Pith. sign in

REVIEW 3 major objections 4 minor 60 references

Multi-View Unified Camera Fields: Geometry-Shaped Action-Facing Representations for RGB-Only Multi-Camera VLA Policies

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Training-only injection of metric depth and cross-view correspondence into a multi-camera VLA's hidden states lets deployment stay RGB-only, with matched gains to 98.9% on LIBERO, +22.4 on LIBERO-Plus, +23.3 on RoboTwin.

desk verdict Good idea and honest evaluation, but the central attribution of the gains to geometry is undercut by a missing control for generic upper-backbone fine-tuning. read the letter →

arxiv 2608.01826 v1 pith:PHEUHRZG submitted 2026-08-03 cs.RO

Jiarui Yang , Yehao Lu , Yuning Su , Yufeng Xie , Yu Zhong , Haiyu Lan , Tianjing Hao , Kaixiang Lu
show 4 more authors
Peiwen Lin Chuang Wang Enyu Li Junwei Liang
This is my paper · ORCID
classification cs.RO
keywords vision-language-actionmulti-camerapolicymetricdepthrecoverycross-viewcorrespondencegeometryinjectionRGB-onlydeploymentrepresentationlearningrobotmanipulation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a vision-language-action (VLA) policy can acquire metric 3D awareness and cross-camera point consistency during training, then lose all geometry inputs and still act as if it had them. MVUCF does this with a training-only geometry injection stage: a coordinate-query depth head teaches the action-facing hidden token grid to encode metric depth at continuous locations, while a preprocessing-aware correspondence head aligns tokens that observe the same physical point from different cameras. After injection, the depth head, calibration, and correspondence projector are discarded and the backbone is frozen, so action training and deployment use the original RGB-only graph with no extra inference cost. The paper's conclusion is that geometry does not need to be an input or a module; it can be a permanent property of the representation itself.

What carries the argument

The central object is the action-facing hidden grid at layer 15 of the VLA backbone, treated as a continuous coordinate field. Two training-only heads read from this grid: a coordinate-query depth head that predicts metric depth $\hat{z}_i(q)$ at continuous grid coordinates from a 3074-dimensional query representation (sampled feature, local $x/y$ differences, and sub-cell phase), and a shared cross-view projector that produces L2-normalized embeddings whose global matching logits $\ell_{mn} = 10\,(e_m^s)^\top e_n^t$ are supervised with Gaussian soft targets centered at geometrically projected token centers $q_j^*$. The load-bearing map is the preprocessing-aware coordinate transform $T_{\text{img}\to\text{grid}}$: raw pixels are back-projected, re-projected, visibility-gated at 10 cm, and only then mapped through letterboxing, resizing, cropping, patchification, and pixel-unshuffle to token coordinates, so labels are correct on the actual grid that the action head consumes. After injection both heads are discarded, which is what preserves the original RGB-only inference graph.

What would settle it

Retrain the geometry-injection stage with camera extrinsics deliberately rotated by, say, 5 degrees during target construction, then evaluate the same RGB-only deployment on LIBERO-Plus camera and layout perturbations; if the +22.4-point average gain collapses or the depth probe MAE stops improving while the training loss stays low, the central claim fails for realistic calibration error.

Watch

Extended reading notes

Core claim

The central discovery is that the layer-15 hidden grid that feeds a VLA's action module can be reshaped so that metric depth is recoverable at arbitrary continuous grid coordinates and so that tokens observing the same world point are mutually retrievable across cameras, even though depth, calibration, and all auxiliary heads are absent at deployment. The depth objective supervises 3072 continuous queries per view per step, half uniform and half near depth discontinuities, through a GridTrunk plus a small MLP; the correspondence objective builds positives by back-projecting raw pixels with ground-truth depth and re-projecting them into other views through the exact preprocessing chain $T_{\text{img}\to\text{grid}}$, keeping only pairs that pass image-bound, depth-validity, and 10 cm z-buffer checks. Under a matched GR00T-N1.6 protocol where geometry injection is the only difference, the paper reports depth MAE falling from about 4.9 cm to 0.44 cm, cross-view Hit@1 rising from near-random to 64%, and success improving from 97.4% to 98.9% on LIBERO, by 22.4 points on LIBERO-Plus, and by 23.3 points across six RoboTwin tasks. The claim is that a multi-view camera field can be injected as a property of action-facing representations rather than as an input.

