Computer Vision and Pattern Recognition · arXiv:2610.05026 v1 · submitted 4 Oct 2026 · 32 references
GeoBridge-VLA: Geometry-Aware Residual Adaptation for Vision-Language-Action Models
The paper says
“Vision-language-action (VLA) models encode semantic information from vision-language pretraining, but manipulation also requires precise spatial reasoning. We present GeoBridge-VLA, a two-stage method for learning geometric features from a pretrained VLA's frozen visual encoder and using them for action prediction.”
more of the abstract
Stage I trains a feature bridge and geometry decoder with depth supervision. Stage II freezes these modules and trains a gated residual interface together with the action-side projections and action expert. The residual augments the existing visual tokens without adding a second image encoder or increasing the token count. Deployment requires RGB, robot state, and language, but no depth observations. Under matched evaluation conditions, GeoBridge-VLA achieves 70.9% success on LIBERO, compared with 60.0% for SmolVLA. Disabling the residual in the same trained checkpoint reduces success from 70.90% to 69.85%, with mixed effects across suites. On a physical ROBOTIS OMY robot, GeoBridge-VLA succeeds in 148 of 200 trials (74.0%) across four tasks, compared with 108 of 200 (54.0%) for SmolVLA.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 6 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. Genuinely distinct geometry-retrofit mechanism, honestly reported, but the load-bearing claim that geometry drives the gains is unsupported without a no-geometry control. This line is wrong
2 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 32 references say.
The paper says, at §V-C (Table IV), §VI (b176-b179)
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.95):
“Because the zero-residual condition retains the weights learned with the geometry pathway, this intervention measures the trained policy’s inference-time sensitivity to the geometry-to-action interface rather than isolating the contribution of geometry during policy learning.”
Pith says: Objection 1 of 2 · major
The central mechanism claim is not supported by the presented evidence. The paper attributes the 10.9-pp LIBERO gain and 20-pp OMY gain to the depth-supervised geometric features (abstract, §I, b186), but the only direct evidence for the geometry-to-action pathway is the same-checkpoint residual ablation: 70.90% vs 69.85% overall, and per-suite signs are mixed (Spatial +3.20, Object -1.20, Goal -0.60, Long +2.80).
more of objection 1
As the paper correctly acknowledges (b178-b179), the zero-residual condition retains all weights learned in Stage II, so the ablation measures inference-time sensitivity only and cannot separate the contribution of geometry from generic effects of the added residual parameters, the altered optimization trajectory, or the jointly trained action-side projections. Because this is precisely the distinction on which the paper's novelty claim rests, the missing separately trained no-geometry control is load-bearing rather than a presentational gap. I recommend adding a Stage-II control that trains the identical residual interface, action projections, and action expert under identical compute and data, but with the geometric bridge/decoder features replaced by a no-geometry input (e.g., a frozen random projection of the same pooled encoder features, or the same bridge trained without depth supervision). Without such a control, the conclusions should be downgraded from 'depth-supervised geometry drives the gains' to 'the complete pipeline improves success', which is substantially weaker and does not establish geometry retrofit as the operative mechanism.
The paper says, at §V-C (Table IV)
The review did not quote this passage.
Pith says: Objection 2 of 2 · major
The per-suite residual effects are small relative to the statistical resolution of the evaluation. Each suite uses 500 episodes (50 per task x 10 tasks), so the standard error of a success rate is about 2 percentage points at values around 70-80%.
more of objection 2
Under that resolution, Spatial +3.20 and Long +2.80 are marginal, and Object -1.20 and Goal -0.60 are consistent with no effect. No multiple-seed results, confidence intervals, or per-episode breakdowns are reported anywhere in the paper (Tables II, IV, V). Given that the residual ablation is the only direct evidence for the proposed mechanism, reporting seed-level variance or confidence intervals would materially change how much weight can be placed on the +1.05-pp overall figure. This also bears on the residual-regularization claim in Eq. 4, where the effect of the penalty lambda_res is never isolated.
5 minor objections
The paper says, at §V-B (Table III, b231)
Pith says: minor objection 1 of 5
The main-text presentation of Table III should carry the caveat the authors already state in the appendix (b231): the SmolVLA shared-encoder and DA3 numbers come from different validation populations and do not define a controlled cross-backbone ranking. As printed, the main-text sentence reads as a favorable comparison against DA3, while delta1 (0.975 vs 0.988) actually favors DA3.
