multi-modality metrics (tip spread σ, DBSCAN modes) that do not reference the executed label. Training supervision is sparse blade-channel occupancy from production poses; evaluation deliberately rejects SE(3) agreement with production and instead scores a hand-designed mesh cost C_total. There are no uniqueness theorems, self-citation chains that force the result, ansatzes smuggled via prior author work, or self-definitional equations. The sole mild circularity is that the binary validity threshold used for Cov5 is set to production mean + 1σ on the same cost; this makes absolute coverage numbers partially relative to production by construction, but does not force the comparative ranking or the ablation contrast, and the paper scopes the claim to geometric validity of the perceptual decision rather than end-to-end task success. Score 2 is therefore proportionate; the derivation chain is otherwise self-contained against the paper’s own baselines and production reference.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The central claim rests on standard 3D vision primitives, domain assumptions about production stow data and free-space geometry, a handful of hand-chosen thresholds and loss weights, and the framing of blade affordances as spatial occupancy objects. No new physical entities are postulated; VulcanVoxel is an architecture, not an ontological invention.

free parameters (6)
  • voxel grid resolution / bounding box = 32^3, 30 cm
    32^3 over (30 cm)^3 (~9.4 mm/voxel) chosen to resolve gaps and blade cross-section; not derived.
  • Focal + Dice loss weights and hyperparameters = w_d=3.0, α=0.85, γ=2.0
    α=0.85, γ=2.0, w_d=3.0 selected by grid search on held-out validation.
  • occupancy threshold τ and pose enumeration N_θ = τ=0.5, N_θ=18
    τ=0.5 for connected components; N_θ=18 at 20° intervals for scoring.
  • validity cost threshold C_total < 1.0 = 1.0 (prod mean 0.63)
    Set one standard deviation above production mean cost 0.63; defines CovK success.
  • distillation mixing weight λ = 0.7
    λ=0.7 balances teacher multi-modal supervision vs ground-truth.
  • collision margin m_e and wall margin = 0.01 m / 0.05 m
    m_e=0.01 m; Wall Insert left/right margin raised to 0.05 m; hand-set safety envelopes.
axioms (5)
  • domain assumption Blade free-space affordances are subsets of 3D Euclidean space defined by geometric feasibility of rigid-body occupancy, complementary to surface-contact affordances.
    Stated in abstract and Sec. 1; frames the entire spatial-inference argument.
  • domain assumption Production stow logs are unimodal (one executed strategy per bin state) even when multiple geometrically valid insertions exist.
    Sec. 1 and 4.3; justifies why pose generative models collapse and why multi-modality recovery is non-trivial.
  • domain assumption Single-view RGB-D + MaskDINO + TSDF fusion + voxel carving yields meshes accurate enough for both training supervision and geometric cost evaluation.
    Sec. 3.1; load-bearing for both training labels and the C_total metric.
  • ad hoc to paper Channel-masked reconstruction of occupancy fields forces local geometric feasibility reasoning without global competition, enabling multi-modal activation.
    Sec. 3.2; the key modeling hypothesis that distinguishes MAE from flow matching on the same voxels.
  • standard math Standard 3D U-Net, Focal/Dice losses, and flow-matching baselines behave as described in the cited literature.
    Architecture and losses imported from prior CV work (He et al. MAE, Lin et al. Focal, Milletari Dice).
invented entities (2)
  • VulcanVoxel (channel-masked 3D occupancy MAE for blade affordance) independent evidence
    purpose: Architecture that reconstructs blade occupancy conditioned on scene channels to recover multi-modal free-space placements.
    Method name and design; not a physical entity. Independent evidence is the empirical Cov5 lift and ablation contrast on held-out real data.
  • StowAffordance dataset framing of free-space blade insertion cycles independent evidence
    purpose: Released real warehouse RGB-D + pose trajectories for free-space affordance research.
    Data product; independent of the model claim once released.

pith-pipeline@v1.1.0-grok45 · 16697 in / 3703 out tokens · 35301 ms · 2026-07-12T11:40:41.324825+00:00 · methodology

0 comments
read the original abstract

Many manipulation tasks require reasoning about free-space affordances: discovering volumes where an extended rigid tool can safely navigate, complementary to surface contact affordances for grasping. Robotic stowing is a canonical instance, where a blade must sweep items aside inside cluttered fabric bins to create insertion space. Production stow systems generate millions of such episodes, but standard approaches with unimodal data infer affordances as SE(3) pose distributions, a geometric question asked in the wrong domain. VulcanVoxel keeps inference spatial: a masked autoencoder over 3D occupancy fields reconstructs blade occupancy conditioned on scene geometry, computing feasibility locally at each voxel and recovering multi-modal predictions from unimodal data. Blade affordances are spatial objects, subsets of 3D space defined by geometric feasibility. Pose parameters carry no structure for reasoning whether unobserved placements are feasible, and standard generative objectives including flow matching faithfully learn the unimodal distribution produced by execution policies and cannot recover geometric alternatives. Trained on 10,000 real warehouse stow episodes without human annotation, VulcanVoxel achieves top-5 coverage of 0.89 versus 0.71 for the best pose-based baseline, with a distilled student providing RGB-to-voxel inference in 30 ms. vs. 1.4 s. for voxel-to-voxel. We have released a dataset of real blade insertion cycles with RGB-D observations and pose trajectories at https://www.armbench.com/blade_insertion. html.

