Pith. sign in

REVIEW 4 major objections 5 minor 37 references

A panorama-aware VLA policy, PanoVLA, achieves 91.3% average stage completion and 73.4% end-to-end success on real-world mobile manipulation, far above local-view baselines.

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

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

T0 review · deepseek-v4-flash

2026-08-04 10:28 UTC pith:GFVOE6EC

load-bearing objection Solid, useful paper with a plausible central claim; the headline numbers are not statistically secured because of expert-size selection and a capacity confound, but the core conclusion about panoramic context survives the criticism. the 4 major comments →

arxiv 2608.02257 v1 pith:GFVOE6EC submitted 2026-08-03 cs.RO

Learning Panorama-Aware VLA for Mobile Manipulation with Whole-Body Teleoperation

Donglin Yang , Haoran Chen , Xingyu Chen , Lixing Liu , Manyi Li , Changhe Tu , Ke Xu , Xiaojian Ma
show 1 more author
Si Liu
This is my paper
classification cs.RO
keywords vision-language-action modelspanoramic perceptionmobile manipulationwhole-body teleoperationMixture-of-Transformersimitation learningclosed-loop control360-degree imaging
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that the limited field of view of local cameras is a key bottleneck for mobile-manipulation vision-language-action (VLA) policies, and that panoramic observations—when encoded with a panorama-specialized encoder and fused through a dedicated expert—supply the global spatial context closed-loop control needs. To test this, the authors build a whole-body teleoperation system that maps sparse human motion to coordinated base-plus-dual-arm commands, collect 800 demonstrations (5.5 hours) across four real-world tasks, and train PanoVLA, a three-expert Mixture-of-Transformers policy. On real-robot trials, PanoVLA reaches 91.3% average stage completion and 73.4% end-to-end success, versus 58.6% and 30.0% for the local-view baseline; ablations show the gains come from both a panorama-geometry-aware encoder and the fusion expert, not from simply adding a wide image. A sympathetic reader would take this as evidence that robot-centric 360-degree context is an effective, cheap complement to local views for long-horizon mobile manipulation.

Core claim

The central discovery is that panoramic context improves mobile manipulation when the architecture respects the panorama's geometry. PanoVLA inserts a panorama expert between the VLM and action experts, using the transformer key–value cache as the fusion interface: the VLM expert produces a cache from local views, language, and proprioceptive state; the panorama expert runs joint self-attention over tokens extracted from equirectangular dual-fisheye images (via a dense scene-parsing encoder) together with that VLM cache; the action expert then predicts flow-matching action chunks from the combined cache. On four tasks—Move Pen, Move Block, Open Curtain, Wipe Table—with 15 closed-loop trials

What carries the argument

The load-bearing mechanism is the panorama-expert fusion path operating through the transformer key–value cache. An equirectangular projection from dual fisheye cameras is encoded by a panorama-specialized dense scene-parsing encoder (semantic and depth features concatenated), pooled into compact tokens, and passed to a 100M-parameter panorama expert that attends jointly to the VLM expert's cached context; the resulting panorama cache is concatenated with the VLM cache for the action expert. This lets panoramic tokens enter the policy in a task-conditioned way rather than as extra vision-encoder image inputs, and the ablation isolating this design (specialized encoder vs. the base VLM's pers

Load-bearing premise

The reported gains assume that the 15 closed-loop trials per task are representative and that the comparisons isolate the panoramic representation—but the paper does not state whether the evaluation trials were disjoint from the 200 demonstrations used for fine-tuning, and the panorama expert adds 100M parameters on top of the local-view baseline.

What would settle it

A concrete falsifier: run the same four tasks with an additional baseline that augments the local-view policy with a 100M-parameter expert processing an ordinary perspective image instead of a panorama. If that baseline matches PanoVLA's 73.4% success rate, the paper's central claim—that panorama-specific encoding drives the improvement—would be refuted. A second check: repeat Wipe Table with the cloth's initial bearing changed across trials; if PanoVLA's advantage disappears, the effect is overfit to the training layouts.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the claim holds, VLA policies for mobile robots should include a 360-degree observation modality; local-view-only pipelines leave a large success margin on the table.
  • The panorama expert's 100M-parameter sweet spot suggests that only modest extra capacity is needed to fuse global context; scaling beyond it does not help with 200 demonstrations per task.
  • The Wipe Table stage analysis implies that panorama fusion helps most at stages requiring relocalization after movement, where raw panorama input collapses (26.7% stage completion vs. 73.3% for PanoVLA).
  • The teleoperation system's motion-retargeting formulation, which models the base as a planar virtual floating base, makes whole-body demonstrations collectible without a separate joystick, enabling the dataset that the policy learns from.