Load-bearing premise

The premise that carries the whole framework is that training-time synchronization, camera calibration, and depth are accurate enough that the 10 cm z-buffer gate keeps the cross-view positives geometrically correct; the paper states that robustness to corrupted or noisy calibration is untested, and biased extrinsics or depth would misalign both supervision signals.

Editorial extensions

If this is right

  • A multi-camera robot can gain metric depth and cross-view consistency by retraining alone; no depth sensor, calibration stream, or geometry module is needed at runtime.
  • The geometry-shaped representation transfers zero-shot to perturbations in robot appearance, language, noise, background, and layout, but the measured pattern implies camera-viewpoint and field-of-view perturbations improve least because calibration is baked into the learned features.
  • Depth-only and cross-view-only supervision each add a small average gain on LIBERO-10 (about +1.1 and +1.3 points), while the combined objective adds about +4.1 points, so the two signals are complementary rather than redundant.
  • The same coordinate-map mechanism carries over to other visual front-ends: for a standard ViT it reduces to resize, crop, and patch-index transforms, so the method is not tied to one backbone's preprocessing.
  • On the real-robot pilot, the direction of improvement matches simulation (49/60 versus 40/60 successes across two tasks), giving feasibility evidence for RGB-only deployment, though the per-task intervals overlap.

Reading between the lines

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

  • A natural extension the paper does not run is to use the frozen geometry-injected backbone as a metric feature extractor for downstream modules such as grasp planning or affordance localization, since the probes show depth and point identity are readable from the hidden grid itself.
  • Because the visibility gate and both objectives share the same calibration, the untested failure mode is silent misalignment under biased extrinsics; an informative experiment would corrupt training-time poses by small rotations and measure depth MAE and cross-view Hit@1 separately to see which supervision breaks first.
  • The two-stage recipe suggests a transfer protocol: inject geometry once on a large calibrated multi-camera corpus, then adapt the action head on small RGB-only datasets, which could give few-shot policies the spatial grounding that sparse action labels cannot provide.
  • The LIBERO-Plus result implies the learned camera field is partly a function of training-time camera parameters, so a deployment robot with a different camera rig would likely need either re-injection under its own calibration or explicit viewpoint augmentation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces MVUCF, a training-only framework that injects geometric structure into the upper backbone layers of a GR00T-N1.6 VLA policy by jointly optimizing a coordinate-query metric-depth objective and a preprocessing-aware cross-view correspondence objective. After this geometry-injection stage, the auxiliary heads are discarded and the backbone is frozen, so deployment remains the original RGB-only inference graph. The authors report held-out representation probes showing improved depth recovery and cross-view matching, matched-policy success-rate improvements on LIBERO (98.9% vs 97.4%), LIBERO-Plus (+22.4 points), and a six-task RoboTwin suite (+23.3 points), plus a small real-robot pilot. The central claim is that the geometry-injection stage, and not any other difference, reshapes the action-facing representations and produces these gains.

Significance. If the attribution is correct, this is a practically valuable result: it improves multi-camera VLA manipulation without changing the inference-time interface, adding no parameters or FLOPs at deployment. The paper has several methodological strengths: the matched GR00T-N1.6 pair shares action data, head, schedule, checkpoint rule, and inference graph; LIBERO results use three independently seeded checkpoints; the RoboTwin task suite was prespecified before training; terminal checkpoints are used without downstream-success selection; and the ablation decomposes the two geometry objectives. The layer-wise VGGT diagnostic and the attention-localization figure are useful additional evidence. The main weakness is that the matched comparison does not control for the presence of the 50k-step upper-backbone fine-tuning stage itself, so the headline gains cannot yet be uniquely attributed to the geometric content of the auxiliary losses.