Figures

Figures reproduced from arXiv: 2607.02549 by Aditya Mehrotra, Harpreet Sawhney, Kunal Mehrotra, Minju Jung, Mudit Agrawal, Tianyu Li.

Figure 1
Figure 1. Figure 1: Blade insertion for stowing. Top: Initial bin state, blade approach, and in￾sertion into a densely cluttered bin. Bottom: Blade navigating tight gaps while sweep￾ing. Precise 3D geometric reasoning avoids damaging items/bin walls while reaching sufficiently deep to start sweeping items to create space. predictions to. Blade insertion for robotic stowing is a canonical instance, studied here with real manip… view at source ↗
Figure 2
Figure 2. Figure 2: VulcanVoxel pipeline. Depth observations are voxelized into multi-channel occupancy grids, then a masked autoencoder reconstructs blade occupancy from scene context. Predicted occupancy is converted to blade pose via geometric scoring. An op￾tional RGB distillation path (bottom) replaces the explicit stages for efficient inference. after which all subsequent motion relies on force feedback. A geometrically… view at source ↗
Figure 3
Figure 3. Figure 3: VulcanVoxel architecture. Point cloud, object/bin masks, and blade geome￾try are used to construct a multi-channel 3D occupancy grid. During training, random channel masking is applied before the grid is passed through a 3D UNet autoencoder that reconstructs the masked channels. matching denoises toward the training distribution, which is unimodal in oc￾cupancy space; operating in the spatial domain is nec… view at source ↗
Figure 4
Figure 4. Figure 4: VulcanVoxel Blade Prediction Visuals. Shown as green dots representing blade voxel occupancy within a rendering of the reconstructed bin-item 3D mesh. Bin walls are colored, items are in gray. Top: Bin image, Blade occupancy predictions for this image and another sample bin. Bottom: Vulcan Distill predictions for bin on the left and another sample bin. where ϕe(p) = minq∈p SDFe(q), me = 0.01 m, we = 1.0, w… view at source ↗
Figure 5
Figure 5. Figure 5: Left: Generalization to August 2025 (1,000 cycles). Method rankings preserved across the two-month temporal gap. Right: High clutter (GCU > 50%, July 2025). VulcanVoxel degrades less in coverage than competing methods; the Cov5 advantage over Voxel FM Pose grows under clutter. (†Wall Insert omits reach cost.) [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Left: Challenging configurations (production failures, August 2025). Vulcan￾Voxel maintains coverage consistent with the successful test set while all baselines degrade substantially. Results represent a generalization test on difficult bin config￾urations, not a failure correction benchmark. Right: Distillation quality, July 2025. VV Distill substantially outperforms RGB FM at comparable speed, preserving… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

24 extracted references · 4 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)

    Bahl, S., Mendonca, R., Chen, L., Jain, U., Pathak, D.: Affordances from human videos as a versatile representation for robotics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)

    2023

  2. [2]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    Cao, A.Q., de Charette, R.: MonoScene: Monocular 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    2022

  3. [3]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)

    Caron, M., Touvron, H., Misra, I., J’egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)

    2021

  4. [4]

    In: Proceedings of SIGGRAPH (1996)

    Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of SIGGRAPH (1996)

    1996

  5. [5]

    https://arxiv.org/abs/2512.12842 (2025)

    Fang, K., Chen, Y., et al.: SAGA: Open-world mobile manipulation via structured affordance grounding. https://arxiv.org/abs/2512.12842 (2025)

    arXiv 2025

  6. [6]

    In: IEEE International Conference on Robotics and Automation (ICRA) (2023)

    Fishman,A.,Murali,A.,Eppner,C.,Peele,B.,Boots,B.,Fox,D.:CabiNet:Scaling neural collision detection for object rearrangement with procedural scene gener- ation. In: IEEE International Conference on Robotics and Automation (ICRA) (2023)

    2023

  7. [7]

    arXiv preprint arXiv:2508.11002 (2025)