Where Pith is reading between the lines

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

  • A testable extension is to fine-tune the frozen panorama encoder instead of keeping it fixed; the paper leaves unknown whether adaptation would further close the gap on unseen layouts.
  • Because no confidence intervals are reported, the 73.4%-vs-30.0% gap is suggestive but unquantified across task randomization; replication with more trials per task and held-out scene configurations would tighten the claim.
  • The KV-cache fusion pattern suggests a general recipe for injecting any scene-wide modality (top-down depth, egocentric video, or a 3D scene graph) into a VLA action expert, beyond panoramas specifically.
  • The task taxonomy hints that difficulty is driven by multi-stage spatial re-localization rather than mere field of view; if so, tasks like Wipe Table and Open Curtain should be where other VLA methods fail first, which matches the reported ordering.

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

4 major / 5 minor

Summary. The paper introduces PanoVLA, a panorama-aware vision-language-action policy for mobile bimanual manipulation, together with a whole-body teleoperation system for collecting demonstrations. The system uses sparse VR tracking and GMR retargeting to control a wheeled bimanual robot; the dataset comprises 800 teleoperated trajectories (about 5.5 hours) with local and panoramic observations. PanoVLA augments a π0.5 backbone with a dedicated panorama encoder (MTPano) and a 100M-parameter panorama expert that fuses panoramic context with VLM features before action generation. In closed-loop experiments on four real-world tasks, the authors report an average stage completion rate of 91.3% and end-to-end success rate of 73.4%, compared with 58.6% and 30.0% for the local-view π0.5 baseline. The paper also reports ablations on expert size and encoder choice (Move Pen only). The central claim is that explicitly modeling panoramic geometry and robot-centric global context yields substantial gains in mobile manipulation.

Significance. If the reported results hold, the paper makes a useful empirical contribution: it demonstrates that a panorama-specialized branch can improve VLA-based mobile manipulation, and it provides a practical teleoperation pipeline and dataset. The controlled baselines (raw panorama, stacked perspective views) and ablations are methodologically helpful, and the consistent improvement over the local-view baseline across all four tasks is encouraging. However, the evaluation lacks statistical rigor in ways that affect the trustworthiness of the headline numbers: no confidence intervals or significance tests are reported, the panorama-expert size is selected on one of the evaluated tasks and that task is then included in the final average, and the main comparison does not hold trainable capacity fixed. These issues are addressable within the scope of the paper, but the quantitative claims need to be substantiated more carefully.