The paper says, at §IV-A
Pith says: minor objection 2 of 5
Please clarify whether the 1,693 demonstrations and 273,465 action frames are totals across the four LIBERO suites or per-suite; the current phrasing is ambiguous and matters for interpreting the 50,000-step Stage-II budget.
show 3 more
The paper says, at §III-A, Eq. 5
Pith says: minor objection 3 of 5
Several design choices are fixed without sensitivity analysis: the selected intermediate layers F^l, the K4 offsets delta_k, the loss coefficients in Eq. 2, and the gate initialization alpha0 in Eq. 3. A short ablation or at least a statement of the chosen values and their stability would strengthen the paper; the K4 offsets in particular are presented as a design improvement but only compared inside the authors' own decoder line.
The paper says, at Fig. 6, Fig. 7
Pith says: minor objection 4 of 5
The depth visualizations are described as 'qualitatively selected'. Please state the selection criterion (best-case? random?) so the reader can calibrate how representative they are of the reported AbsRel/rmse/delta1 averages.
The paper says, at §VI, b179 (text near Table V)
Pith says: minor objection 5 of 5
The sentence beginning 'A separately trained no-geometry control would be needed' is split mid-sentence by the Table V header, which interrupts the argument at the exact point where the paper makes its most important caveat. Please reflow.
What would settle it
Train a no-geometry control that runs Stage II identically (same data, steps, batch size, and regularization) but with the feature bridge and geometry decoder randomly initialized and frozen, so the residual interface receives random or uninformative features rather than depth-supervised geometry. Evaluate this control on the same LIBERO evaluation protocol (2000 episodes) and, if feasible, on the OMY tasks with matched trials. If the control achieves success close to GeoBridge-VLA (within a few points), the depth-supervised geometric signal is not the cause of the gains; if it falls well below, the geometry retrofitting is supported. Additionally, report per-task OMY results with confidence intervals or error bars given the 50-trial-per-task sample size.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 6 Oct 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 2 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The paper's actual contribution is narrower and more specific than the headline numbers suggest, and the authors mostly know this. What is new is the shared-hierarchy retrofit: decode depth from frozen intermediate VLA features, then inject a gated residual into the native visual tokens. No second encoder, no extra tokens. That is distinct from DepthVLA and SpatialVLA, which bolt on modules, and from IVRA and RetoVLA, which repurpose existing signals. The claim that a frozen VLA encoder retains recoverable geometric structure is worth taking seriously, and the depth numbers (AbsRel 0.030 on the shared encoder) are plausible for a frozen backbone.
The paper is also well run on its own terms: two-stage training is specified clearly, evaluation conditions are matched across methods, and the reporting is honest. The residual ablation in Table IV is presented for what it is — an inference-time probe — and the authors explicitly say (b179) that a separately trained no-geometry control is needed. I am not catching them hiding anything.
But that acknowledgment does not fix the problem. The ablation is a 1.05-point effect with mixed per-suite signs: it helps Spatial and Long, hurts Object and Goal. Because it disables the residual in the same trained checkpoint, it retains every training-time influence — the extra residual parameters, the altered optimization trajectory, the jointly trained action projections. So it cannot separate the depth-supervised geometry from generic adaptation effects. The 10.9-point LIBERO and 20-point OMY gains over SmolVLA are real measurements of the full pipeline, but attributing them to geometry is exactly what the evidence does not support. One control experiment — Stage II trained without the geometry pathway — would settle it. If most of the gain survives without geometry, the retrofit hypothesis collapses into a fine-tuning story; if it dies, the paper is a solid contribution.
Minor soft spots, in proportion: the OMY results have no error bars over 50 trials per task; the DA3 pseudo-depth supervision quality is untested; and everything runs on one backbone. The self-cited SPACE-CLIP decoder comparison has author overlap, but the paper is careful to frame it as decoder-config evidence, not action gains, so I would not weight that heavily.
Who this is for: robotics VLA researchers working on spatial reasoning and efficient retrofitting of frozen backbones. It deserves a serious referee — the mechanism is novel enough, the evaluation is mostly solid, and the blocking issue is fixable with one experiment. I would not cite it until that control exists; right now the mechanism claim is unproven.
