pith. sign in

arxiv: 2606.26800 · v2 · pith:QH6BLQLGnew · submitted 2026-06-25 · 💻 cs.RO

SSI-Policy: Learning Structured Scene Interfaces for Vision-Language Robotic Manipulation

Kaijun Wang , Zikai Ouyang , Xuping Wu , Jinyi Hong , Wei Pan , Haibo Lu , Jia Pan , Wei Zhang
show 1 more author
Linfang Zheng
This is my paper

Pith reviewed 2026-06-30 09:59 UTC · model grok-4.3

classification 💻 cs.RO
keywords robotic manipulationvision-language policyfew-shot learningstructured scene representationmonocular depthmotion trajectoriesLIBERO benchmark
0
0 comments X

The pith

A robot-agnostic Structured Scene Interface lets policies learn manipulation tasks from only 10 demonstrations by training on action-free video.

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

The paper presents SSI-Policy, a framework that inserts an intermediate Structured Scene Interface between perception and control to handle spatial grounding and task reasoning in low-data robotic manipulation. The interface combines monocular depth cues, language-grounded object layouts, and instruction-conditioned 2D motion trajectories into a single RGB-only representation that can be learned without action labels. By decoupling the interface from the downstream policy, the method allows the policy to train effectively on small demonstration sets while avoiding geometric drift over long horizons. On the LIBERO benchmark this yields nearly 15 percent gains over prior methods with 10 demos per task and matches the performance of 50-demo approaches that rely on large external pretraining. The same interface transfers to 13 real-world tasks involving spatial reasoning, cross-embodiment transfer, and contact-rich contact.

Core claim

The central claim is that a unified RGB-only Structured Scene Interface encoding monocular depth features, language-grounded object layouts, and instruction-conditioned 2D motion trajectories can be trained from action-free video, remains robot-agnostic, and supplies a sufficiently structured representation for a downstream policy to solve vision-language manipulation tasks from few demonstrations.

What carries the argument

The Structured Scene Interface (SSI), a modular RGB-only intermediate representation that jointly encodes monocular depth, language-grounded layouts, and conditioned motion trajectories to decouple perception from control.

If this is right

  • Geometric depth cues and motion trajectories supply complementary information inside the shared interface.
  • The robot-agnostic interface supports cross-embodiment transfer on real hardware.
  • Performance remains competitive without large-scale external pretraining.
  • The modular split allows the interface to be trained separately from the policy on video data.

Where Pith is reading between the lines

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

  • The same interface could be pre-trained on large unlabeled video corpora to further reduce demonstration needs.
  • Replacing the monocular depth branch with stereo or depth-sensor input might improve precision on contact-rich tasks.
  • The explicit layout and trajectory channels may ease debugging of spatial failures compared with opaque end-to-end models.

Load-bearing premise

The Structured Scene Interface trained on action-free video produces a representation that transfers to a policy without introducing geometric drift or control errors on downstream tasks.

What would settle it

A controlled comparison in which an SSI-trained policy exhibits measurably higher failure rates or larger spatial errors on long-horizon LIBERO tasks than an otherwise identical end-to-end policy trained on the same 10 demonstrations would falsify the transfer claim.

Figures

Figures reproduced from arXiv: 2606.26800 by Haibo Lu, Jia Pan, Jinyi Hong, Kaijun Wang, Linfang Zheng, Wei Pan, Wei Zhang, Xuping Wu, Zikai Ouyang.

Figure 1
Figure 1. Figure 1: Overview of SSI-Policy. From RGB observations and a language instruction, SSI constructs a structured intermediate in￾terface that encodes monocular geometry cues, language-grounded object layouts, and instruction-conditioned motion trajectories. SSI is robot-agnostic (cross-embodiment) and can be learned from action-free videos alone (e.g., human-hand videos or robot videos). A diffusion policy is then tr… view at source ↗
Figure 2
Figure 2. Figure 2: Framework overview. The Perception Composer con￾verts RGB images and language instructions into three structured signals: monocular depth features, task-relevant layout maps, and instruction-conditioned motion trajectories. The Diffusion Action Planner then integrates these signals—together with proprioception and optional RGB inputs—to generate multi-step action sequences. B. Framework Overview We propose… view at source ↗
Figure 4
Figure 4. Figure 4: Real-world setup and task suite. Top: time-lapse of a representative task. Middle/Bottom-left: representative scenes illustrating the diversity of the 13 real-world tasks, annotated with colored markers indicating task categories. Bottom-right: full view of the experimental platform, including the 6-DoF robotic arm and dual RGB cameras (side-view and eye-in-hand). modal fusion. The Diffusion Action Planner… view at source ↗
Figure 5
Figure 5. Figure 5: Cross-embodiment results on LIBERO-Spatial. Success rates under few-shot and zero-shot settings. shown in Table III, our method outperforms Diffusion Policy by up to 75 percentage points on directional tasks and 10– 20% on disambiguation tasks (80.0% vs. 43.3% on average). We attribute this to structured spatial grounding: depth cues help resolve directional references, while the SSI com￾bines object local… view at source ↗
Figure 6
Figure 6. Figure 6: Evaluating SSI as a policy interface. Success rates across LIBERO suites comparing the full model (SSI + RGB) with an SSI￾only variant conditioned on SSI and proprioception. variants are specialized and do not generalize across suites. The unified interface performs consistently well across suites, indicating complementary benefits from jointly encoding geometric structure and task-conditioned motion withi… view at source ↗
read the original abstract

Real-world robotic manipulation demands spatial grounding, task-aware reasoning, and precise control. Learning such capabilities becomes particularly challenging in the low-data regime. Prior methods often trade off scalable task-level reasoning and explicit physical structure: video-based approaches can drift geometrically over long horizons, 3D approaches often require depth sensing, and many flow/trajectory interfaces emphasize motion without an explicit RGB-only geometric representation. We introduce SSI-Policy, a modular framework built around a Structured Scene Interface (SSI) -- a unified, RGB-only intermediate representation that jointly encodes monocular depth features, language-grounded object layouts, and instruction-conditioned 2D motion trajectories. Critically, SSI is robot-agnostic and trainable from action-free video, decoupling perception from control so that the downstream policy can learn from few demonstrations. On the LIBERO benchmark with only 10 demonstrations per task, SSI-Policy improves over the strongest prior method by nearly 15\% and remains competitive with 50-demo methods that leverage large-scale external pretraining. Ablations show that geometric and motion cues provide complementary benefits within the shared interface. We further validate on 13 real-world tasks spanning spatial reasoning, cross-embodiment transfer, and contact-rich manipulation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces SSI-Policy, a modular framework for vision-language robotic manipulation built around a Structured Scene Interface (SSI). SSI is an RGB-only unified representation that jointly encodes monocular depth features, language-grounded object layouts, and instruction-conditioned 2D motion trajectories. It is designed to be robot-agnostic and trainable from action-free video, decoupling perception from control to enable policy learning from few demonstrations. The central empirical claim is that on the LIBERO benchmark with only 10 demonstrations per task, SSI-Policy improves over the strongest prior method by nearly 15% while remaining competitive with 50-demo methods that use large-scale external pretraining; ablations indicate complementary benefits from geometric and motion cues, with further validation on 13 real-world tasks spanning spatial reasoning, cross-embodiment transfer, and contact-rich manipulation.

Significance. If the performance gains and transfer properties hold under rigorous controls, the work could meaningfully advance low-data robotic manipulation by offering a structured, transferable RGB-only interface that avoids depth sensors and action labels during pretraining. The robot-agnostic design and explicit separation of perception and control are strengths that could support broader applicability across embodiments.

major comments (2)
  1. [Experimental Results / Setup] The provided abstract states performance numbers (nearly 15% gain on LIBERO-10) but supplies no details on training procedure, error bars, data splits, or ablation controls. This absence prevents verification of the central claim; the full manuscript must include these in the experimental section for the result to be load-bearing.
  2. [Method / Ablations] The weakest assumption—that the SSI trained from action-free video transfers without introducing geometric drift or control errors—requires explicit quantitative support (e.g., drift metrics or ablation on transfer error) in the results or method sections, as this is central to the few-demonstration claim.
minor comments (1)
  1. Clarify notation for the three SSI components (depth, layouts, trajectories) and ensure consistent use across figures and text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. We address each major point below and have revised the manuscript to strengthen the experimental reporting and provide additional quantitative support for the core assumptions.

read point-by-point responses

  1. Referee: [Experimental Results / Setup] The provided abstract states performance numbers (nearly 15% gain on LIBERO-10) but supplies no details on training procedure, error bars, data splits, or ablation controls. This absence prevents verification of the central claim; the full manuscript must include these in the experimental section for the result to be load-bearing.