    Gkanatsios, N., Xu, J., Bronars, M., Mousavian, A., Ke, T.W., Fragkiadaki, K.: 3d flowmatch actor: Unified 3d policy for single-and dual-arm manipulation. arXiv preprint arXiv:2508.11002 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    He, K., Chen, X., Xie, S., Li, Y., Doll’ar, P., Girshick, R.: Masked autoencoders are scalablevisionlearners.In:ProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition (CVPR) (2022)

    2022

  9. [9]

    arXiv preprint arXiv:2505.04572 (2025)

    Hudson, N., Hooks, J., et al.: Stow: Robotic packing of items into fabric pods. arXiv preprint arXiv:2505.04572 (2025)

    Pith/arXiv arXiv 2025

  10. [10]

    In: Conference on Robot Learning (CoRL) (2024)

    Ke, T.W., Gkanatsios, N., Fragkiadaki, K.: 3D diffuser actor: Policy diffusion with 3d scene representations. In: Conference on Robot Learning (CoRL) (2024)

    2024

  11. [11]

    In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition (CVPR) (2023)

    Li, F., Zhang, H., Liu, H., Guo, J., Ni, L.M., Zhang, L.: Mask DINO: Towards a unified transformer-based framework for object detection and segmentation. In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition (CVPR) (2023)

    2023

  12. [12]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2017)

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Doll’ar, P.: Focal loss for dense object de- tection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2017)

    2017

  13. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)

    Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)

    2019

  14. [14]

    In: International Conference on 3D Vision (3DV) (2016) 16 T

    Milletari, F., Navab, N., Ahmadi, S.A.: V-Net: Fully convolutional neural networks for volumetric medical image segmentation. In: International Conference on 3D Vision (3DV) (2016) 16 T. Li et al

    2016

  15. [15]

    In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV) (2021)

    Mo, K., Guibas, L., Mukadam, M., Gupta, A., Tulsiani, S.: Where2act: From pixels to actions for articulated 3d objects. In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV) (2021)

    2021

  16. [16]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Qian, S., Mo, K., Blukis, V., Fouhey, D.F., Fox, D., Goyal, A.: 3d-mvp: 3d multi- view pretraining for manipulation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 22530–22539 (2025)

    2025

  17. [17]

    In: IEEE International Conference on Robotics and Automation (ICRA) (2022)

    Simeonov, A., Du, Y., Tagliasacchi, A., Tenenbaum, J.B., Rodriguez, A., Agrawal, P., Sitzmann, V.: Neural descriptor fields: SE(3)-equivariant object representations for manipulation. In: IEEE International Conference on Robotics and Automation (ICRA) (2022)

    2022

  18. [18]

    https://arxiv.org/abs/2506.09284 (2025)

    Tang, Y., Huang, W., et al.: UAD: Unsupervised affordance distillation for gener- alization in robotic manipulation. https://arxiv.org/abs/2506.09284 (2025)

    Pith/arXiv arXiv 2025

  19. [19]

    In: Robotics: Science and Systems (RSS) (2023)

    Turpin, D., Wang, L., Bhatt, R., Garg, A.: GIFT: Generalizable interaction-aware functional tool affordances without labels. In: Robotics: Science and Systems (RSS) (2023)

    2023

  20. [20]

    In: Conference on Robot Learning (CoRL) (2024)

    Wang, Y., Li, Z., Zheng, M., Driggs-Campbell, K., Li, F.F., Savarese, S., Li, Y.: D3Fields: Dynamic 3d descriptor fields for zero-shot generalizable rearrangement. In: Conference on Robot Learning (CoRL) (2024)

    2024

  21. [21]

    https://arxiv.org/abs/2503.04877 (2025)

    Wilcox, A., Ghanem, M., et al.: Adapt3R: Adaptive 3d scene representation for domain transfer in imitation learning. https://arxiv.org/abs/2503.04877 (2025)

    Pith/arXiv arXiv 2025

  22. [22]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2023)

    Wu, R., Chen, K., Geng, C., Dong, H.: Learning environment-aware affordance for 3d articulated object manipulation under occlusions. In: Advances in Neural Information Processing Systems (NeurIPS) (2023)

    2023

  23. [23]

    In: International Conference on Learning Representations (ICLR) (2022)

    Wu, R., Zhao, Y., Gupta, A., Huang, D.A., Fang, H.S., Li, J., Guibas, L., Dong, H.: VAT-Mart: Learning visual action trajectory proposals for manipulating 3d artic- ulated objects. In: International Conference on Learning Representations (ICLR) (2022)

    2022

  24. [24]

    In: Conference on Robot Learning (CoRL) (2020)

    Zeng, A., Florence, P., Tompson, J., Welker, S., Chien, J., Attarian, M., Arm- strong, T., Krasin, I., Duong, D., Sindhwani, V., Lee, J.: Transporter networks: Rearranging the visual world for robotic manipulation. In: Conference on Robot Learning (CoRL) (2020)

    2020