Pith. sign in

REVIEW 2 major objections 5 minor 42 references

Lifting RGB images into a 3D volumetric representation closes the 2D–3D gap that weakens robot imitation policies and raises success rates by double digits.

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

2026-07-13 22:59 UTC pith:JBBGWIQ3

load-bearing objection Solid engineering paper: RGB-only volumetric lift + learned spatial tokens + multi-token DiT decoder delivers large, consistent gains on LIBERO/ManiSkill/real robot; novelty is the combination, not the pieces. the 2 major comments →

arxiv 2603.17720 v2 pith:JBBGWIQ3 submitted 2026-03-18 cs.RO

VolumeDP: Modeling Volumetric Representation for Manipulation Policy Learning

Tianxing Zhou , Feiyang Xue , Zhangchen Ye , Tianyuan Yuan , Hang Zhao , Tao Jiang This is my paper
classification cs.RO
keywords imitation learningdiffusion policyvolumetric representationrobotic manipulationspatial tokensRGB-only 3D reasoningLIBERO
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.

Robot imitation policies that watch ordinary cameras still have to output 3-D actions, so they are forced to reason across a dimensional mismatch. VolumeDP removes that mismatch by first lifting 2-D image features into an explicit volumetric grid, then selecting only the task-relevant voxels and feeding the resulting compact spatial tokens into a multi-token diffusion decoder. On the standard LIBERO suite the method reaches 88.8 % average success—14.8 points above the strongest prior baseline—and records similar gains on ManiSkill, LIBERO-Plus, and real-robot tasks that vary layout, camera angle and background. The practical claim is that ordinary RGB cameras already contain enough geometry for strong spatial reasoning once the geometry is made explicit rather than left implicit in 2-D feature maps.

Core claim

By constructing a volumetric feature volume from RGB alone via deformable volume–image cross-attention, then distilling that volume into a small set of task-conditioned spatial tokens that a multi-token decoder can condition on without collapsing them, a diffusion policy can recover the spatial alignment that pure 2-D pipelines lose, producing large and consistent gains in success rate and out-of-distribution robustness.

What carries the argument

Volumetric Representation built by Volume–Image Cross-Attention: each voxel query samples a few learnable image locations around its projected camera coordinate, yielding a dense 3-D feature field that subsequent Spatial Token Generation and Multi-Token Decoder stages can exploit.

Load-bearing premise

The method needs a fixed, task-specific axis-aligned workspace box and known camera calibration so that every voxel can be correctly projected into the image plane; if either is wrong the lifted features become misaligned.

What would settle it

Retrain and re-evaluate the identical architecture on LIBERO-Spatial after deliberately shifting the volumetric bounds by 10–20 cm or injecting a few degrees of camera extrinsic noise; a collapse of the reported 90.7 % success rate would falsify the claim that the volumetric lift is the source of the spatial gains.

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

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 / 5 minor

Summary. VolumeDP proposes an RGB-only imitation policy that restores 2D–3D spatial alignment for robotic manipulation. Image features are lifted into a fixed volumetric grid via deformable Volume–Image Cross-Attention (Sec. III-A, Eqs. 1–3), task-relevant voxels are distilled into a compact set of spatial tokens by a goal-aware TokenLearner-style module with proprioceptive auxiliary supervision (Sec. III-B), and a multi-token DiT-style diffusion decoder conditions on the full token set via adaLN-Zero (Sec. III-C). On LIBERO the method reports 88.8% average success (+14.8 pp over the strongest re-implemented baseline), with large gains on ManiSkill (+19.3 pp) and LIBERO-Plus OOD suites (+21.3 pp); real-robot trials on four tasks further show higher in-distribution success and improved robustness to layout, viewpoint and background shifts. Ablations (Table IV) attribute gains to each of the three modules.