The strongest objection
The central claim is that depth-supervised geometric features, decoded from the frozen VLA visual encoder and injected as a gated residual, drive the reported LIBERO and OMY gains over SmolVLA. The only direct evidence for this mechanism is the same-checkpoint residual intervention (Table IV): disabling the residual at inference changes overall success from 70.90% to 69.85% (+1.05 pp), with negative per-suite effects on Object (−1.20) and Goal (−0.60). The authors explicitly state that this intervention does not isolate the contribution of geometry to policy learning (b179) and call for a separately trained no-geometry control. That control is absent. Consequently, the 10.9-pp LIBERO and 20-pp OMY improvements cannot be attributed to geometry rather than to generic effects of the additional residual parameters, the altered optimization trajectory, or the action-side projections trained jointly in Stage II. The paper's own residual ablation, being an inference-time probe of the same checkpoint, retains all training-time influences and therefore cannot distinguish these alternatives. While the paper is transparent about this limitation, the load-bearing claim—that geometry is the operative mechanism—remains unsupported by the presented evidence.
The reader's weakest assumption correctly identifies the central gap: the absence of a no-geometry control leaves the attribution of the headline gains to depth-supervised geometry unsupported. The residual ablation measures only inference-time sensitivity, not the training-time effect of geometry, and its small and mixed effect highlights that the mechanism is not established. This is precisely the sort of control that would move the paper from CONDITIONAL to ACCEPT. The paper is otherwise well-specified with matched evaluation conditions, and the depth-prediction results provide supporting evidence that geometric information is accessible from the frozen encoder, but that does not bridge the gap to action conditioning. The proposed concrete test—a random-initialized frozen bridge/decoder control—directly isolates the geometry signal. I agree with the reader's verdict of CONDITIONAL and recommend no change. The concern is load-bearing because the paper's main claim is mechanistic; without the control, the reported improvements could plausibly stem from extra trainable parameters or optimization effects rather than geometric content.
The referee's report
If the attribution holds, the core idea is valuable: a frozen VLA visual hierarchy may contain recoverable geometric signal that can be retrofitted into action conditioning without a second image encoder, extra tokens, or depth at deployment. The paper is strong on several fronts. The two-stage protocol is precisely specified; evaluation conditions are matched between methods for both LIBERO and OMY; the real-robot evaluation (200 matched trials across four tasks) is a genuine asset rare in this literature; the shared-encoder depth numbers are plausible; and the authors are unusually transparent, explicitly stating (b178-b179) that their residual intervention does not isolate geometry's training-time contribution and that a separately trained no-geometry control is required. The obstacle to publication is not presentation but one load-bearing, unresolved attribution gap: the headline gains are attributed to depth-supervised geometry even though the only directly presented evidence for that mechanism is an inference-time intervention showing a roughly one-point effect with mixed per-suite signs.
Pith's reading of the paper (a machine's summary, not the authors')
This paper tries to establish that the frozen visual encoder of a pretrained vision-language-action model already holds recoverable geometric signal, and that this signal, decoded with depth supervision and re-injected as a gated residual into the VLA's own visual tokens, improves closed-loop manipulation. The claim matters because it offers a retrofit path: geometry is added without a second image encoder, without extra visual tokens, and without depth observations at deployment, so the deployed policy sees only RGB, robot state, and language. Under matched conditions the full pipeline lifts LIBERO success from 60.0% to 70.9% and physical-robot success on ROBOTIS OMY from 54.0% to 74.0%. The paper's own control, disabling the residual in the same trained checkpoint, moves success only from 70.90% to 69.85%, so what the geometric pathway itself contributes during training remains open.
What carries the argument. The load-bearing mechanism is the gated residual interface of Eq. (3): the depth-supervised feature map $g$ is pooled to the native patch grid, projected by a $1\times 1$ convolution $R_\omega$, scaled by a learned gate $\alpha = \sigma(\eta)$ initialized near $0.002$, and added to the native visual tokens as $\tilde{e}_z = z + \Delta z$. Because the readout is initialized to zero, Stage-II optimization begins exactly at the native visual representation, and the residual respects the existing conditioning pathway without adding tokens. The geometric side is a K4-based geometry decoder that aggregates features from four fixed sampling positions with content-dependent weights, preserving local spatial structure; a feature bridge aligns the frozen encoder's layer outputs to that decoder. The two-stage split — geometry acquisition under depth supervision, then geometry-to-action transfer under a flow-matching objective with an explicit residual penalty — is what separates the roles the mechanism claims to play.