major comments (4)
  1. [Sec. V-B, Table II] The paper reports success rates and stage completion rates from 15 trials per task with no confidence intervals, standard errors, or significance tests. For binary success metrics with n=15, the standard error around a 50% proportion is roughly 13 percentage points; even around 30% it is about 12 points. Several reported differences are within this noise, e.g., PanoVLA and π0.5 w/ Pano both achieve 86.7% SR on Move Pen, and the 73.3% vs 46.7% difference on Wipe Table is based on 15 trials. The aggregate averages across 60 trials are more stable, but the per-task comparisons and the claim that PanoVLA 'stabilizes closed-loop policy execution' are not statistically supported as reported. Please include confidence intervals (e.g., Clopper-Pearson), treat the per-task results as primary, and avoid over-interpreting small differences.
  2. [Sec. V-C, Table III vs Table II] The 100M panorama expert is selected as the best performer on Move Pen in the ablation study, and the same Move Pen results are then included in the headline aggregate in Table II. This is a form of test-set-based model selection: a hyperparameter is tuned on a task that contributes 25% of the reported average. Since the differences among 50M, 100M, 200M, and 300M experts on Move Pen are small (95.6/86.7 vs 93.3/86.7 vs 91.1/73.3), the selection is fragile, and the aggregate numbers are inflated in a way that is not acknowledged. The authors should either hold out Move Pen from the final average after selecting the expert size, select the size on a separate validation task, or report expert-size sweeps on all four tasks. At minimum, state clearly whether the selection was made on a held-out split and justify why the same data can be used for both selection and final evaluation.
  3. [Sec. V-A, Table II] The main comparison does not control for trainable capacity. PanoVLA adds a 100M-parameter panorama expert, whereas the π0.5, π0.5 w/ Pano, and π0.5 w/ Stacked Pano baselines have no additional capacity. Thus the observed gains could partly stem from having more parameters rather than from the panoramic representation. The SigLIP variant in Table III (PanoVLA w/ SigLIP 100M) controls for capacity between encoders, but it still has the panorama expert. To isolate the contribution of panoramic geometry, add a matched-capacity baseline such as π0.5 with an extra 100M expert that consumes the same local views, or a PanoVLA variant with the panorama expert but a non-panoramic global input. Without such a control, the paper's attribution of improvements to 'panoramic spatial understanding' is not fully secured.
  4. [Sec. V-C] All ablations are conducted on Move Pen only, and the paper generalizes the conclusions (e.g., 'MTPano provides more informative cues') to the other three tasks. Move Pen is the task where the raw-panorama baseline already ties PanoVLA, and the claimed advantages are largest on Wipe Table and Open Curtain, which require multi-stage state tracking. The encoder and expert-size choices should be at least sanity-checked on one of those tasks. As written, the design decisions rest on evidence from the least discriminating task, which weakens the support for the architecture's main benefit.
minor comments (5)
  1. [Sec. III-A, Eq. (1)] The GMR objective terms L_pos, L_rot, L_base, L_reg are described in words but their explicit definitions and the weights λ are not given. This makes the retargeting system difficult to reproduce. Please provide at least the functional forms or a reference to the exact formulation used.
  2. [Sec. III-B and V-A] The paper does not explicitly state that the 200 demonstration trajectories per task are disjoint from the 15 closed-loop evaluation trials per task. While this is likely the case, the text should say so directly to avoid ambiguity.
  3. [Sec. II, Related Work] Some references appear with future dates (e.g., [18], [21], [35]). If these are preprints under review, please mark them clearly; if they are intended as journal versions, ensure the citations are complete.
  4. [Figure 2] The architecture figure is dense and the flow of tokens/caches is hard to follow. A cleaner separation of the three experts, the cache concatenation, and the training/fine-tuning modules would improve readability.
  5. [Sec. V-B, paragraph 3] The sentence 'The substantial gain in SCR indicates that PanoVLA performs better across intermediate stages' conflates SCR improvements with per-stage behavior. It would be clearer to point to Figure 4 for stage-level evidence, which is only shown for Wipe Table.

Circularity Check

0 steps flagged

No significant circularity: central claims are empirical closed-loop comparisons, not derived from fitted constants or self-citations.

full rationale

The paper's central claim—that panoramic context improves mobile manipulation—is supported by closed-loop real-robot trials (Sec. V-B, Table II) comparing PanoVLA against three π0.5-based baselines. The evaluation protocol (15 trials per task) is an independent measurement, not a quantity constructed from the model's training objective or from a fitted parameter. The panorama encoder MTPano [35] is an external pretrained model with no author overlap with the present paper, and the π0.5 backbone is an external baseline, so no load-bearing result reduces to a self-citation. The expert-size ablation (Sec. V-C, Table III) does select the 100M panorama expert using Move Pen results, and the same task is included in the aggregate in Table II; this is a test-set-selection/statistical concern about the reported magnitudes, but it is not a circularity in the derivation chain: the final numbers are still observed successes and failures, not predictions forced by construction. Similarly, the lack of confidence intervals and the question of whether the 200 demonstrations per task overlap the evaluation trials are evaluation-validity concerns, not self-referential reasoning. On the specific circularity criteria (self-definition, fitted input called prediction, self-citation load-bearing, imported uniqueness, ansatz-by-citation, renaming), no step in the paper reduces to its own inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The model is trained end-to-end; besides learned weights, several design constants are fixed by experiment or convention, and the key empirical assumption is that panoramic observations carry task-relevant global context.

free parameters (4)
  • Panorama expert size = 100M parameters
    Selected from closed-loop results on Move Pen (Table III); 50M underperforms, 200M/300M do not improve. This is a hand-tuned model capacity choice, not derived from theory.
  • GMR retargeting weights = not reported
    λ_pos, λ_rot, λ_base, λ_reg in Eq. (1) must be set manually; the paper does not give values or sensitivity analysis.
  • Action chunk horizon = 32 steps
    Design choice for action chunking in Eq. (2) and training; no sweep or justification reported.
  • Proprio/action normalization percentiles = 1st and 99th
    Used to normalize 17-D state/action; conventional but hand-chosen.
axioms (4)
  • domain assumption A robot-centric 360-degree panoramic view supplies task-relevant global spatial context beyond local cameras.
    Central premise of the paper (Sec. I, IV); supported only by the empirical comparison, not by an independent argument.
  • domain assumption MTPano's pretrained semantic and depth features transfer to the robot's top-mounted panorama without domain adaptation.
    Used in panorama encoder (Eq. 3); the authors freeze MTPano, so performance depends on this transfer.
  • domain assumption The GMR retargeting yields demonstrations that are coherent enough for imitation learning.
    Teleoperation system in Sec. III is not separately evaluated; its quality is an input to all downstream results.
  • domain assumption Mixed-expert KV-cache fusion preserves the information needed for action prediction.
    Architecture choice in Sec. IV-B; no ablation of the fusion interface itself.

pith-pipeline@v1.3.0-daily-deepseek · 12013 in / 10718 out tokens · 89628 ms · 2026-08-04T10:28:28.091158+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Learning Panorama-Aware VLA for Mobile Manipulation with Whole-Body Teleoperation." pith.science (2026). https://pith.science/paper/GFVOE6EC

@misc{pith2026260802257,
  author       = {Pith},
  title        = {Pith review of: Learning Panorama-Aware VLA for Mobile Manipulation with Whole-Body Teleoperation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GFVOE6EC}},
  note         = {Machine review of arXiv:2608.02257}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Mobile manipulation is a key capability for embodied intelligence, enabling robots to accomplish complex multi-stage tasks in open-world environments. However, mobile manipulation poses two key challenges for vision-language-action (VLA) policies: At the data level, the efficient collection of high-quality whole-body demonstrations demands the coordinated control of both the mobile base and the robotic arms; at the model level, existing VLA models predominantly rely on local camera observations, whose limited field of view hinders global spatial understanding. To address these challenges, we develop a whole-body teleoperation system and a panoramic-aware VLA policy. The system enables coordinated control of a wheeled bimanual robot through a single VR interface and supports the acquisition of a real-world mobile manipulation dataset comprising 5.5 hours of multimodal demonstrations. Building upon this dataset, we propose PanoVLA, a panorama-aware vision-language-action policy for mobile bimanual manipulation. Built upon a Mixture-of-Transformers architecture, PanoVLA introduces global spatial context through dedicated panorama encoding and fusion modules, enabling effective integration of panoramic observations with language instructions and robot states for action generation. Evaluation on four real-world mobile manipulation tasks demonstrates that PanoVLA achieves an average stage completion rate of 91.3\% and an end-to-end success rate of 73.4\%, substantially outperforming local-view baselines. These results demonstrate that incorporating panoramic spatial context improves spatial understanding and closed-loop manipulation performance in mobile robots.

Figures

Figures reproduced from arXiv: 2608.02257 by Changhe Tu, Donglin Yang, Haoran Chen, Ke Xu, Lixing Liu, Manyi Li, Si Liu, Xiaojian Ma, Xingyu Chen.

Figure 1
Figure 1. Figure 1: Overview of the whole-body teleoperation pipeline. Sparse [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed PanoVLA framework, which incorporates panoramic context into a vision-language-action model for mobile manipulation. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative real-world closed-loop rollouts across the four mobile manipulation tasks. Across multiple large-workspace, multi-stage tasks, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Stage-wise performance on Wipe Table task. Bars summarize per [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] 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

37 extracted references · 16 linked inside Pith

  1. [1]

    Open x-embodiment: Robotic learning datasets and rt-x models : Open x-embodiment collaboration0,

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, Mandlekar,et al., “Open x-embodiment: Robotic learning datasets and rt-x models : Open x-embodiment collaboration0,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 6892–6903

    2024

  2. [2]

    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. Sanketi,et al., “Open- vla: An open-source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    π0.5: a vision-language-action model with open-world generaliza- tion,

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia,et al., “π0.5: a vision-language-action model with open-world generaliza- tion,” 2025. [Online]. Available: https://arxiv.org/abs/2504.16054

    Pith/arXiv arXiv 2025

  4. [4]

    Learning fine-grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine-grained bimanual manipulation with low-cost hardware,” inRobotics: Science and Systems (RSS), 2023

    2023

  5. [5]

    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), 2024, pp. 12 156–12 163

    2024

  6. [6]

    Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots,

    C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song, “Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots,” inRobotics: Science and Systems (RSS), 2024

    2024

  7. [7]

    Open teach: A versatile teleoperation system for robotic manipulation,

    A. Iyer, Z. Peng, Y . Dai, I. Guzey, S. Haldar, S. Chintala, and L. Pinto, “Open teach: A versatile teleoperation system for robotic manipulation,”arXiv preprint arXiv:2403.07870, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Mobile aloha: Learning bimanual mobile manipulation using low-cost whole-body teleoperation,

    Z. Fu, T. Z. Zhao, and C. Finn, “Mobile aloha: Learning bimanual mobile manipulation using low-cost whole-body teleoperation,” inProceedings of The 8th Conference on Robot Learning, ser. Proceedings of Machine Learning Research, P. Agrawal, O. Kroemer, and W. Burgard, Eds., vol. 270. PMLR, 06–09 Nov 2025, pp. 4066–4083. [Online]. Available: https://procee...

    2025

  9. [9]

    Telemoma: A modular and versatile teleoperation system for mobile manipulation,

    S. Dass, W. Ai, Y . Jiang, S. Singh, J. Hu, R. Zhang, P. Stone, B. Abbatematteo, and R. Mart ´ın-Mart´ın, “Telemoma: A modular and versatile teleoperation system for mobile manipulation,”arXiv preprint arXiv:2403.07869, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Whole-body teleoperation for mobile manipulation at zero added cost,

    D. Honerkamp, H. Mahesheka, J. O. von Hartz, T. Welschehold, and A. Valada, “Whole-body teleoperation for mobile manipulation at zero added cost,”IEEE Robotics and Automation Letters, 2025

    2025

  11. [11]

    Behavior robot suite: Streamlining real-world whole-body manipulation for everyday household activi- ties,

    Y . Jiang, R. Zhang, J. Wong, C. Wang, Y . Ze, H. Yin, C. Gokmen, S. Song, J. Wu, and L. Fei-Fei, “Behavior robot suite: Streamlining real-world whole-body manipulation for everyday household activi- ties,” inConference on Robot Learning (CoRL), 2025

    2025

  12. [12]

    Robopanoptes: The all-seeing robot with whole-body dexterity,

    X. Xu, D. Bauer, and S. Song, “Robopanoptes: The all-seeing robot with whole-body dexterity,” inRobotics: Science and Systems (RSS), 2025

    2025

  13. [13]

    Retargeting matters: General motion retargeting for humanoid motion tracking,

    J. P. Araujo, Y . Ze, P. Xu, J. Wu, and C. K. Liu, “Retargeting matters: General motion retargeting for humanoid motion tracking,” arXiv preprint arXiv:2510.02252, 2025

    arXiv 2025

  14. [14]

    Open-television: Teleoperation with immersive active visual feedback,

    X. Cheng, J. Li, S. Yang, G. Yang, and X. Wang, “Open-television: Teleoperation with immersive active visual feedback,” inProceedings of The 8th Conference on Robot Learning, ser. Proceedings of Machine Learning Research, P. Agrawal, O. Kroemer, and W. Burgard, Eds., vol. 270. PMLR, 06–09 Nov 2025, pp. 2729–2749. [Online]. Available: https://proceedings.m...