Significance. If the reported gains hold under independent re-implementation, the work supplies a practical, depth-sensor-free route to spatially grounded visuomotor policies that measurably improves both in-distribution precision and OOD robustness. The combination of volumetric lifting, learned spatial tokenization, and multi-token diffusion decoding is a clear architectural contribution relative to pure 2D diffusion policies and to prior 3D methods that require point clouds or depth. Systematic ablations, multi-seed evaluation, and real-robot OOD tests strengthen the empirical case. The approach is immediately usable in settings where only calibrated RGB cameras are available, and the code/video release further aids reproducibility.

major comments (2)
  1. [Sec. III-A, Implementation Details] Sec. III-A and Implementation Details: the volumetric bounds r are stated to be “task-specific and tunable” and are discretized once per suite/task into a 40³ grid. The manuscript never reports how sensitive final success rates are to modest expansions/contractions or translations of these bounds, nor whether a single fixed workspace volume can be shared across the four LIBERO suites. Because the entire spatial-alignment claim rests on correct projection of every voxel (Eqs. 1–2), a short sensitivity study (or an explicit statement that bounds were chosen once from demonstration statistics and never retuned) is needed to establish that the gains are not an artifact of oracle workspace knowledge.
  2. [Tables I–III] Tables I–III: the strongest baselines (Diffusion Policy and DiT-Block Policy) are marked “*” as author re-implementations. The paper does not supply the hyper-parameter search budget, training schedule, or verification that the re-implemented DiT-Block Policy recovers the numbers originally reported by Dasari et al. Without that verification, the absolute 14.8 pp / 19.3 pp / 21.3 pp margins cannot be fully trusted as comparisons against the published state of the art. A short appendix table comparing re-implemented versus published baseline numbers (where available) would close this gap.
minor comments (5)
  1. [Abstract, Introduction] Abstract and Introduction claim “state-of-the-art” and “outperforming the strongest baseline by 14.8%”; the tables correctly report percentage-point differences. Align the wording so that “14.8 percentage points” is used consistently.
  2. [Fig. 2] Fig. 2 is rendered as a dense block of placeholder characters in the supplied manuscript; the architecture diagram is therefore unreadable. Ensure the final PDF contains a clear vector figure.
  3. [Sec. III-C, Eq. (4)] Eq. (4) re-uses the symbol α for the DDPM noise-schedule coefficient; the same symbol is commonly reserved for the cumulative product of (1−β). A brief clarification or alternative notation would avoid confusion with the standard DDPM literature.
  4. [Sec. V-B] The real-world OOD camera-view experiment rotates the camera by only ±5°. While useful, the text should note that this is a modest perturbation relative to the larger viewpoint shifts that can occur in deployment.
  5. [Sec. II-B] Related Work cites DP3, PerAct, GNFactor and RVT but does not discuss more recent RGB-only 3D lifting or occupancy-style policies; a sentence situating VolumeDP relative to those lines would improve completeness.

Circularity Check

0 steps flagged

No circularity: empirical success rates on held-out simulation and real-robot trials do not reduce to fitted inputs or self-definitional claims.

full rationale

VolumeDP's central claims are measured success rates (88.8% average on LIBERO suites, +14.8 over DiT-Block Policy; +19.3 on ManiSkill; +21.3 on LIBERO-Plus OOD; real-world ID/OOD gains) obtained by training a diffusion policy on expert demonstrations and evaluating on independent held-out episodes and physical trials. The architecture (Volume-Image Cross-Attention lifting via known camera K/P and fixed task-specific bounds r into a 40^3 grid, TokenLearner-style spatial tokens with optional proprioceptive BCE auxiliary, multi-token DiT-style decoder) is a design choice whose free parameters (r, N=200, voxel size ~1 cm) are stated as suite/task hyperparameters chosen once, not fitted to the reported success numbers. Ablations replace modules and re-measure SR; no equation equates a claimed prediction to a training loss term or to a self-cited uniqueness result. Self-citations (e.g., Occ3D/CVT-Occ occupancy work by overlapping authors) appear only as related 3D representation background and are not load-bearing for the performance claims. The derivation chain is therefore ordinary empirical ML evaluation, fully self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 3 invented entities