    Authors: We agree that the abstract alone is insufficient for verification and that the experimental section must contain these details. The full manuscript already reports training procedures, data splits, and ablation controls in Section 4, but we have revised the experimental section to add explicit error bars across all LIBERO-10 runs, clearer descriptions of the 10-demonstration splits, and consolidated ablation tables with statistical controls. These changes make the central performance claim directly verifiable from the text. revision: yes

  2. Referee: [Method / Ablations] The weakest assumption—that the SSI trained from action-free video transfers without introducing geometric drift or control errors—requires explicit quantitative support (e.g., drift metrics or ablation on transfer error) in the results or method sections, as this is central to the few-demonstration claim.

    Authors: We acknowledge that explicit quantification of geometric drift during transfer from action-free video is important for supporting the few-demonstration claim. The current manuscript provides indirect evidence through complementary ablations on geometric and motion cues (Section 4.3) and real-world transfer results, but does not include dedicated drift metrics. We have added a new ablation subsection with quantitative transfer-error metrics (e.g., endpoint drift on held-out video sequences and policy performance degradation when SSI is frozen vs. fine-tuned) to directly address this assumption. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The abstract and context present SSI-Policy as a modular framework whose central claims rest on empirical benchmark gains (LIBERO-10) and ablations showing complementary cues, without any equations, fitted parameters renamed as predictions, or load-bearing self-citations. No derivation chain is described that reduces by construction to its inputs; results are framed as external evidence rather than self-referential definitions. The paper is therefore self-contained against the supplied material.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the SSI is introduced as a new representation without detailing its internal construction or any fitted constants.

pith-pipeline@v0.9.1-grok · 5769 in / 1145 out tokens · 32352 ms · 2026-06-30T09:59:36.898159+00:00 · methodology

Review history (2 revisions) →

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

61 extracted references · 24 canonical work pages · 14 internal anchors

  1. [1]

    Learning universal policies via text-guided video generation,

    Y . Du,et al., “Learning universal policies via text-guided video generation,” inAdvances in Neural Information Processing Systems, A. Oh,et al., Eds., vol. 36. Curran Associates, Inc., 2023, pp. 9156– 9172

    2023

  2. [2]

    Video pretraining (vpt): Learning to act by watching unlabeled online videos,

    B. Baker,et al., “Video pretraining (vpt): Learning to act by watching unlabeled online videos,”Advances in Neural Information Processing Systems, vol. 35, pp. 24 639–24 654, 2022

    2022

  3. [3]

    Dreamitate: Real-world visuomotor policy learning via video generation,

    J. Liang,et al., “Dreamitate: Real-world visuomotor policy learning via video generation,” in8th Annual Conference on Robot Learning, 2024

    2024

  4. [4]

    Depth helps: Improving pre-trained rgb-based policy with depth information injection,

    X. Pang,et al., “Depth helps: Improving pre-trained rgb-based policy with depth information injection,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 7251–7256

    2024

  5. [5]

    SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Models,

    D. Qu,et al., “SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Models,” inProceedings of Robotics: Science and Systems, LosAngeles, CA, USA, June 2025

    2025

  6. [6]

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

    M. Shridhar,et al., “Perceiver-actor: A multi-task transformer for robotic manipulation,” inProceedings of the 6th Conference on Robot Learning (CoRL), 2022

    2022

  7. [7]

    Visual robotic manipulation with depth-aware pretraining,

    W. Wang,et al., “Visual robotic manipulation with depth-aware pretraining,” 2024. [Online]. Available: https://arxiv.org/abs/2401. 09038