    2025

  15. [15]

    Bunny-visionpro: Real-time bimanual dexterous teleoperation for imitation learning,

    R. Ding, Y . Qin, J. Zhu, C. Jia, S. Yang, R. Yang, X. Qi, and X. Wang, “Bunny-visionpro: Real-time bimanual dexterous teleoperation for imitation learning,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025, pp. 12 248–12 255

    2025

  16. [16]

    Dexcap: Scalable and portable mocap data collection system for dexterous manipulation,

    C. Wang, H. Shi, W. Wang, R. Zhang, L. Fei-Fei, and C. K. Liu, “Dexcap: Scalable and portable mocap data collection system for dexterous manipulation,” inRobotics: Science and Systems (RSS), 2024

    2024

  17. [17]

    Fastumi: A scalable and hardware- independent universal manipulation interface with dataset,

    Z. Zhaxizhuoma, K. Liu, C. Guan, Z. Jia, Z. Wu, X. Liu, T. Wang, S. Liang, P. Chen, P. Zhang,et al., “Fastumi: A scalable and hardware- independent universal manipulation interface with dataset,” inConfer- ence on Robot Learning. PMLR, 2025, pp. 3069–3093

    2025

  18. [18]

    Hommi: Learning whole-body mobile manipulation from human demonstrations,

    X. Xu, J. Park, H. Zhang, E. Cousineau, A. Bhat, J. Barreiros, D. Wang, J. Bohg, and S. Song, “Hommi: Learning whole-body mobile manipulation from human demonstrations,”arXiv preprint arXiv:2603.03243, 2026

    Pith/arXiv arXiv 2026

  19. [19]

    Octo: An open-source generalist robot policy,

    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

  20. [20]

    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

  21. [21]

    π 0: A vision-language-action flow model for general robot control,

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al., “π 0: A vision-language-action flow model for general robot control,” 2026. [Online]. Available: https://arxiv.org/abs/2410.24164

    Pith/arXiv arXiv 2026

  22. [22]

    π 0.7: a steerable generalist robotic foundation model with emergent capabilities,

    P. Intelligence, B. Ai, A. Amin, R. Aniceto, A. Balakrishna, G. Balke, K. Black,et al., “π 0.7: a steerable generalist robotic foundation model with emergent capabilities,” 2026. [Online]. Available: https://arxiv.org/abs/2604.15483

    Pith/arXiv arXiv 2026

  23. [23]

    Depthvla: Enhancing vision-language-action models with depth- aware spatial reasoning,

    T. Yuan, Y . Liu, C. Lu, Z. Chen, T. Jiang, and H. Zhao, “Depthvla: Enhancing vision-language-action models with depth- aware spatial reasoning,”arXiv preprint arXiv:2510.13375, 2025. [Online]. Available: https://arxiv.org/abs/2510.13375

    arXiv 2025

  24. [24]

    From foundation to application: Improving vla models in practice,

    W. Wu, F. Wang, F. Lu, H. Sun, S. Liu, Y . Wang, Y . Yan, Y . Wang, S. Ma, X. Wang, Y . Liu, S. Yang, T. Zhou, K. Zhang, L. Zhou, C. Su, N. Xue, B. Tan, H. Zhang, Y . Zhang, F. Liao, X. Zhu, Y . Shen, and K. Zheng, “From foundation to application: Improving vla models in practice,”arXiv preprint arXiv:2607.06403, 2026, lingBot-VLA 2.0 technical report. [O...

    Pith/arXiv arXiv 2026

  25. [25]

    Tactilealoha: Learning bimanual manipulation with tactile sensing,

    N. Gu, K. Kosuge, and M. Hayashibe, “Tactilealoha: Learning bimanual manipulation with tactile sensing,”IEEE Robotics and Automation Letters, vol. 10, no. 8, pp. 8348–8355, 2025. [Online]. Available: https://doi.org/10.1109/LRA.2025.3585396

    arXiv 2025

  26. [26]

    Tact: Humanoid whole- body contact manipulation through deep imitation learning with tactile modality,

    M. Murooka, T. Hoshi, K. Fukumitsu, S. Masuda, M. Hamze, T. Sasaki, M. Morisawa, and E. Yoshida, “Tact: Humanoid whole- body contact manipulation through deep imitation learning with tactile modality,”IEEE Robotics and Automation Letters, 2025, accepted June 2025; arXiv preprint 2506.15146. [Online]. Available: https://arxiv.org/abs/2506.15146

    Pith/arXiv arXiv 2025

  27. [27]

    Momanipvla: Transfer- ring vision-language-action models for general mobile manipulation,

    Z. Wu, Y . Zhou, X. Xu, Z. Wang, and H. Yan, “Momanipvla: Transfer- ring vision-language-action models for general mobile manipulation,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  28. [28]

    Harmonic mobile manipulation,

    R. Yang, Y . Kim, R. Hendrix, A. Kembhavi, X. Wang, and K. Ehsani, “Harmonic mobile manipulation,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 3658–3665

    2024

  29. [29]

    Spin: Simultaneous perception, interaction and navigation,

    S. Uppal, A. Agarwal, H. Xiong, K. Shaw, and D. Pathak, “Spin: Simultaneous perception, interaction and navigation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2024, pp. 18 133–18 142

    2024

  30. [30]

    Sg-vla: Learning spatially-grounded vision-language-action models for mobile manipulation,

    R. Tu, A. Shukla, S. Yoo, X. Li, J. Li, J. Xie, H. Su, and Z. Tu, “Sg-vla: Learning spatially-grounded vision-language-action models for mobile manipulation,”arXiv preprint arXiv:2603.22760, 2026

    arXiv 2026

  31. [31]

    Ac-dit: Adaptive coordination diffusion transformer for mobile manipulation,

    S. Chen, J. Liu, S. Qian, H. Jiang, L. Li, R. Zhang, Z. Liu, C. Gu, C. Hou, P. Wang, Z. Wang, and S. Zhang, “Ac-dit: Adaptive coordination diffusion transformer for mobile manipulation,” 2025. [Online]. Available: https://arxiv.org/abs/2507.01961

    Pith/arXiv arXiv 2025

  32. [32]

    Learning human-to-humanoid real-time whole-body teleoperation,

    T. He, Z. Luo, W. Xiao, C. Zhang, K. Kitani, C. Liu, and G. Shi, “Learning human-to-humanoid real-time whole-body teleoperation,” arXiv preprint arXiv:2403.04436, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Omnih2o: Universal and dexterous human- to-humanoid whole-body teleoperation and learning,

    T. He, Z. Luo, X. He, W. Xiao, C. Zhang, W. Zhang, K. Kitani, C. Liu, and G. Shi, “Omnih2o: Universal and dexterous human- to-humanoid whole-body teleoperation and learning,”arXiv preprint arXiv:2406.08858, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    Twist: Teleoperated whole-body imitation system,

    Y . Ze, Z. Chen, J. P. Ara ´ujo, Z. ang Cao, X. B. Peng, J. Wu, and C. K. Liu, “Twist: Teleoperated whole-body imitation system,”arXiv preprint arXiv:2505.02833, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Mtpano: Multi-task panoramic scene understanding via label-free integration of dense prediction priors,

    J. Zhang, X. Zhan, L. Zhang, Y . Wang, Z. Yu, J. Wang, W. Wang, and X. Li, “Mtpano: Multi-task panoramic scene understanding via label-free integration of dense prediction priors,” 2026. [Online]. Available: https://arxiv.org/abs/2602.05330

    Pith/arXiv arXiv 2026

  36. [36]

    Aha: A vision- language-model for detecting and reasoning over failures in robotic manipulation,

    J. Duan, W. Pumacay, N. Kumar, Y . R. Wang, S. Tian, W. Yuan, R. Krishna, D. Fox, A. Mandlekar, and Y . Guo, “Aha: A vision- language-model for detecting and reasoning over failures in robotic manipulation,”arXiv preprint arXiv:2410.00371, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Rethinking the practicality of vision- language-action model: A comprehensive benchmark and an improved baseline,

    W. Song, J. Chen, X. Sun, H. Lei, Y . Qin, W. Zhao, P. Ding, H. Zhao, T. Wang, P. Hou,et al., “Rethinking the practicality of vision- language-action model: A comprehensive benchmark and an improved baseline,”arXiv preprint arXiv:2602.22663, 2026

    arXiv 2026