major comments (3)
  1. [Appendix A / Tables 7 and 9 / Table 4] The central attribution claim is not yet supported because the matched GR00T-N1.6 comparison does not isolate geometry injection. Ours receives 50k gradient steps that update VLM layers 8–15 under the depth and cross-view losses, while Base receives no such upper-backbone update before action training (Appendix A, Tables 7 and 9). Consequently, any benefit from additional feature adaptation, feature smoothing, or continued pretraining of those layers is attributed to geometry. The ablation in Table 4 cannot close this gap: Depth-only and Cross-view-only are both geometry-supervision objectives, so their improvements over Base are also consistent with a generic auxiliary-fine-tuning effect. A control that runs the same 50k update on a non-geometric objective (for example, masked token reconstruction, random-view contrastive matching, or next-token prediction on the same data) is required to support the claim that the camera field itself, rather than extra backbone fine-tuning, drives the LIBERO-Plus (+22.4) and RoboTwin (+23.3) gains.
  2. [Table 2 / Appendix F] The LIBERO-Plus robustness result is reported for a single seed-0 terminal checkpoint, without per-seed variance or additional seeds. The 22.4-point average improvement and the claim that Ours performs better across all seven categories are therefore based on one training run. Since the paper already runs three-seed evaluations for standard LIBERO, the same protocol should be applied to LIBERO-Plus (or at least two additional seeds) before the robustness improvement can be considered established.
  3. [Appendix G / Section 'Evaluation protocol'] The paper states that the supervision tap layer (layer 15) and the update range (layers 8–15) were selected during method development based on observed same-point discrimination, with the layer-12 variant degrading token-level separation. This selection happened before downstream evaluation, which is good, but it is a form of model selection on an auxiliary metric. The report would be strengthened by stating explicitly how many such auxiliary-metric choices were tried and whether any of them interacted with the downstream results; as written, the reader cannot fully rule out that the choice was influenced by knowledge of the action outcomes.
minor comments (4)
  1. [Scope and outlook / Appendix H] The manuscript explicitly acknowledges that robustness to corrupted or noisy calibration labels is untested and that the real-world depth pipeline relies on LingBot-Depth completion. This limitation is appropriately stated, and for the simulated benchmarks the depth and calibration are exact, so it does not undercut the simulation claims; however, the real-robot pilot should be described more cautiously than as 'practical effectiveness' given that calibration accuracy and depth completion are unvalidated in that setting.
  2. [Figure 8] The real-robot pilot reports 49/60 versus 40/60 successes with overlapping Wilson confidence intervals and only 30 trials per policy per task. The paper correctly frames this as pilot-scale evidence, but the abstract's phrase "real-world humanoid experiments further provide evidence of its practical effectiveness" slightly overstates the strength of the result.
  3. [Representation diagnostics (Figure 6)] The held-out depth and cross-view retrieval probes largely measure the same properties that were directly optimized by the auxiliary losses, so their improvement on held-out data is expected and should be interpreted as evidence that the injection worked, not as independent evidence that the action improvements stem from geometry. The paper's own framing is mostly careful about this, but Figure 6's caption could be read as implying more.
  4. [Abstract / formatting] The abstract contains 'Multi-ViewUnifiedCameraFields' and 'Multi-ViewUnifiedCameraFields' without a space between 'View' and 'Unified'; this appears to be a LaTeX/formatting artifact and should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; downstream gains are not fitted by the geometry objectives and the benchmark comparisons are external to the training targets.

full rationale

The paper's central derivation is self-contained rather than circular. The geometry-injection stage supervises upper VLM layers with a coordinate-query depth loss (Eqs. 3-4, 18) and a reprojection-based correspondence loss (Eqs. 5-6, 22-26), using ground-truth depth, intrinsics, extrinsics, and a z-buffer visibility gate. These targets are then discarded, the auxiliary heads are removed, and the action module is trained on the frozen, geometry-shaped backbone with RGB-only inputs. No downstream success metric is used to select the auxiliary losses, the 10 cm visibility threshold, the query-sampling strategy, or the training horizon; the paper states explicitly that the visibility gate is fixed 'without using downstream success to select the value,' and that settings are 'fixed across datasets rather than tuned per benchmark.' The held-out depth and cross-view retrieval probes do measure quantities closely related to the training objectives, so their improvement is expected, but they are evaluated on held-out data and serve as generalization checks rather than as fitted predictions presented as novel results. The main manipulation claims rest on matched Base-vs-Ours comparisons on LIBERO, LIBERO-Plus, and RoboTwin, and the reported success rates are not derivable from the auxiliary losses by construction. The acknowledged limitations, including untested calibration robustness (Scope and outlook; Appendix H) and the absence of a matched non-geometric 50k-step fine-tuning control (Appendix A, Tables 7 and 9), are experimental-confounding or robustness concerns, not circularity: they do not make any benchmark number equivalent to the training inputs. There is also no load-bearing self-citation chain or imported uniqueness theorem; citations to GR00T, VGGT, Selfi, and other prior work are external and used as context or baselines.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on open simulation benchmarks and a matched GR00T pair, not on a fitted equation. However, the reported gains depend on several manually chosen hyperparameters, the untested calibration-accuracy assumption, and the assumption that upper-VLM fine-tuning itself is not the cause of improvement. The paper introduces no new physical entity.

free parameters (6)
  • Depth loss weights (SILog/inverse/gradient/seam/uncertainty) = 0.5 / 1.25 / 1.0 / 0.1 / 0.05
    Manually fixed loss weighting for the coordinate-query depth objective (Eq. 4, Appendix C); affects how strongly geometry shapes the hidden states and is not derived from any principle.
  • Cross-view loss weights and outer anneal = 0.1 / 0.1 / 0.1 with outer CV weight 1.0 to 0.1 over 40k steps
    Manual weights for soft cross-entropy, hard-negative InfoNCE, and margin terms (Eq. 6, Appendix D); the annealed schedule is chosen by hand.
  • Visibility z-buffer threshold delta_vis = 0.10 m
    Empirically selected operating point between positive-pair yield and reprojection precision (Appendix D); all cross-view positives pass through this gate.
  • Soft-label sigma and margin gamma = sigma = 0.75, gamma = 0.2
    Fix the spatial softness of correspondence targets and the required logit gap (Appendix D); chosen without downstream-success tuning.
  • Global logit scale tau and local window = tau = 10.0, 3x3 window with scale annealed 50 to 30
    Fixed non-learnable scales for matching logits and local refinement; manually chosen constants in the matching objective.
  • Supervision tap layer and update range = layer 15, layers 8-15
    Selected after internal VLM probing because layer 12 degraded token-level separation (Appendix G); not tuned on downstream success, but a manual architectural choice that affects the reported results.
assumptions (5)
  • domain assumption Training-time depth maps, intrinsics, and extrinsics are synchronized and accurate for all target construction.
    Cross-view positives are defined by back-projection with ground-truth depth and calibrated poses (Eq. 2, Algorithm 2, Appendix D); the paper states robustness to noisy calibration is untested in Scope and outlook.
  • domain assumption Ground-truth or completed depth preserves metric scale within [0.05, 5.0] m in the training data.
    Depth supervision and z-buffer checks consume these depths (Appendix C, H); invalid holes are masked or completed, but biased depth would bias geometry targets.
  • domain assumption The full GR00T preprocessing chain is exactly representable as T_img->grid in Eq. (11).
    Preprocessing-aware mapping is central to avoiding coordinate mismatch between raw pixels and token-grid targets; the paper relies on code-level detail of Eagle/SigLIP preprocessing.
  • ad hoc to paper LingBot-Depth pretrained completion does not bias metric depth for transparent, reflective, or textureless regions.
    Used only for real-world depth completion before target construction (Appendix H); no validation on this robot's distribution is reported.
  • domain assumption The observed gains are attributable to the geometry objectives rather than to generic fine-tuning of layers 8-15 on the same data.
    Ablations compare Depth-only, Cross-view-only, and Full, but no condition updates the same layers with a neutral auxiliary objective, so this attribution is assumed rather than controlled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-View Unified Camera Fields: Geometry-Shaped Action-Facing Representations for RGB-Only Multi-Camera VLA Policies." pith.science (2026). https://pith.science/paper/PHEUHRZG

@misc{pith2026260801826,
  author       = {Pith},
  title        = {Pith review of: Multi-View Unified Camera Fields: Geometry-Shaped Action-Facing Representations for RGB-Only Multi-Camera VLA Policies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PHEUHRZG}},
  note         = {Machine review of arXiv:2608.01826}
}
read the original abstract

Vision-Language-Action (VLA) models have shown strong generalization in robotic manipulation, yet complex contact-rich tasks often benefit from multi-camera observations that jointly capture the end effector, objects, and targets under occlusion. Existing multi-camera VLAs usually concatenate view tokens, leaving action representations weak in metric depth and inconsistent across cameras. We introduce Multi-View Unified Camera Fields (MVUCF), a training-only framework that forms a shared action-facing latent field across views. A coordinate-query depth objective makes metric depth recoverable, while a preprocessing-aware correspondence objective aligns tokens observing the same physical point from different cameras. Both directly shape the hidden states consumed by the action module. After geometry injection, depth, camera calibration, and auxiliary heads are removed, so deployment uses the original RGB-only graph with no extra inference FLOPs. Held-out probes confirm stronger depth recovery and cross-view matching. Under matched GR00T-N1.6 settings, MVUCF reaches 98.9% on LIBERO, improves LIBERO-Plus by 22.4 points, and raises success by 23.3 points across six RoboTwin tasks spanning three action families: touch, move-and-place, and contact interaction. Real-world humanoid experiments further provide evidence of its practical effectiveness under RGB-only deployment.

Figures

Figures reproduced from arXiv: 2608.01826 by the authors.

Figure 1
Figure 1. Overview of Multi-View Unified Camera Fields. Calibrated head-and-wrist views provide complementary observations; training-time depth and reprojection supervision reshape action-facing tokens, while deployment remains RGB-only with the native inference graph [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Geometry-injection diagnostics. (a) With the same depth-head probe, metric depth is difficult to extract from the native [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. System architecture of the proposed Multi-View Unified Camera Fields framework. During geometry injection, the [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Cross-view target construction. Depth lifts a source [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Layer-wise VGGT diagnostic. Shallow features pre [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Full-evaluation diagnostics. Unlike the single-frame per-view visualization in Figure 2, the results here report full [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Action-to-image cross-attention for move_pillbottle_pad at matched rollout steps. Independent normalization highlights the spatial support within each panel; the comparison concerns localization rather than absolute attention magnitude. Attention extrac￾tion details ar…
Figure 8
Figure 8. Figure 8: Real-robot RGB-only pilot on Agibot Expedition [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 48 canonical work pages

  1. [1]

    Zitkovich, Brianna and others , booktitle =

  2. [2]

    Kim, Moo Jin and others , booktitle =

  3. [3]

    and Sadigh, Dorsa and Finn, Chelsea and Levine, Sergey , booktitle =

    Ghosh, Dibya and Walke, Homer Rich and Pertsch, Karl and Black, Kevin and Mees, Oier and Dasari, Sudeep and Hejna, Joey and Kreiman, Tobias and Xu, Charles and Luo, Jianlan and Tan, You Liang and Chen, Lawrence Yunliang and Vuong, Quan and Xiao, Ted and Sanketi, Pannag R. and Sadigh, Dorsa and Finn, Chelsea and Levine, Sergey , booktitle =. 2024 , doi =

    work page 2024

  4. [4]

    Proceedings of Robotics: Science and Systems , year =

    _0 : A Vision-Language-Action Flow Model for General Robot Control , author =. Proceedings of Robotics: Science and Systems , year =

    work page

  5. [6]

    2025 , howpublished =

    2025

  6. [7]

    Proceedings of Robotics: Science and Systems , year =

    Diffusion Policy: Visuomotor Policy Learning via Action Diffusion , author =. Proceedings of Robotics: Science and Systems , year =

  7. [8]

    Proceedings of Robotics: Science and Systems , year =

    Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware , author =. Proceedings of Robotics: Science and Systems , year =

  8. [9]

    Zhen, Haoyu and others , booktitle =

    work page

Show all 60 references
  1. [10]

    Wang, Jianyuan and others , booktitle =

  2. [11]

    Wang, Shuzhe and others , booktitle =

  3. [12]

    Proceedings of the International Conference on Learning Representations , year =

    Spatial Forcing: Implicit Spatial Representation Alignment for Vision-Language-Action Model , author =. Proceedings of the International Conference on Learning Representations , year =

  4. [13]

    Deng, Youming and Peng, Songyou and Zhang, Junyi and Heal, Kathryn and Sun, Tiancheng and Flynn, John and Marschner, Steve and Chai, Lucy , booktitle =

  5. [14]

    Proceedings of the Conference on Robot Learning , series =

    Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation , author =. Proceedings of the Conference on Robot Learning , series =

  6. [15]

    Liu, Bo and others , booktitle =

  7. [16]

    Fei, Senyu and Wang, Siyin and Shi, Junhao and Dai, Zihao and Cai, Jikun and Qian, Pengfang and Ji, Li and He, Xinzhe and Zhang, Shiduo and Fei, Zhaoye and Fu, Jinlan and Gong, Jingjing and Qiu, Xipeng , booktitle =

  8. [17]

    Mu, Yao and others , booktitle =

  9. [18]

    Peng, Yue and others , eprint =

  10. [19]

    Zheng, Ruijie and others , booktitle =

  11. [20]

    Hou, Zhi and others , booktitle =

  12. [21]

    Zhao, Qingqing and others , booktitle =

  13. [22]

    Pertsch, Karl and others , booktitle =

  14. [23]

    Proceedings of Robotics: Science and Systems , year =

    Learning to Act Anywhere with Task-Centric Latent Actions , author =. Proceedings of Robotics: Science and Systems , year =

  15. [24]

    Proceedings of Robotics: Science and Systems , year =

    Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success , author =. Proceedings of Robotics: Science and Systems , year =

  16. [25]

    Qu, Delin and others , booktitle =

  17. [26]

    Sun, Lin and others , eprint =

  18. [27]

    Bhat, Vineet and Lan, Yu-Hsiang and Krishnamurthy, Prashanth and Karri, Ramesh and Khorrami, Farshad , eprint =

  19. [28]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Probing the 3D Awareness of Visual Foundation Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  20. [29]

    Lin, Tao and others , eprint =

  21. [31]

    Proceedings of the International Conference on Learning Representations , year =

    Actions as Language: Fine-Tuning VLMs into VLAs Without Catastrophic Forgetting , author =. Proceedings of the International Conference on Learning Representations , year =

  22. [32]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle =

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle =

  23. [33]

    Bhat, V.; Lan, Y.-H.; Krishnamurthy, P.; Karri, R.; and Khorrami, F. 2025. 3D-CAVLA : Leveraging Depth and 3D Context to Generalize Vision-Language-Action Models for Unseen Tasks. Accepted at the 1st Workshop on 3D-LLM/VLA, CVPR 2025, arXiv:2505.05800

    2025

  24. [34]

    Black, K.; et al. 2025. _0 : A Vision-Language-Action Flow Model for General Robot Control. In Proceedings of Robotics: Science and Systems

    2025

  25. [35]

    Bu, Q.; et al. 2025. Learning to Act Anywhere with Task-Centric Latent Actions. In Proceedings of Robotics: Science and Systems

    2025

  26. [36]

    Chi, C.; et al. 2023. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. In Proceedings of Robotics: Science and Systems

    2023

  27. [37]

    Deng, Y.; Peng, S.; Zhang, J.; Heal, K.; Sun, T.; Flynn, J.; Marschner, S.; and Chai, L. 2026. Selfi : Self-improving Reconstruction Engine via 3D Geometric Feature Alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7351--7361

    2026

  28. [38]

    El Banani, M.; et al. 2024. Probing the 3D Awareness of Visual Foundation Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21795--21806

    2024

  29. [39]

    Fei, S.; Wang, S.; Shi, J.; Dai, Z.; Cai, J.; Qian, P.; Ji, L.; He, X.; Zhang, S.; Fei, Z.; Fu, J.; Gong, J.; and Qiu, X. 2026. LIBERO-Plus : A Progressive Robustness Benchmark for Visual-Language-Action Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and ...

    2026

  30. [40]

    R.; Pertsch, K.; Black, K.; Mees, O.; Dasari, S.; Hejna, J.; Kreiman, T.; Xu, C.; Luo, J.; Tan, Y

    Ghosh, D.; Walke, H. R.; Pertsch, K.; Black, K.; Mees, O.; Dasari, S.; Hejna, J.; Kreiman, T.; Xu, C.; Luo, J.; Tan, Y. L.; Chen, L. Y.; Vuong, Q.; Xiao, T.; Sanketi, P. R.; Sadigh, D.; Finn, C.; and Levine, S. 2024. Octo : An Open-Source Generalist Robot Policy. In Proceeding...

    2024

  31. [41]

    J.; Wu, X.; Zha, L.; Russakovsky, O.; and Majumdar, A

    Hancock, A. J.; Wu, X.; Zha, L.; Russakovsky, O.; and Majumdar, A. 2026. Actions as Language: Fine-Tuning VLMs into VLAs Without Catastrophic Forgetting. In Proceedings of the International Conference on Learning Representations

    2026

  32. [42]

    Hou, Z.; et al. 2025. Dita : Scaling Diffusion Transformer for Generalist Vision-Language-Action Policy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 7686--7697

    2025

  33. [43]

    J.; Finn, C.; and Liang, P

    Kim, M. J.; Finn, C.; and Liang, P. 2025. Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. In Proceedings of Robotics: Science and Systems

    2025

  34. [44]

    J.; et al

    Kim, M. J.; et al. 2025. OpenVLA : An Open-Source Vision-Language-Action Model. In Proceedings of the Conference on Robot Learning, volume 270 of PMLR, 2679--2713

    2025

  35. [45]

    Li, F.; Song, W.; Zhao, H.; Wang, J.; Ding, P.; Wang, D.; Zeng, L.; and Li, H. 2026. Spatial Forcing: Implicit Spatial Representation Alignment for Vision-Language-Action Model. In Proceedings of the International Conference on Learning Representations

    2026

  36. [46]

    Lin, T.; et al. 2026. Evo-Depth : A Lightweight Depth-Enhanced Vision-Language-Action Model. Concurrent public preprint, arXiv:2605.14950

    2026 arXiv

  37. [47]

    Liu, B.; et al. 2023. LIBERO : Benchmarking Knowledge Transfer for Lifelong Robot Learning. In Advances in Neural Information Processing Systems, volume 36, 44776--44791

    2023

  38. [48]

    Mu, Y.; et al. 2025. RoboTwin : Dual-Arm Robot Benchmark with Generative Digital Twins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 27649--27660

    2025

  39. [49]

    NVIDIA . 2025. NVIDIA Isaac GR00T N1.6-3B . Hugging Face model card, https://huggingface.co/nvidia/GR00T-N1.6-3B

    2025

  40. [50]

    NVIDIA ; Bjorck, J.; Casta \ n eda, F.; et al. 2025. GR00T N1 : An Open Foundation Model for Generalist Humanoid Robots. arXiv:2503.14734

    2025 arXiv

  41. [51]

    Peng, Y.; et al. 2026. G ^3 VLA : Geometric Inductive Bias for Vision-Language-Action Models. Concurrent public preprint, arXiv:2606.24472

    2026 arXiv

  42. [52]

    Pertsch, K.; et al. 2025. FAST : Efficient Action Tokenization for Vision-Language-Action Models. In Proceedings of Robotics: Science and Systems

    2025

  43. [53]

    Qu, D.; et al. 2025. SpatialVLA : Exploring Spatial Representations for Visual-Language-Action Models. In Proceedings of Robotics: Science and Systems

    2025

  44. [54]

    Shridhar, M.; Manuelli, L.; and Fox, D. 2023. Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation. In Proceedings of the Conference on Robot Learning, volume 205 of PMLR, 785--799

    2023

  45. [55]

    Sun, L.; et al. 2025. GeoVLA : Empowering 3D Representations in Vision-Language-Action Models. arXiv:2508.09071

    2025 arXiv

  46. [56]

    Wang, J.; et al. 2025. VGGT : Visual Geometry Grounded Transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5294--5306

    2025

  47. [57]

    Wang, S.; et al. 2024. DUSt3R : Geometric 3D Vision Made Easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20697--20709

    2024

  48. [58]

    Zhao, Q.; et al. 2025. CoT-VLA : Visual Chain-of-Thought Reasoning for Vision-Language-Action Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1702--1713

    2025

  49. [59]

    Z.; et al

    Zhao, T. Z.; et al. 2023. Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. In Proceedings of Robotics: Science and Systems

    2023

  50. [60]

    Zheng, R.; et al. 2025. TraceVLA : Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies. In Proceedings of the International Conference on Learning Representations

    2025

  51. [61]

    Zitkovich, B.; et al. 2023. RT -2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. In Proceedings of the Conference on Robot Learning, volume 229 of PMLR, 2165--2183

    2023

  52. [62]

    Tan, B.; Sun, C.; Qin, X.; Adai, H.; Fu, Z.; Zhou, T.; Zhang, H.; Xu, Y.; Zhu, X.; Shen, Y.; and Xue, N. 2026. Masked Depth Modeling for Spatial Perception. Accepted to ECCV 2026; public code and pretrained weights released, arXiv:2601.17895

    2026

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.