    2024

  8. [8]

    Any-point Trajectory Modeling for Policy Learning,

    C. Wen,et al., “Any-point Trajectory Modeling for Policy Learning,” inProceedings of Robotics: Science and Systems, Delft, Netherlands, July 2024

    2024

  9. [9]

    Flow as the cross-domain manipulation interface,

    M. Xu,et al., “Flow as the cross-domain manipulation interface,” in 8th Annual Conference on Robot Learning, 2024

    2024

  10. [10]

    General flow as foundation affordance for scalable robot learning,

    C. Yuan,et al., “General flow as foundation affordance for scalable robot learning,” in8th Annual Conference on Robot Learning, 2024

    2024

  11. [11]

    Flip: Flow- centric generative planning as general-purpose manipulation world model,

    C. Gao,et al., “Flip: Flow-centric generative planning as general- purpose manipulation world model,” 2025. [Online]. Available: https://arxiv.org/abs/2412.08261

    work page arXiv 2025

  12. [12]

    LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning

    B. Liu,et al., “Libero: Benchmarking knowledge transfer for lifelong robot learning,”arXiv preprint arXiv:2306.03310, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  13. [13]

    Imitating latent policies from observation,

    A. Edwards,et al., “Imitating latent policies from observation,” in Proceedings of the 36th International Conference on Machine Learn- ing, ser. Proceedings of Machine Learning Research, K. Chaudhuri and R. Salakhutdinov, Eds., vol. 97. PMLR, 09–15 Jun 2019, pp. 1755–1763

    2019

  14. [14]

    Learning to act without actions,

    D. Schmidt and M. Jiang, “Learning to act without actions,” inSecond Agent Learning in Open-Endedness Workshop, 2023

    2023

  15. [15]

    Genie: generative interactive environments,

    J. Bruce,et al., “Genie: generative interactive environments,” inPro- ceedings of the 41st International Conference on Machine Learning, ser. ICML’24. JMLR.org, 2024

    2024

  16. [16]

    Latent action pretraining from videos,

    S. Ye,et al., “Latent action pretraining from videos,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  17. [17]

    Learning to Act Anywhere with Task-centric Latent Ac- tions,

    Q. Bu,et al., “Learning to Act Anywhere with Task-centric Latent Ac- tions,” inProceedings of Robotics: Science and Systems, LosAngeles, CA, USA, June 2025

    2025

  18. [18]

    Generative visual foresight meets task-agnostic pose estimation in robotic table-top manipulation,

    C. Zhang,et al., “Generative visual foresight meets task-agnostic pose estimation in robotic table-top manipulation,” in9th Annual Conference on Robot Learning, 2025

    2025

  19. [19]

    Grounding video models to actions through goal conditioned exploration,

    Y . Luo and Y . Du, “Grounding video models to actions through goal conditioned exploration,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  20. [20]

    Learning to act from actionless videos through dense correspondences,

    P.-C. Ko,et al., “Learning to act from actionless videos through dense correspondences,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  21. [21]

    SKIL: Semantic Keypoint Imitation Learning for Generalizable Data-efficient Manipulation,

    S. Wang,et al., “SKIL: Semantic Keypoint Imitation Learning for Generalizable Data-efficient Manipulation,” inProceedings of Robotics: Science and Systems, LosAngeles, CA, USA, June 2025

    2025

  22. [22]

    Coarse-to-Fine Q-attention: Efficient Learning for Visual Robotic Manipulation via Discretisation,

    S. James,et al., “ Coarse-to-Fine Q-attention: Efficient Learning for Visual Robotic Manipulation via Discretisation,” in2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos, CA, USA: IEEE Computer Society, June 2022, pp. 13 729– 13 738

    2022

  23. [23]

    Lift3d: Synthesize 3d training data by lifting 2d gan to 3d generative radiance field,

    L. Li,et al., “Lift3d: Synthesize 3d training data by lifting 2d gan to 3d generative radiance field,” inProc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  24. [24]

    Frame mining: a free lunch for learning robotic manipulation from 3d point clouds,

    M. Liu,et al., “Frame mining: a free lunch for learning robotic manipulation from 3d point clouds,” 2022. [Online]. Available: https://arxiv.org/abs/2210.07442

    work page arXiv 2022

  25. [25]

    Polarnet: 3d point clouds for language- guided robotic manipulation,

    S. Chen,et al., “Polarnet: 3d point clouds for language- guided robotic manipulation,” 2023. [Online]. Available: https: //arxiv.org/abs/2309.15596

    work page arXiv 2023

  26. [26]

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

    Y . Ze,et al., “3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,” inProceedings of Robotics: Science and Systems (RSS), 2024

    2024

  27. [27]

    Rise: 3d perception makes real-world robot imitation simple and effective,

    C. Wang,et al., “Rise: 3d perception makes real-world robot imitation simple and effective,”arXiv preprint arXiv:2404.12281, 2024

    work page arXiv 2024

  28. [28]

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

    T. Gervet,et al., “Act3d: 3d feature field transformers for multi-task robotic manipulation,” 2023. [Online]. Available: https: //arxiv.org/abs/2306.17817

    work page arXiv 2023

  29. [29]

    3D-VLA: A 3D Vision-Language-Action Generative World Model

    H. Zhen,et al., “3d-vla: A 3d vision-language-action generative world model,” 2024. [Online]. Available: https://arxiv.org/abs/2403.09631

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  30. [30]

    RT-1: Robotics Transformer for Real-World Control at Scale

    A. Brohan,et al., “Rt-1: Robotics transformer for real-world control at scale,” inarXiv preprint arXiv:2212.06817, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  31. [31]

    RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control

    ——, “Rt-2: Vision-language-action models transfer web knowledge to robotic control,” inarXiv preprint arXiv:2307.15818, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  32. [32]

    Octo: An Open-Source Generalist Robot Policy,

    D. Ghosh,et al., “Octo: An Open-Source Generalist Robot Policy,” inProceedings of Robotics: Science and Systems, Delft, Netherlands, July 2024

    2024

  33. [33]

    OpenVLA: An Open-Source Vision-Language-Action Model

    M. Kim,et al., “Openvla: An open-source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  34. [34]

    BC-z: Zero-shot task generalization with robotic imitation learning,

    E. Jang,et al., “BC-z: Zero-shot task generalization with robotic imitation learning,” in5th Annual Conference on Robot Learning, 2021

    2021

  35. [35]

    A Generalist Agent

    S. Reed,et al., “A generalist agent,”arXiv preprint arXiv:2205.06175, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  36. [36]

    Open x-embodiment: Robotic learning datasets and rt-x models,

    A. O’Neill,et al., “Open x-embodiment: Robotic learning datasets and rt-x models,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 6892–6903

    2024

  37. [37]

    Instruction-following agents with multimodal transformer,

    H. Liu,et al., “Instruction-following agents with multimodal transformer,” 2023. [Online]. Available: https://arxiv.org/abs/2210. 13431

    2023

  38. [38]

    $\pi_0$: A Vision-Language-Action Flow Model for General Robot Control

    K. Black,et al., “π0: A vision-language-action flow model for general robot control, 2024,”URL https://arxiv. org/abs/2410.24164

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  39. [39]

    Palm-e: An embodied multimodal language model,

    D. Driess,et al., “Palm-e: An embodied multimodal language model,”

  40. [40]

    PaLM-E: An Embodied Multimodal Language Model

    [Online]. Available: https://arxiv.org/abs/2303.03378

    work page internal anchor Pith review Pith/arXiv arXiv

  41. [41]

    Otter: A vision-language-action model with text- aware visual feature extraction,

    H. Huang,et al., “Otter: A vision-language-action model with text- aware visual feature extraction,”arXiv preprint arXiv:2503.03734, 2025

    work page arXiv 2025

  42. [42]

    Copa: General robotic manipulation through spatial constraints of parts with foundation models,

    ——, “Copa: General robotic manipulation through spatial constraints of parts with foundation models,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 9488– 9495

    2024

  43. [43]

    ReKep: Spatio-Temporal Reasoning of Relational Keypoint Constraints for Robotic Manipulation

    W. Huang,et al., “Rekep: Spatio-temporal reasoning of rela- tional keypoint constraints for robotic manipulation,”arXiv preprint arXiv:2409.01652, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  44. [44]

    Robo-abc: Affordance generalization beyond categories via semantic correspondence for robot manipulation,

    Y . Ju,et al., “Robo-abc: Affordance generalization beyond categories via semantic correspondence for robot manipulation,” 2024. [Online]. Available: https://arxiv.org/abs/2401.07487

    work page arXiv 2024

  45. [45]

    An affordance keypoint detection network for robot manipulation,

    R. Xu,et al., “An affordance keypoint detection network for robot manipulation,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 2870–2877, 2021

    2021

  46. [46]

    Depth Anything V2

    L. Yang,et al., “Depth anything v2,”arXiv:2406.09414, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  47. [47]

    Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection

    S. Liu,et al., “Grounding dino: Marrying dino with grounded pre-training for open-set object detection,”arXiv preprint arXiv:2303.05499, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  48. [48]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi,et al., “Diffusion policy: Visuomotor policy learning via action diffusion,” inProceedings of Robotics: Science and Systems (RSS), 2023

    2023

  49. [49]

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

    J. Devlin,et al., “Bert: Pre-training of deep bidirectional transformers for language understanding,” 2019. [Online]. Available: https: //arxiv.org/abs/1810.04805

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  50. [50]

    R3m: A universal visual representation for robot manipulation,

    S. Nair,et al., “R3m: A universal visual representation for robot manipulation,” inConference on Robot Learning. PMLR, 2023, pp. 892–909

    2023

  51. [51]

    Mail: Improving imitation learning with mamba,

    X. Jia,et al., “Mail: Improving imitation learning with mamba,”

  52. [52]

    Available: https://arxiv.org/abs/2406.08234

    [Online]. Available: https://arxiv.org/abs/2406.08234

    work page arXiv

  53. [53]

    Vla-cache: Towards efficient vision-language-action model via adaptive token caching in robotic manipulation,

    S. Xu,et al., “Vla-cache: Towards efficient vision-language-action model via adaptive token caching in robotic manipulation,”arXiv preprint arXiv:2502.02175, 2025

    work page arXiv 2025

  54. [54]

    Diffusion transformer policy,

    Z. Hou,et al., “Diffusion transformer policy,”arXiv preprint arXiv:2410.15959, 2024

    work page arXiv 2024

  55. [55]

    TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies

    R. Zheng,et al., “Tracevla: Visual trace prompting enhances spatial- temporal awareness for generalist robotic policies,”arXiv preprint arXiv:2412.10345, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  56. [56]

    Grape: Generalizing robot policy via preference alignment,

    Z. Zhang,et al., “Grape: Generalizing robot policy via preference alignment,” 2025. [Online]. Available: https://arxiv.org/abs/2411. 19309

    2025

  57. [57]

    Bridgedata v2: A dataset for robot learning at scale,

    H. Walke,et al., “Bridgedata v2: A dataset for robot learning at scale,” inConference on Robot Learning (CoRL), 2023

    2023

  58. [58]

    Multimodal diffusion transformer: Learning versatile behavior from multimodal goals,

    M. Reuss,et al., “Multimodal diffusion transformer: Learning versatile behavior from multimodal goals,” inRobotics: Science and Systems, 2024

    2024

  59. [59]

    Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success

    M. J. Kim,et al., “Fine-tuning vision-language-action models: Opti- mizing speed and success,”arXiv preprint arXiv:2502.19645, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  60. [60]

    FAST: Efficient Action Tokenization for Vision- Language-Action Models,

    K. Pertsch,et al., “FAST: Efficient Action Tokenization for Vision- Language-Action Models,” inProceedings of Robotics: Science and Systems, LosAngeles, CA, USA, June 2025

    2025

  61. [61]

    CLIP-RT: Learning Language-Conditioned Robotic Policies from Natural Language Supervision,

    G.-C. Kang,et al., “CLIP-RT: Learning Language-Conditioned Robotic Policies from Natural Language Supervision,” inProceedings of Robotics: Science and Systems, LosAngeles, CA, USA, June 2025

    2025