Core claim. The central claim is that a frozen VLA hierarchy can be made to support precise spatial reasoning by training a depth-supervised readout on its intermediate features and then routing that readout into the action policy through the native visual-token stream. Stage I trains a feature bridge and a K4-based geometry decoder against depth targets — simulator ground truth for LIBERO, teacher pseudo-depth for the physical robot — while the visual encoder stays frozen. Stage II freezes those modules and trains a gated residual interface that adds a geometrically derived perturbation to the native visual tokens, together with the action-side projections and the action expert. The deployed policy consumes only RGB, robot state, and language. If the results hold, the paper establishes that recoverable geometric signal sits inside a frozen VLA's visual hierarchy and that injecting it improves manipulation success, while the same-checkpoint residual ablation leaves exactly how much of the gain the geometric signal itself carries to future work.
The premise it rests on. The benchmark gains are caused by the depth-supervised geometric signal rather than by the jointly trained action pathway, yet the paper's own same-checkpoint residual ablation shows only about a one-point inference-time effect with mixed per-suite signs, and the authors acknowledge that no separately trained no-geometry control exists to distinguish the two.
If the paper is right (Pith's inferences, not the authors')
Depth-supervised geometric features can be decoded from a frozen VLA visual encoder with competitive quality: AbsRel 0.030, RMSE 0.1316 m, and $\delta_1$ 0.975 on the LIBERO front-view validation set.
Geometry can be added to a VLA policy without a second image encoder, additional visual tokens, or depth at deployment, keeping the native token count intact.
The retrofit improves all four LIBERO suites and all four physical OMY tasks, with the largest gains where spatial precision matters: Object +20.8 points, Long +14.4, and block stacking +24.0.
The same-checkpoint residual ablation shows the geometric-to-action pathway helps Spatial and Long but slightly hurts Object and Goal, so the benefit is task-dependent rather than uniform.
The method is demonstrated only on the SmolVLA backbone and a single robot platform, so the retrofit's portability to other VLA architectures is untested.
The gap between the 10.9-point pipeline gain and the 1.05-point inference-time residual effect suggests most of the improvement arises from joint Stage-II training rather than from the depth-supervised readout itself; a separately trained no-geometry control, which the authors note is missing, would settle which mechanism drives the gain.
The physical-robot result bundles the retrofit with a teacher-generated pseudo-depth target, so part of the 20-point OMY gain could be an artifact of that supervision's quality rather than of the method; sensitivity of the approach to the teacher is untested.
If the retrofit holds across backbones, it implies a cheap upgrade path for deployed VLA policies — tuning a depth readout and residual interface on existing frozen weights instead of re-pretraining — a direction the paper leaves open and that a reader could test directly.
The mixed per-suite signs of the residual ablation hint that geometry benefits tasks with explicit spatial relations and long horizons, while semantic-heavy tasks may not need it; identifying that boundary is a natural next experiment.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Vision-language-action (VLA) models encode semantic information from vision-language pretraining, but manipulation also requires precise spatial reasoning. We present GeoBridge-VLA, a two-stage method for learning geometric features from a pretrained VLA's frozen visual encoder and using them for action prediction. Stage I trains a feature bridge and geometry decoder with depth supervision. Stage II freezes these modules and trains a gated residual interface together with the action-side projections and action expert. The residual augments the existing visual tokens without adding a second image encoder or increasing the token count. Deployment requires RGB, robot state, and language, but no depth observations. Under matched evaluation conditions, GeoBridge-VLA achieves 70.9% success on LIBERO, compared with 60.0% for SmolVLA. Disabling the residual in the same trained checkpoint reduces success from 70.90% to 69.85%, with mixed effects across suites. On a physical ROBOTIS OMY robot, GeoBridge-VLA succeeds in 148 of 200 trials (74.0%) across four tasks, compared with 108 of 200 (54.0%) for SmolVLA.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-06
The authors' main claim, as the classifier picked it out 0.66
“Under matched evaluation conditions, GeoBridge-VLA achieves 70.9% success on LIBERO, compared with 60.0% for SmolVLA.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.71 · human participants 0.03 · controlled comparison 0.96 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Vision-language-action (VLA) models encode semantic information from vision-language pretraining, but manipulation also requires precise spatial reasoning. | background 0.89 | 0.07 | 0.04 | asserted 1.00 |
| We present GeoBridge-VLA, a two-stage method for learning geometric features from a pretrained VLA's frozen visual encoder and using them for action prediction. | contribution 0.96 | 0.91 | 0.12 | asserted 1.00 |
| Stage I trains a feature bridge and geometry decoder with depth supervision. | method 1.00 | 0.91 | 0.03 | asserted 0.99 |
| Stage II freezes these modules and trains a gated residual interface together with the action-side projections and action expert. | method 1.00 | 0.89 | 0.04 | asserted 0.95 |
| The residual augments the existing visual tokens without adding a second image encoder or increasing the token count. | method 0.92 | 0.79 | 0.06 | asserted 1.00 |
| Deployment requires RGB, robot state, and language, but no depth observations. | method 0.87 | 0.64 | 0.04 | asserted 1.00 |
| Under matched evaluation conditions, GeoBridge-VLA achieves 70.9% success on LIBERO, compared with 60.0% for SmolVLA. | result 1.00 | 0.95 | 0.40 | asserted 1.00 |
| Disabling the residual in the same trained checkpoint reduces success from 70.90% to 69.85%, with mixed effects across suites. | result 0.98 | 0.95 | 0.10 | asserted 1.00 |
| On a physical ROBOTIS OMY robot, GeoBridge-VLA succeeds in 148 of 200 trials (74.0%) across four tasks, compared with 108 of 200 (54.0%) for SmolVLA. | result 1.00 | 0.95 | 0.23 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 1.00
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-06: 1 request, 6,889 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (7)
Showing 6 of the paper's 7 figures. Figures are the authors', reproduced from the paper.