The central performance claim rests on a small set of architectural hyper-parameters (volume bounds, voxel size, token count) that are chosen by hand per task suite, on the domain assumption that calibrated RGB cameras and proprioception are available, and on three invented modules whose only evidence is the empirical gains they produce inside this paper.

free parameters (4)
  • volumetric bounds r = [(xmin,ymin,zmin),(xmax,ymax,zmax)]
    Task- or suite-specific axis-aligned box that defines the 3D grid; chosen by hand to cover the workspace (Sec. III-A). Directly determines which geometry can be represented.
  • voxel resolution / size s (40×40×40, ~1 cm³)
    Discretization of the volume; fixed once for all experiments but still a free design choice that trades accuracy against compute.
  • number of spatial tokens N=200
    Output size of the Spatial Token Generation module; selected for efficiency and kept constant across ablations.
  • feature dimension C=32 (volume) / 512 (tokens)
    Channel widths of the volumetric features and the projected tokens; architectural free parameters.
axioms (4)
  • domain assumption Camera intrinsic K and extrinsic P matrices are known and accurate enough for voxel-to-pixel projection (Eqs. 1-2).
    Required for Volume-Image Cross-Attention; never learned or estimated inside the pipeline.
  • domain assumption Proprioceptive end-effector pose and gripper state are available at every timestep and can be used both as policy input and as auxiliary supervision for the token weights.
    Stated in Sec. III-B; common in lab settings but not free in every real deployment.
  • domain assumption A fixed third-person (and optional wrist) RGB camera at 256×256 resolution supplies sufficient visual information for the lift.
    Experimental setup in Sec. IV-A; the method is not claimed to work from arbitrary or moving cameras without recalibration.
  • ad hoc to paper Deformable attention with a small set of learned offsets is an adequate approximation to full volume-image attention.
    Adopted from DETR-style literature (Sec. III-A) without theoretical guarantee that the sampled points recover the necessary 3D geometry.
invented entities (3)
  • Volumetric Representation obtained by Volume-Image Cross-Attention no independent evidence
    purpose: Lift 2D image features into a dense 3D feature grid that restores spatial alignment.
    Core architectural novelty; no independent physical measurement or external dataset validates the lifted features outside the policy success rates.
  • Spatial Token Generation module with goal-aware weights and proprioceptive auxiliary loss no independent evidence
    purpose: Compress the volume into a compact set of task-relevant tokens while discarding empty space.
    Inspired by TokenLearner but newly supervised by end-effector and gripper-change masks; evidence is only the ablation drop when removed.
  • Multi-Token Decoder that conditions every DiT block on the full set of spatial tokens via adaLN-Zero no independent evidence
    purpose: Avoid lossy aggregation of spatial tokens into a single descriptor before action prediction.
    Adaptation of existing DiT-Block Policy; the claim that multi-token conditioning is essential is supported only by the paper’s own ablations.

pith-pipeline@v1.1.0-grok45 · 17887 in / 3083 out tokens · 30538 ms · 2026-07-13T22:59:49.340260+00:00 · methodology

0 comments
read the original abstract

Imitation learning is a prominent paradigm for robotic manipulation. However, existing visual imitation methods map 2D image observations directly to 3D action outputs, imposing a 2D-3D mismatch that hinders spatial reasoning and degrades robustness. We present VolumeDP, a policy architecture that restores spatial alignment by explicitly reasoning in 3D. VolumeDP first lifts image features into a Volumetric Representation via cross-attention. It then selects task-relevant voxels with a learnable module and converts them into a compact set of spatial tokens, markedly reducing computation while preserving action-critical geometry. Finally, a multi-token decoder conditions on the entire token set to predict actions, thereby avoiding lossy aggregation that collapses multiple spatial tokens into a single descriptor. VolumeDP achieves a state-of-the-art average success rate of 88.8% on the LIBERO simulation benchmark, outperforming the strongest baseline by a substantial 14.8% improvement. It also delivers large performance gains over prior methods on the ManiSkill and LIBERO-Plus benchmarks. Real-world experiments further demonstrate higher success rates and robust generalization to novel spatial layouts, camera viewpoints, and environment backgrounds. Code and videos are available on the project page: https://yzc0731.github.io/VolumeDP/

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

42 extracted references · 16 linked inside Pith

  1. [1]

    Implicit behavioral cloning,

    P. Florence, C. Lynch, A. Zeng, O. A. Ramirez, A. Wahid, L. Downs, A. Wong, J. Lee, I. Mordatch, and J. Tompson, “Implicit behavioral cloning,” inConference on robot learning. PMLR, 2022, pp. 158– 168

    2022

  2. [2]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,”The International Journal of Robotics Research, vol. 44, no. 10-11, pp. 1684–1704, 2025

    2025

  3. [3]

    3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu, “3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,”arXiv preprint arXiv:2403.03954, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Act3d: 3d feature field transformers for multi-task robotic manipulation,

    T. Gervet, Z. Xian, N. Gkanatsios, and K. Fragkiadaki, “Act3d: 3d feature field transformers for multi-task robotic manipulation,”arXiv preprint arXiv:2306.17817, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    Learning human-to-robot handovers from point clouds,

    S. Christen, W. Yang, C. P ´erez-D’Arpino, O. Hilliges, D. Fox, and Y .-W. Chao, “Learning human-to-robot handovers from point clouds,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9654–9664

    2023

  6. [6]

    The ingredients for robotic diffusion transformers,

    S. Dasari, O. Mees, S. Zhao, M. K. Srirama, and S. Levine, “The ingredients for robotic diffusion transformers,” in2025 IEEE Interna- tional Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 15 617–15 625

    2025

  7. [7]

    Libero: Benchmarking knowledge transfer for lifelong robot learn- ing,

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone, “Libero: Benchmarking knowledge transfer for lifelong robot learn- ing,”Advances in Neural Information Processing Systems, vol. 36, pp. 44 776–44 791, 2023

    2023

  8. [8]

    Using proba- bilistic movement primitives in robotics,

    A. Paraschos, C. Daniel, J. Peters, and G. Neumann, “Using proba- bilistic movement primitives in robotics,”Autonomous Robots, vol. 42, no. 3, pp. 529–551, 2018

    2018

  9. [9]

    Efficient reductions for imitation learning,

    S. Ross and D. Bagnell, “Efficient reductions for imitation learning,” inProceedings of the thirteenth international conference on artificial intelligence and statistics. JMLR Workshop and Conference Pro- ceedings, 2010, pp. 661–668

    2010

  10. [10]

    A reduction of imitation learning and structured prediction to no-regret online learning,

    S. Ross, G. Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” inProceedings of the fourteenth international conference on artificial intelligence and statistics. JMLR Workshop and Conference Proceedings, 2011, pp. 627–635

    2011

  11. [11]

    Domain generalization: A survey,

    K. Zhou, Z. Liu, Y . Qiao, T. Xiang, and C. C. Loy, “Domain generalization: A survey,”IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 4, pp. 4396–4415, 2022

    2022

  12. [12]

    What matters in learning from offline human demonstrations for robot manipula- tion,

    A. Mandlekar, D. Xu, J. Wong, S. Nasiriany, C. Wang, R. Kulkarni, L. Fei-Fei, S. Savarese, Y . Zhu, and R. Mart´ın-Mart´ın, “What matters in learning from offline human demonstrations for robot manipula- tion,”arXiv preprint arXiv:2108.03298, 2021

    Pith/arXiv arXiv 2021

  13. [13]

    Behavior transformers: Cloningkmodes with one stone,

    N. M. Shafiullah, Z. Cui, A. A. Altanzaya, and L. Pinto, “Behavior transformers: Cloningkmodes with one stone,”Advances in neural information processing systems, vol. 35, pp. 22 955–22 968, 2022

    2022

  14. [14]

    Spatial action maps for mobile manipulation,

    J. Wu, X. Sun, A. Zeng, S. Song, J. Lee, S. Rusinkiewicz, and T. Funkhouser, “Spatial action maps for mobile manipulation,” in16th Robotics: Science and Systems, RSS 2020. MIT Press Journals, 2020

    2020

  15. [15]

    Diffusion model-augmented behavioral cloning,

    S.-F. Chen, H.-C. Wang, M.-H. Hsu, C.-M. Lai, and S.-H. Sun, “Diffusion model-augmented behavioral cloning,” inInternational Conference on Machine Learning. PMLR, 2024, pp. 7486–7510

    2024

  16. [16]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,”Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

    2020

  17. [17]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,”arXiv preprint arXiv:2011.13456, 2020

    Pith/arXiv arXiv 2011

  18. [18]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,”arXiv preprint arXiv:2010.02502, 2020

    Pith/arXiv arXiv 2010

  19. [19]

    Crossway diffu- sion: Improving diffusion-based visuomotor policy via self-supervised learning,

    X. Li, V . Belagali, J. Shang, and M. S. Ryoo, “Crossway diffu- sion: Improving diffusion-based visuomotor policy via self-supervised learning,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 16 841–16 849

    2024

  20. [20]

    Octo: An open-source generalist robot policy,

    O. Mees, D. Ghosh, K. Pertsch, K. Black, H. R. Walke, S. Dasari, J. Hejna, T. Kreiman, C. Xu, J. Luoet al., “Octo: An open-source generalist robot policy,” inFirst Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024, 2024

    2024

  21. [21]

    Aloha unleashed: A simple recipe for robot dexterity,

    T. Z. Zhao, J. Tompson, D. Driess, P. Florence, S. K. S. Ghasemipour, C. Finn, and A. Wahid, “Aloha unleashed: A simple recipe for robot dexterity,” inConference on Robot Learning. PMLR, 2025, pp. 1910– 1924

    2025

  22. [22]

    Open- vla: An open-source vision-language-action model,

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

    Pith/arXiv arXiv 2024

  23. [23]

    Perceiver-actor: A multi- task transformer for robotic manipulation,

    M. Shridhar, L. Manuelli, and D. Fox, “Perceiver-actor: A multi- task transformer for robotic manipulation,” inConference on Robot Learning. PMLR, 2023, pp. 785–799

    2023

  24. [24]

    Gnfactor: Multi-task real robot learning with generalizable neural feature fields,

    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,” inConference on robot learning. PMLR, 2023, pp. 284–301

    2023

  25. [25]

    Rvt: Robotic view transformer for 3d object manipulation,

    A. Goyal, J. Xu, Y . Guo, V . Blukis, Y .-W. Chao, and D. Fox, “Rvt: Robotic view transformer for 3d object manipulation,” inConference on Robot Learning. PMLR, 2023, pp. 694–710

    2023

  26. [26]

    Deformable detr: Deformable transformers for end-to-end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable detr: Deformable transformers for end-to-end object detection,”arXiv preprint arXiv:2010.04159, 2020

    Pith/arXiv arXiv 2010

  27. [27]

    Bev- former: learning bird’s-eye-view representation from lidar-camera via spatiotemporal transformers,

    Z. Li, W. Wang, H. Li, E. Xie, C. Sima, T. Lu, Q. Yu, and J. Dai, “Bev- former: learning bird’s-eye-view representation from lidar-camera via spatiotemporal transformers,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 3, pp. 2020–2036, 2024

    2020

  28. [28]

    Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving,

    X. Tian, T. Jiang, L. Yun, Y . Mao, H. Yang, Y . Wang, Y . Wang, and H. Zhao, “Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving,”Advances in Neural Information Processing Systems, vol. 36, pp. 64 318–64 330, 2023

    2023

  29. [29]

    Cvt-occ: Cost volume temporal fusion for 3d occupancy prediction,

    Z. Ye, T. Jiang, C. Xu, Y . Li, and H. Zhao, “Cvt-occ: Cost volume temporal fusion for 3d occupancy prediction,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 381–397

    2024

  30. [30]

    Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,

    V . Sanh, L. Debut, J. Chaumond, and T. Wolf, “Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,”arXiv preprint arXiv:1910.01108, 2019

    Pith/arXiv arXiv 1910

  31. [31]

    Tokenlearner: What can 8 learned tokens do for images and videos?

    M. S. Ryoo, A. Piergiovanni, A. Arnab, M. Dehghani, and A. An- gelova, “Tokenlearner: What can 8 learned tokens do for images and videos?”arXiv preprint arXiv:2106.11297, 2021

    Pith/arXiv arXiv 2021

  32. [32]

    Rethinking latent redundancy in behavior cloning: An information bottleneck approach for robot manipulation,

    S. Bai, W. Zhou, P. Ding, W. Zhao, D. Wang, and B. Chen, “Rethinking latent redundancy in behavior cloning: An information bottleneck approach for robot manipulation,”arXiv preprint arXiv:2502.02853, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    Maniskill3: Gpu parallelized robot simulation and rendering for generalizable embodied ai,

    S. Tao, F. Xiang, A. Shukla, Y . Qin, X. Hinrichsen, X. Yuan, C. Bao, X. Lin, Y . Liu, T.-K. Chanet al., “Maniskill3: Gpu parallelized robot simulation and rendering for generalizable embodied ai,” in 7th Robot Learning Workshop: Towards Robots with Human-Level Abilities, 2025

    2025

  34. [34]

    Libero-plus: In-depth robustness analysis of vision-language-action models,

    S. Fei, S. Wang, J. Shi, Z. Dai, J. Cai, P. Qian, L. Ji, X. He, S. Zhang, Z. Fei, J. Fu, J. Gong, and X. Qiu, “Libero-plus: In-depth robustness analysis of vision-language-action models,”arXiv preprint arXiv:2510.13626, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Rdt-1b: a diffusion foundation model for bimanual manipulation,

    S. Liu, L. Wu, B. Li, H. Tan, H. Chen, Z. Wang, K. Xu, H. Su, and J. Zhu, “Rdt-1b: a diffusion foundation model for bimanual manipulation,”arXiv preprint arXiv:2410.07864, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Multimodal diffusion transformer: Learning versatile behavior from multimodal goals,

    M. Reuss, ¨O. E. Ya˘gmurlu, F. Wenzel, and R. Lioutikov, “Multimodal diffusion transformer: Learning versatile behavior from multimodal goals,”arXiv preprint arXiv:2407.05996, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Any-point trajectory modeling for policy learning,

    C. Wen, X. Lin, J. So, K. Chen, Q. Dou, Y . Gao, and P. Abbeel, “Any-point trajectory modeling for policy learning,”arXiv preprint arXiv:2401.00025, 2023

    Pith/arXiv arXiv 2023

  38. [38]

    Translating flow to policy via hindsight online imitation,

    Y . Zheng, Z. Ye, W. Dong, S. Wang, Y . Liu, C. Zhang, C. Wen, and Y . Gao, “Translating flow to policy via hindsight online imitation,” in The Fourteenth International Conference on Learning Representations, 2026

    2026

  39. [39]

    Droid: A large-scale in-the-wild robot manipulation dataset,

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

    Pith/arXiv arXiv 2024

  40. [40]

    Vilt: Vision-and-language transformer without convolution or region supervision,

    W. Kim, B. Son, and I. Kim, “Vilt: Vision-and-language transformer without convolution or region supervision,” inInternational conference on machine learning. PMLR, 2021, pp. 5583–5594

    2021

  41. [41]

    Galaxea open-world dataset and g0 dual-system vla model,

    T. Jiang, T. Yuan, Y . Liu, C. Lu, J. Cui, X. Liu, S. Cheng, J. Gao, H. Xu, and H. Zhao, “Galaxea open-world dataset and g0 dual-system vla model,”arXiv preprint arXiv:2509.00576, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    Gello: A general, low- cost, and intuitive teleoperation framework for robot manipulators,

    P. Wu, Y . Shentu, Z. Yi, X. Lin, and P. Abbeel, “Gello: A general, low- cost, and intuitive teleoperation framework for robot manipulators,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 12 156–12 163

    2024