References (32)
- [2] M. J. Kim, K. Pertsch, S. Karamchetiet al., “OpenVLA: An open-source vision-language-action model,” arXiv:2406.09246, 2024. [Online]. Available: https://arxiv.org/abs/2406.09246 on PithPith's reviewarXiv
- [3] K. Black, N. Brown, D. Driesset al., “π 0: A vision-language-action flow model for general robot control,” arXiv:2410.24164, 2024. [Online]. Available: https://arxiv.org/abs/2410.24164 on PithPith's reviewarXiv
- [4] M. Shukor, D. Aubakirova, F. Capuanoet al., “SmolVLA: A vision-language-action model for affordable and efficient robotics,” arXiv:2506.01844, 2025 on PithPith's reviewarXiv
- [5] D. Qu, H. Song, Q. Chenet al., “SpatialVLA: Exploring spatial representations for visual-language-action model,” arXiv:2501.15830, 2025 on PithPith's reviewarXiv
- [6] 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:2510.13375, 2025 on PitharXiv
- [7] Z. Zhang, H. Li, Y . Daiet al., “From spatial to actions: Grounding vision-language-action model in spatial foundation priors,” inInternational Conference on Learning Representations, 2026. [Online]. Available: https://arxiv.org/abs/2510.17439 on PitharXiv
- [8] Y . Peng, Y . Zhao, A. Habuda, K. Pham, Y . Zhu, T. N. Le, F. Abu-Dakka, and L. Guo, “G 3VLA: Geometric inductive bias for vision-language-action models,” arXiv:2606.24472, 2026. [Online]. Available: https://arxiv.org/abs/2606.24472 on PithPith's reviewarXiv
- [9] Y . Zheng, X. Li, S. Guet al., “GIFT: Guided intermediate feature training via action-oriented structural supervision for robotic manipulation,” arXiv:2609.04193, 2026. [Online]. Available: https://arxiv.org/abs/2609.04193 on PithPith's reviewarXiv
- [10] R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision transformers for dense prediction,” arXiv:2103.13413, 2021 on PithPith's reviewarXiv
- [11] S. F. Bhat, R. Birkl, D. Wofk, P. Wonka, and M. Müller, “ZoeDepth: Zero-shot transfer by combining relative and metric depth,” arXiv:2302.12288, 2023 on PithPith's reviewarXiv
- [12] L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,” arXiv:2406.09414, 2024 on PithPith's reviewarXiv
- [13] H. Lin, S. Chen, J. Liew, D. Y . Chen, Z. Li, G. Shi, J. Feng, and B. Kang, “Depth Anything 3: Recovering the visual space from any views,” arXiv:2511.10647, 2025. [Online]. Available: https://arxiv.org/abs/2511.10647 on PithPith's reviewarXiv
- [14] R. Zhang, Z. Zeng, Z. Guo, and Y . Li, “Can language understand depth?” arXiv:2207.01077, 2022 on PithPith's reviewarXiv
- [15] T. Cho, T. Kim, and A. J. Choi, “SPACE-CLIP: Spatial perception via adaptive CLIP embeddings for monocular depth estimation,” arXiv:2601.17657, 2026 on PithPith's reviewarXiv
- [16] M. J. Kim, C. Finn, and P. Liang, “Fine-tuning vision-language-action models: Optimizing speed and success,” arXiv:2502.19645, 2025 on PithPith's reviewarXiv
- [17] J. Wen, Y . Zhu, J. Li, M. Zhu, K. Wu, Z. Xu, N. Liu, R. Cheng, C. Shen, Y . Peng, F. Feng, and J. Tang, “TinyVLA: Towards fast, data-efficient vision-language-action models for robotic manipulation,” arXiv:2409.12514, 2024 on PithPith's reviewarXiv
- [18] H. Zhen, X. Qiu, P. Chen, J. Yang, X. Yan, Y . Du, Y . Hong, and C. Gan, “3D-VLA: A 3d vision-language-action generative world model,” arXiv:2403.09631, 2024 on PithPith's reviewarXiv
- [19] R. Yang, G. Chen, C. Wen, and Y . Gao, “FP3: A 3d foundation policy for robotic manipulation,” arXiv:2503.08950, 2025 on PithPith's reviewarXiv
- [20] C. Li, J. Wen, Y . Peng, Y . Peng, F. Feng, and Y . Zhu, “PointVLA: Injecting the 3d world into vision-language-action mod- els,” arXiv:2503.07511, 2025 on PithPith's reviewarXiv
- [21] V . Bhat, Y .-H. Lan, P. Krishnamurthy, R. Karri, and F. Khorrami, “3D CA VLA: Leveraging depth and 3d context to generalize vision language action models for unseen tasks,” arXiv:2505.05800, 2025 on PitharXiv
show 12 more
- [22] J. Park, K. Ranasinghe, J. Jang, C. Mata, Y . S. Jang, and M. S. Ryoo, “IVRA: Improving visual-token relations for robot action policy with training-free hint-based guidance,” arXiv:2601.16207, 2026 on PitharXiv
- [23] J. Koo, T. Cho, H. Kang, E. Pyo, T. G. Oh, T. Kim, and A. J. Choi, “Retovla: Reusing register tokens for spatial reasoning in vision- language-action models,” in2026 IEEE International Conference on Robotics and Automation (ICRA), 2026, pp. 2624–2631 on Pith
- [24] Q. Qian, G. Zhao, G. Zhang, J. Wang, R. Xu, J. Gao, and D. Zhao, “GP3: A 3d geometry-aware policy with multi-view images for robotic manipulation,” arXiv:2509.15733, 2025 on PitharXiv
- [25] Z. Rao, W. Chen, L. Xie, X. Hua, D. Yin, Z. Tian, and F. R. Yu, “AugVLA-3D: Depth-driven feature augmentation for vision- language-action models,” arXiv:2602.10698, 2026 on PithPith's reviewarXiv
- [26] Y . Li, Y . Chen, M. Zhou, H. Li, Z. Zhang, and D. Zhao, “QDepth- VLA: Quantized depth prediction as auxiliary supervision for vision- language-action models,” arXiv:2510.14836, 2025 on PithPith's reviewarXiv
- [27] T. Lin, Y . Du, J. Liu, N. Zhu, Y . Li, Y . Fu, Y . Chen, H. Cai, Z. Ye, B. Cheng, K. Ye, Y . Mao, Y . Zhong, M. Dong, J. Yan, G. Li, and B. Zhao, “Evo-Depth: A lightweight depth-enhanced vision-language- action model,” arXiv:2605.14950, 2026 on PithPith's reviewarXiv
- [28] M. S. Sarowar, M. T. Islam, S. Kim, and S. Ahn, “GaussVLA: Geometry-aware spatial reasoning for vision-language-action model,” arXiv:2608.24959, 2026 on PithPith's reviewarXiv
- [29] D. Eigen, C. Puhrsch, and R. Fergus, “Depth map prediction from a single image using a multi-scale deep network,” inAdvances in Neural Information Processing Systems, vol. 27, 2014. [Online]. Available: https://arxiv.org/abs/1406.2283 on PithPith's reviewarXiv
- [30] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,”IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004 on Pith
- [31] B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone, “LIBERO: Benchmarking knowledge transfer for lifelong robot learn- ing,” arXiv:2306.03310, 2023 on PithPith's reviewarXiv
- [32] Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” arXiv:2210.02747, 2022. APPENDIX SPACE-CLIPV2 GEOMETRYDECODER This appendix describes the decoder internals and stan- dalone depth results. The main method concerns connectin... on PithPith's reviewarXiv
- [2023] [Online]. Available: https://arxiv.org/abs/2307.15818 on PithPith's reviewarXiv
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment