acceptodds
Under review as a conference paper at ICLR 2027

GeoLACE: Geometric Layerwise Action-Conditioned Enhancement for Vision-Language-Action Models

Abstract

Generalist vision–language–action (VLA) models benefit from vision–language pretraining but have limited geometric understanding for precise manipulation. Promising approaches use geometric foundation models (GFMs) to provide VLA policies with geometric priors from RGB observations. However, these methods often lack a systematic layer-selection strategy that jointly considers geometric complementarity and relevance to action prediction. This can lead to redundant or underutilized geometric features, limiting gains in manipulation performance. Incorporating geometric knowledge can also disrupt pretrained VLA representations, potentially compromising vision–language capabilities. To address these challenges, we propose GeoLACE, a plug-and-play framework that selects complementary GFM layers and action-relevant VLA layers. It performs layerwise geometric fusion using multi-level features extracted from current RGB observations. We further introduce a two-stage training strategy that enables the policy to better exploit these priors while preserving its pretrained capabilities. We integrate GeoLACE into representative VLA backbones with different architectures. Experiments on LIBERO, RoboTwin, and two distinct real-world bimanual platforms demonstrate improved manipulation performance, with particularly substantial gains in real-world manipulation. GQA evaluation further supports better retention of pretrained vision–language capabilities than the compared methods.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Geometric Foundation Models, Action-Conditioned Layer Selection, Layerwise Feature Fusion, Pretrained Capability Retention

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7UT9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeoLACE, a framework that integrates geometric features from pretrained geometric foundation models into VLA policies. The method introduces ActSelect to identify complementary geometric features and action-relevant VLA layers, followed by layer-wise feature fusion and a two-stage training strategy to preserve pretrained vision-language capabilities. Experiments on LIBERO, RoboTwin, and real-world robotic systems demonstrate improvements in manipulation performance.

Overall, it presents a well-motivated approach to incorporating geometric priors into VLA models, supported by extensive simulation and real-world experiments. Therefore, I support the acceptance of this paper.

Critical strengths
  1. The proposed action-conditioned layer selection provides a systematic way to integrate geometric features into VLA models, and demonstrates improvements across multiple benchmarks and two real-world robotic platforms, supported by ablation studies on layer selection, feature fusion, and training strategies.

  2. The paper evaluates the preservation of pretrained vision-language capabilities through GQA experiments, addressing an important concern when incorporating geometric features into VLA models.

Critical weaknesses
  1. The method relies on online VGGT features and uses feature caching to reduce inference latency. However, it remains unclear how cached geometric features affect performance in dynamic scenes where objects or robot viewpoints change rapidly.
Questions for authors
  1. Could the authors compare the proposed CKA-based GFM source-layer selection with simpler strategies, such as random or uniformly spaced layer selection, under the same training settings?

  2. How sensitive is the proposed method to the geometric feature refresh frequency during real-world execution? Could the authors further analyze the trade-off between inference efficiency and manipulation performance?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 7UT9 (1/2)

Thank you for your careful review and constructive suggestions. We appreciate your recognition of our systematic layer-selection approach, the simulation and real-world evaluations, and the assessment of pretrained vision–language capability preservation. Below, we address your concern and questions in the order raised and indicate the corresponding manuscript revisions.

[W1] The method relies on online VGGT features and uses feature caching to reduce inference latency. However, it remains unclear how cached geometric features affect performance in dynamic scenes where objects or robot viewpoints change rapidly.

Response: The original submission describes visual-change-triggered updates to reduce temporal mismatch; we now test the default feature-reuse configuration under dynamic perturbations. Specifically, GeoLACE achieves 70.6% overall success in Setting A and 60.0% in Setting B, with modest decreases from the original conditions while matching or exceeding π₀.₅ on all five tasks. Visual-change-triggered refresh ratios increase on every task, consistent with geometric updates responding to scene changes.

Evidence in the original submission. This aspect of feature reuse was partially discussed in Appendix D.4 of the original manuscript: the visual-change criterion compares current observations with those associated with cached geometry and can trigger an earlier update. Waiting for an ongoing refresh and rechecking the updated cache before synchronous recomputation are designed to reduce temporal mismatch.

Additional dynamic-scene experiments. We have added real-world dynamic-scene experiments and analysis in Appendix D.4 (Table 16), with a corresponding reference in Sec. 4.4. We evaluate GeoLACE using its default feature-reuse configuration. Both π₀.₅ and GeoLACE are evaluated over 40 trials per task under three types of perturbations:

  • Object position changes: In starfruit handover and rag pickup, we rapidly reposition the object as the gripper approaches it.
  • Object orientation changes: In phone placement and cable organization, we change the object’s orientation as the gripper approaches it.
  • Scene content changes: In item packing, we introduce additional objects onto the table after the initially presented items have been packed.

The last two columns report visual-change-triggered refreshes as a percentage of total VLA policy requests for GeoLACE under the original and dynamic conditions.

TaskDynamic success: π₀.₅Dynamic success: GeoLACERefresh ratio: Original (%)Refresh ratio: Dynamic (%)
Starfruit handover20/4034/4023.4125.01
Rag pickup36/4036/4025.1027.33
Item packing19/4029/4027.5129.52
Phone placement9/4014/4019.3420.70
Overall (Setting A)84/160 (52.5%)113/160 (70.6%)——
Cable organization (Setting B)10/4024/4020.4322.85

Despite modest decreases relative to the original conditions in Table 4, GeoLACE maintains an overall advantage over the baseline in both settings. This advantage extends across object repositioning, rotation, and the introduction of additional objects, demonstrating effective manipulation when the scene changes during execution.

The visual-change-triggered refresh ratio increases on every task, indicating that geometric updates are triggered more often per policy request under dynamic conditions. The refresh criterion compares current observations with the reference observations associated with the cached geometry, tying update decisions to changes relative to the geometric features being reused. This enables updates in response to object motion and newly introduced scene content, helping reduce temporal mismatch during feature reuse. Together, the task results and observed refresh behavior support the effectiveness of the default feature-reuse configuration under the tested dynamic scene changes.

AuthorsOpenReview ↗

Response to Reviewer 7UT9 (2/2)

[Q1] Could the authors compare the proposed CKA-based GFM source-layer selection with simpler strategies, such as random or uniformly spaced layer selection, under the same training settings?

Response: Yes. We have added uniform and random source-selection comparisons. Specifically, under the same branch count, VLA mapping, training data, training budget, and Safe-Add schedule, CKA-based selection achieves 98.1% average LIBERO success, compared with 97.4% for uniform selection and 97.2% for random selection, and performs better on all four suites.

Additional controlled comparison. We compare the three source-selection strategies while keeping the three fusion branches, VLA mapping (1, 7, 9), training data, training budget, and Safe-Add schedule fixed.

Source selectionSource layersAverage LIBERO success (%)
Uniform(6, 12, 18)97.4
Random(4, 10, 12)97.2
CKA-based(8, 16, 21)98.1

Analysis. The CKA analysis identifies three representational stages: layers 0–13, 14–17, and 18–23. Uniform selection omits the middle stage, while random selection remains entirely within the first stage. CKA-based coverage of all three stages provides a mechanism consistent with the observed downstream gains; equal spacing in layer indices does not ensure such coverage.

We have added the complete per-suite results and this analysis to Appendix A.1 (GFM Source Selection), Table 5.

[Q2] 2.1 How sensitive is the proposed method to the geometric feature refresh frequency during real-world execution?

Response: Our additional comparison shows a modest success increase from 74.4% with asynchronous 2 Hz refresh to 76.9% with synchronous encoding targeting 10 Hz, while mean latency rises from 90.8 to 169.1 ms per action chunk.

Additional experiment. To examine sensitivity to geometric feature updates during real-world execution, we compare our default configuration, which reuses features with scheduled asynchronous 2 Hz VGGT refresh, against synchronous encoding targeting 10 Hz. In the synchronous configuration, every policy request waits for VGGT features computed from the current observations before VLA inference proceeds. In the default configuration, policy requests arrive at approximately 10 Hz and reuse geometric features refreshed asynchronously at a scheduled rate of 2 Hz, with visual changes able to trigger earlier updates. The VLA continues to receive current RGB observations and proprioception. Each configuration is evaluated over 40 trials on each of the four Setting A tasks.

TaskAsync. (2 Hz) success ↑Sync. (10 Hz) success ↑Async. (2 Hz) latency (ms) ↓Sync. (10 Hz) latency (ms) ↓
Starfruit handover31/4033/40104.3167.2
Rag pickup37/4036/4081.7168.7
Item packing32/4033/4086.7171.4
Phone placement19/4021/4090.6168.9
Overall74.4%76.9%90.8169.1

[Q2] 2.2 Could the authors further analyze the trade-off between inference efficiency and manipulation performance?

Response: Yes. Our additional comparison with synchronous VGGT encoding targeting 10 Hz quantifies the efficiency–performance trade-off: asynchronous 2 Hz refresh reduces mean latency by 46.3%, with an observed decrease of 2.5 percentage points in success. This measured trade-off motivates our default asynchronous configuration.

Additional analysis. Across these two evaluated configurations, computing fresh geometry for every policy request provides a modest improvement in manipulation success but incurs substantially higher latency. Feature reuse avoids repeated geometric encoding, while asynchronous refresh reduces blocking; visual-change-triggered updates allow earlier refreshes when the scene changes.

We have reported this comparison and the reason for choosing the default configuration in Appendix D.4 (Table 15), together with the execution and visual-change-triggered refresh protocols. Reported latency includes all recorded requests, including cache initialization and requests involving geometric refreshes.

Reviewer_hsdjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoLACE, a framework that integrates multi-layer features from a frozen geometric foundation model (VGGT) into pretrained Vision-Language-Action (VLA) policies. It introduces an offline layer-pair selection procedure (ActSelect) combining inter-layer CKA clustering with action-loss gate-gradient probes, injects the selected VGGT features into VLA visual tokens via zero-initialized residual MLPs, and applies a two-stage training schedule (Safe-Add) that penalizes large residual norms in the final training phase. Experiments with pi0.5 and GR00T N1.6 on LIBERO, RoboTwin 2.0, GQA, and two bimanual robot setups report improvements over native 2D policies and training-time alignment baselines.

Critical strengths
  1. Enhancing pretrained VLA policies with multi-level 3D geometric priors without degrading their pretrained vision-language capabilities is a practical problem, and the method is validated on two mainstream flow-matching VLA backbones (π0.5\pi _ {0.5} and GR00T N1.6).
  2. The empirical evaluation covers two simulation benchmarks (LIBERO and RoboTwin 2.0), five real-world bimanual manipulation tasks across two hardware setups, and zero-shot GQA capability retention checks.
Critical weaknesses
  1. The offline ActSelect procedure feels overly complex relative to its empirical gain. In Table 2, picking three VLA layers at random already reaches 97.3%–97.5% on LIBERO versus 98.1% for ActSelect (with Random 2 and two-layer L17 reaching 99.2% vs. 98.8% on Spatial), and in Appendix Table 6 the fourth-ranked mapping (0, 8, 9) yields a larger action-loss reduction (ΔLact=6.28×10−4\Delta \mathcal{L} _ {\mathrm{act}} = 6.28 \times 10^{-4}) than the selected (1, 7, 9) (ΔLact=5.31×10−4\Delta \mathcal{L} _ {\mathrm{act}} = 5.31 \times 10^{-4}).
  2. The motivation for the two-stage Safe-Add scheme is not fully convincing. In Table 3(a), one-stage training already achieves 97.5%, whereas the 20k+10k Safe-Add schedule drops to 96.4% (below native π0.5\pi _ {0.5} at 96.9%), and Figure 6 omits the one-stage variant on GQA, leaving it unclear whether VLM retention comes from Stage II's τk\tau _ k penalty or simply from avoiding an explicit alignment loss (Lalign\mathcal{L} _ {\mathrm{align}}).
  3. Unlike training-only alignment methods (SF [1] and ROCKET [2]), GeoLACE runs VGGT online at test time, yet omits online VGGT-injection baselines (Evo-0 [3] and 3D-Mix [4]) across all benchmarks. Moreover, Appendix D.4 reveals that the +9.4 ms latency in Table 4 relies on reusing cached VGGT features at 2 Hz for a 10 Hz policy, which conflicts with preserving online patch correspondence as the robot moves.
  4. From a presentation standpoint, Sections 3 and 4 introduce many intricate procedures while deferring essential definitions to the appendix, including how raw gate gradients qkℓq _ {k\ell} in Equation (2) are normalized into q~kℓ(p)\tilde{q} _ {k\ell}^{(p)}, the formula for τk\tau _ k, the sampling probabilities for α\alpha, and the 2 Hz feature caching behind Table 4.

[1] Li et al. 2026. Spatial Forcing: Implicit Spatial Representation Alignment for Vision-Language-Action Model. ICLR 2026.

[2] Sun et al. 2026. ROCKET: Residual-Oriented Multi-Layer Alignment for Spatially-Aware Vision-Language-Action Models. arXiv:2602.17951.

[3] Lin et al. 2025. Evo-0: Vision-Language-Action Model with Implicit Spatial Understanding. arXiv:2507.00416.

[4] Yu et al. 2026. 3D-Mix for VLA: A Plug-and-Play Module for Integrating VGGT-Based 3D Information into Vision-Language-Action Models. arXiv:2603.24393.

Questions for authors
  1. Why does the fourth-ranked mapping (0, 8, 9) achieve a larger action-loss reduction than (1, 7, 9) in Appendix Table 6, and how does (0, 8, 9) perform after full fine-tuning on LIBERO?
  2. How does one-stage GeoLACE (97.5% in Table 3(a)) perform on the GQA benchmark in Figure 6, and why does the 20k+10k Safe-Add schedule drop LIBERO success to 96.4%?
  3. What is the real-world latency and success rate in Table 4 when VGGT is run synchronously at 10 Hz rather than cached at 2 Hz, and how does GeoLACE compare to online injection baselines like Evo-0 [3] or 3D-Mix [4]?
7 comments
AuthorsOpenReview ↗

Response to Reviewer hsdj (1/5)

Thank you for your careful review and detailed suggestions. We appreciate your recognition of the practical problem addressed by GeoLACE and its evaluation across two policy backbones, simulation benchmarks, and real-world robotic systems. We respond to each concern below with the corresponding comparisons, additional evaluations, and manuscript clarifications, grouping related weaknesses and questions.

[W1&Q1] 1.1 The offline ActSelect procedure feels overly complex relative to its empirical gain.

Response: ActSelect scores candidate pairings using the native action loss with frozen backbones and shared temporary projectors, avoiding full policy fine-tuning for each candidate. Complete selection costs only 0.82 H20 GPU-hours, just 0.98% of a full 30k-step fine-tuning run’s compute cost.

Under matched settings, ActSelect achieves 98.1% average LIBERO success, outperforming random target selection (97.3%–97.5%) and the manually specified SpatialStack-style mapping (97.1%).

Beyond the near-ceiling LIBERO results, GeoLACE achieves larger gains over π₀.₅, raising average RoboTwin success by 7.5 and 11.0 percentage points under easy and hard conditions and improving success on all five real-world tasks across two platforms.

Evidence in the original submission. With three GFM sources fixed, ActSelect ranks all (183)=816\binom{18}{3}=816 monotonic VLA target triples by their summed source–target LCB scores (Sec. 3.2; Appendix A.2). Sec. 4.2 compares ActSelect with three random target-layer configurations under matched settings, showing higher average LIBERO success for the selected mapping. The original experiments also cover six RoboTwin tasks under easy and randomized hard conditions and five real-world tasks across two platforms, providing evidence beyond near-ceiling LIBERO performance.

Computational cost. Based on the recorded runs, five-probe ActSelect takes 0.746 GPU-hours and three-seed CKA takes 0.077 GPU-hours on a single NVIDIA H20-3e GPU, totaling 0.823 GPU-hours. A full 30k-step policy fine-tuning run requires 83.82 GPU-hours on four GPUs of the same type. The complete offline selection pipeline therefore adds approximately 0.98% to the compute cost of one full fine-tuning run.

Analysis. Random target selection and the manually specified SpatialStack-style mapping do not evaluate source–target correspondences using the action objective. ActSelect combines complementary sources identified by CKA with targets chosen through normalized action-loss utility and an LCB correction for calibration uncertainty. Random 2 and L17 score higher on Spatial (99.2% vs. 98.8%), but ActSelect achieves higher average success across the four suites (98.1% vs. 97.3% and 97.5%), supporting the benefit of this action-oriented selection.

RoboTwin and real-world manipulation gains. GeoLACE improves average RoboTwin success by 7.5 and 11.0 percentage points under easy and randomized hard conditions. On real robots, success increases from 54.4% to 74.4% in Setting A and from 32.5% to 65.0% in Setting B, complementing the matched layer-selection evidence with broader manipulation gains.

AuthorsOpenReview ↗

Response to Reviewer hsdj (2/5)

[W1&Q1] 1.2 Why does the fourth-ranked mapping (0, 8, 9) achieve a larger action-loss reduction than (1, 7, 9) in Appendix Table 6?

Response: Thank you for raising this question. We use the LCB score as our layer-selection criterion, while joint action-loss reduction serves as an auxiliary measure to verify the effectiveness of the selected layers. Their rankings need not coincide because joint verification additionally involves residual RMS rescaling and cross-layer interactions. The joint action-loss measurement in Table 9 evaluates the candidate mappings with the VLA and temporary probe projectors frozen before full fine-tuning.

Evidence in the original submission. Appendix B.3 states that the full LCB ranking determines the mapping, while finite-scale joint verification checks its action-loss effect. Nonlinear cross-layer interactions can cause the joint ranking to differ from the additive first-order ranking.

LCB-based selection. ActSelect computes normalized first-order action-loss utilities for source–target pairings on the full calibration set. After averaging across five probes, it combines mean utility with an LCB correction for uncertainty across calibration evaluations and selects the mapping with the highest summed score. Accordingly, (1, 7, 9) is selected with a score of 1.8519, while (0, 8, 9) ranks fourth with 1.7733. We have clarified this selection criterion in Sec. 3.2.

Joint action-loss verification (Table 9). With the VLA and temporary probe projectors frozen, the three residuals are jointly injected at controlled RMS ratios on a 10-batch subset before full fine-tuning. Both mappings reduce action loss across all five probes and tested scales, verifying their combined contribution to action prediction. RMS rescaling and cross-layer interactions affect these finite-scale joint gains, whereas LCB aggregates normalized first-order utilities with an uncertainty correction. Their orderings can therefore differ while providing complementary evidence for action-relevant geometric fusion. The temporary probe projectors are discarded before two-stage fusion training. We have expanded this discussion in Appendix B.3 and clarified the verification setting in the caption of Table 9.

[W1&Q1] 1.3 How does (0, 8, 9) perform after full fine-tuning on LIBERO?

Response: To directly evaluate the downstream effectiveness of the LCB-selected mapping, we conducted full policy fine-tuning with the (0, 8, 9) mapping and ran 2,000 LIBERO rollouts, with 500 per suite. The mapping achieves 97.8% average LIBERO success, compared with 98.1% for the LCB-selected mapping (1, 7, 9).

Additional fine-tuning and analysis. We conducted full policy fine-tuning with the (0, 8, 9) mapping and ran 2,000 LIBERO rollouts, with 500 per suite. The average success comparison and corresponding analysis have been added to Appendix B.3 (Table 10).

These results show that the larger joint action-loss reduction achieved by (0, 8, 9) in frozen-model verification does not necessarily translate into better policy performance after full fine-tuning. The higher average success of the LCB-selected mapping further supports the selection criterion’s effectiveness in the evaluated setting.

MappingSpatial (%)Object (%)Goal (%)Long (%)Average (%)
(1, 7, 9)98.899.099.295.298.1
(0, 8, 9)98.498.898.695.497.8
AuthorsOpenReview ↗

Response to Reviewer hsdj (3/5)

[W2&Q2] 2.1 How does one-stage GeoLACE perform on GQA?

Response: We additionally evaluated the one-stage GeoLACE model (Ours-I), trained with Stage I alone for 30k steps, on GQA. Ours-I achieves 84.71% retention, 60.70% full-answer accuracy, and 0.0377 JS divergence. In comparison, GeoLACE with the 25k+5k two-stage schedule achieves 87.40%, 62.80%, and 0.0325, respectively, performing better on all three metrics.

Additional GQA evaluation of Ours-I. Ours-I uses one-stage training for 30k steps, corresponding to the 97.5% LIBERO success rate in Table 3(a). We evaluate it on the same 4,000 held-out GQA questions without further parameter updates and retain the original GQA results for the other models. We have added Ours-I to Figure 6 and included the corresponding analysis in Sec. 4.3.

MethodRetention (%) ↑Acquisition (%) ↑Full-answer accuracy (%) ↑JS divergence ↓
π₀.₅-FT86.60‾\underline{86.60}13.2061.92‾\underline{61.92}0.0304
SF83.6016.0060.820.0437
ROCKET84.1712.2261.480.0403
Ours-I: Stage I 30k84.7113.7360.700.0377
GeoLACE: Stage I 25k + Stage II 5k87.4014.00‾\underline{14.00}62.800.0325‾\underline{0.0325}

Bold and underlined values indicate the best and second-best results in each column, respectively.

[W2&Q2] 2.2 The motivation for the two-stage Safe-Add scheme is not fully convincing. Figure 6 omits the one-stage variant on GQA, leaving it unclear whether VLM retention comes from Stage II’s Lpen\mathcal{L}_{\text{pen}} penalty or simply from avoiding an explicit alignment loss (Lalign\mathcal{L}_{\text{align}}).

Response: Safe-Add is designed to preserve the VLA’s pretrained representations and vision–language capabilities while introducing action-relevant geometry. Stage I learns geometric residuals, and Stage II controls their contribution through soft regularization and stochastic scaling.

To evaluate this design, we add a controlled GQA comparison between one-stage GeoLACE (Ours-I) and full two-stage GeoLACE. Switching to Stage II improves GQA retention and full-answer accuracy and reduces JS divergence, whereas continuing Stage I degrades all three metrics. Since both variants start from the same model after 25k steps of Stage I and avoid explicit feature alignment, their GQA comparison supports Stage II’s contribution to capability preservation beyond simply avoiding an explicit alignment loss.

Training configurationRetention (%) ↑Full-answer accuracy (%) ↑JS divergence ↓
Stage I 25k87.0462.330.0347
Ours-I: Stage I 30k84.7160.700.0377
GeoLACE: Stage I 25k + Stage II 5k87.4062.800.0325

Relative to the shared model after 25k steps of Stage I, Ours-I shows lower retention and full-answer accuracy and higher JS divergence. GeoLACE instead improves all three metrics after switching to Stage II. Since both variants use the same training budget without explicit feature alignment, this comparison supports Stage II’s contribution to capability preservation. Stage I learns task-relevant residuals using the native action loss, and Stage II continues action training with soft regularization and stochastic scaling to control their contribution. The observed changes are consistent with controlled residual adaptation helping preserve pretrained vision–language capabilities during continued action training. The complete results and analysis have been added to Appendix C.3 (Table 14).

[W2&Q2] 2.3 Why does the 20k+10k Safe-Add schedule drop LIBERO success to 96.4%?

Response: Compared with 25k+5k, the 20k+10k schedule shortens the geometry-learning phase that uses only the action loss and applies residual regularization and stochastic scaling earlier and for longer. Our interpretation is that these constraints may limit task adaptation before the geometric residuals are sufficiently learned, contributing to the performance decline.

Evidence in the original submission. This sensitivity was partially discussed in Sec. 4.2 of the original manuscript, which compared schedules under the same 30k-step budget and noted that introducing regularization too early or maintaining it for too long may limit adaptation.

Analysis of the original schedule ablation. The 20k+10k schedule achieves 96.4% success, below both one-stage training (97.5%) and the native baseline (96.9%), whereas 25k+5k achieves 98.1%. Compared with 25k+5k, the 20k+10k schedule shortens the initial geometry-learning phase and introduces residual regularization and stochastic scaling earlier and for longer. The lower success is consistent with residual regularization and stochastic scaling constraining task adaptation before the geometric residuals are sufficiently learned. Under the same 30k-step training budget, the results support the 25k+5k schedule, which provides a longer geometry-learning phase followed by a shorter capability-preservation phase.

Reviewer_31bHOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

GeoLACE augments a VLA with patch-aligned residual features from a frozen VGGT. It selects source layers using representation similarity and target layers using action-loss gradients, then applies two-stage Safe-Add training. The method improves LIBERO success from 96.9% to 98.1% for the evaluated π backbone and from 97.0% to 98.6% for GR00T. It also reports gains on six RoboTwin tasks and real-robot tasks.

Critical strengths
  1. The method addresses a relevant limitation of VLA representations and transfers across two policy backbones. Its real-robot results make the contribution more practically meaningful than simulation-only evidence.
  2. The analysis is unusually detailed: matched random-layer controls, probe repetitions, selection bootstrap, residual diagnostics, and VQA retention tests provide evidence beyond aggregate task scores.
  3. The paper has a clear structure, a logical flow, and rich experimental details, making it easy to follow.
Critical weaknesses
  1. The incremental contribution is not fully isolated from closely related geometric fusion. SpatialStack [1] already uses multi-layer VGGT features and additive visual-token fusion. GeoLACE adds action-conditioned selection and Safe-Add, but the current random-layer controls do not replace a matched fixed hierarchical mapping. Stage II also combines residual regularization and stochastic scaling without separately establishing their contributions.
  2. The Related Work section omits several relevant studies that incorporate VGGT into VLA models [1–3]. Including these works and clarifying how the proposed method differs from them would make the discussion more complete.
  3. The calibration/probe repetitions and image-level VQA bootstrap do not quantify uncertainty over complete policy training runs. Given the near-ceiling LIBERO results, the six-task RoboTwin subset, and small real-robot trial counts, independent policy seeds would strengthen the evidence for consistent gains.

[1] IntentVLA: Short-Horizon Intent Modeling for Aliased Robot Manipulation, EMNLP 2026 Oral [2] Learning Action Manifold with Multi-view Latent Priors for Robotic Manipulation, arXiv:2605.11832 [3] GeoAware-VLA: Implicit Geometry Aware Vision-Language-Action Model,arXiv:2509.14117

Questions for authors
  1. Can you compare against a fixed SpatialStack-style layer mapping under the same branch count, training budget, and Safe-Add schedule, and separate the two Stage II components?
  2. How well do selection scores correlate with fully trained policy success across candidate mappings? Would alternative source-layer choices change the conclusion?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 31bH (1/3)

Thank you for your careful review and constructive suggestions. We appreciate your recognition of the method’s transfer across policy backbones, practical real-world results, and detailed supporting analyses. We address your concerns point by point below, grouping closely related weaknesses and questions and indicating the corresponding manuscript revisions.

[W1&Q1] 1.1 Can you compare against a fixed SpatialStack-style layer mapping under the same branch count, training budget, and Safe-Add schedule?

Response: Yes. We have added the requested fixed SpatialStack-style comparison. Specifically, with the same three branches, 25k+5k training budget, fusion architecture, and Safe-Add schedule, ActSelect achieves 98.1% average LIBERO success versus 97.1% for the SpatialStack-style mapping, with higher success on every suite.

Additional experiment. We have added a matched SpatialStack-style mapping comparison in Appendix A.4 (Table 6), with a corresponding reference in Sec. 4.2. All configurations use three branches and the 25k+5k training schedule. Each configuration is evaluated on 500 rollouts per suite, totaling 2,000 rollouts across the four LIBERO suites.

SpatialStack demonstrates effective hierarchical geometric fusion for VLM spatial reasoning using a manually specified mapping. Its source–target correspondences are designed for spatial reasoning without an action-conditioned selection criterion. We evaluate their transfer to VLA manipulation while holding the projector architecture, additive visual-token fusion, training data, optimizer, and Safe-Add schedule fixed.

ConfigurationSpatialObjectGoalLongAverage
SpatialStack-style + Safe-Add98.697.898.293.897.1
ActSelect + Safe-Add98.899.099.295.298.1

This comparison shows that hierarchical feature coverage alone does not determine the most effective correspondences for action prediction. ActSelect evaluates candidate source–target correspondences using the native action loss, combining complementary source representatives with action-relevant target layers. The improvement with fusion and Safe-Add held fixed supports the additional contribution of our action-oriented source–target selection. Complete per-suite results and analysis are provided in Appendix A.4, Table 6.

[W1&Q1] 1.2 Can you separate the two Stage II components?

Response: Yes. We have added separate Stage II component ablations. Specifically, removing soft regularization reduces average success to 96.8%, while removing stochastic scaling reduces it to 97.4%, compared with 98.1% for full Safe-Add under otherwise identical settings. Both components contribute to the full configuration’s average success.

Additional experiment. We have added separate Stage II component ablations in Appendix A.4 (Table 7), with a corresponding reference in Sec. 4.2. All variants retain the same ActSelect-selected source–target mapping. Each ablation removes one Stage II component while keeping all other training and evaluation settings unchanged. All configurations use the 25k+5k training schedule and 500 evaluation rollouts per suite.

The configuration without soft regularization sets λ=0\lambda=0 and retains the original stochastic scaling distribution. The configuration without stochastic scaling fixes α=1\alpha=1 and retains λ=0.01\lambda=0.01. At inference, all variants use α=1\alpha=1. Table 7 reports the following success rates (%).

ConfigurationSpatialObjectGoalLongAverage
w/o soft regularization98.497.697.693.696.8
w/o stochastic scaling99.298.697.294.697.4
Full Safe-Add98.899.099.295.298.1

Soft regularization penalizes residual ratios above the Stage-I reference τk\tau_k, while stochastic scaling trains action prediction with full, attenuated, or zero geometry. The penalty is computed from unscaled residuals and remains active across sampled scales. These results support the individual contributions of both components within Safe-Add.

AuthorsOpenReview ↗

Response to Reviewer 31bH (2/3)

[W2] The Related Work section omits several relevant studies that incorporate VGGT into VLA models [1–3]. Including these works and clarifying how the proposed method differs from them would make the discussion more complete.

Response: We have added the three suggested studies and SA-VLA to Sec. 2 and clarified how their geometric integration differs from GeoLACE’s action-conditioned source–target selection and controlled layerwise residual fusion.

Additional related work and clarification. We have added IntentVLA and Multi-view-VLA to the “Single-Level” paragraph and GeoAware-VLA to the “Multi-Level” paragraph of Related Work. Beyond the three suggested studies, we have also included SA-VLA [1] in the “Single-Level” discussion. The revised discussion distinguishes their geometric feature fusion or aggregation from GeoLACE’s action-conditioned source–target selection and controlled layerwise residual fusion.

[W3] The calibration/probe repetitions and image-level VQA bootstrap do not quantify uncertainty over complete policy training runs. Given the near-ceiling LIBERO results, the six-task RoboTwin subset, and small real-robot trial counts, independent policy seeds would strengthen the evidence for consistent gains.

Response: We now report mean ± standard deviation of GQA results across three independent complete policy training runs. For manipulation, we add controlled LIBERO comparisons and explain how the six RoboTwin tasks cover varied geometric requirements and provide evidence beyond near-ceiling LIBERO. We also increase real-world evaluation to 40 trials per policy per task and add dynamic-scene experiments and a synchronous VGGT comparison.

Additional GQA evaluation across independent training runs. For each method, we independently fine-tune policies on LIBERO using three different training seeds and evaluate the resulting models on the same 4,000 held-out GQA questions without further parameter updates. Ours-I denotes one-stage GeoLACE trained for 30k steps. Results are reported as mean ± standard deviation across the three independent training runs.

MethodRetention (%) ↑Acquisition (%) ↑Full-answer accuracy (%) ↑JS divergence ↓
π₀.₅-FT86.77±0.37‾\underline{86.77 \pm 0.37}11.61 ± 0.6861.12 ± 0.400.03234 ± 0.00104
SF83.48 ± 0.6015.80 ± 0.6060.00 ± 0.350.03804 ± 0.00057
ROCKET85.44 ± 0.7210.72 ± 0.8961.23±0.15‾\underline{61.23 \pm 0.15}0.03855 ± 0.00121
Ours-I85.19 ± 0.6113.00±1.05‾\underline{13.00 \pm 1.05}60.89 ± 0.400.03867 ± 0.00098
GeoLACE87.37 ± 0.8612.44 ± 0.7461.88 ± 0.440.03297±0.00114‾\underline{0.03297 \pm 0.00114}

Bold and underlined values indicate the best and second-best mean results in each column, respectively.

GeoLACE achieves the highest mean retention and full-answer accuracy. Compared with Ours-I under the same 30k-step training budget, it improves mean retention by 2.18 percentage points and full-answer accuracy by 0.99 percentage points, while reducing mean JS divergence from 0.03867 to 0.03297. These results support the capability-preservation benefit of Stage II while quantifying GQA variation across complete policy training runs.

Additional LIBERO comparisons. We have added SA-VLA (SFT) [1] to Table 1 and controlled comparisons covering the SpatialStack-style mapping, Stage II components, GFM source selection, and VLA target mappings. These comparisons, detailed in our responses to W1/Q1 and Q2, strengthen the evidence for the individual design choices.

RoboTwin task coverage. The six tasks already reported in Figure 5 cover several geometric requirements: object relocation and placement (Move Playingcard and Place Shoe), interaction with switches and doors (Turn Switch and Open Microwave), tool–target coordination (Beat Block Hammer), and coordinated bimanual lifting (Lift Pot). This coverage provides a focused evaluation of how object poses and spatial relations inform precise manipulation. Each task is evaluated over 100 trials in both easy and hard settings. GeoLACE’s average advantage over π₀.₅ is 7.5 percentage points under easy conditions and 11.0 under randomized hard conditions, providing evidence beyond the near-ceiling LIBERO results.

Expanded real-world evaluation. We have increased evaluation from 20 to 40 trials per policy per task across all five tasks on two robot platforms (Table 4; Appendix D.3). We also add dynamic-scene experiments across the five tasks and a synchronous VGGT comparison in Setting A (Appendix D.4, Tables 15–16), extending evaluation to changing scenes and different execution conditions.

Together, these additions provide evidence across independent training runs for capability preservation and broader controlled evaluation of manipulation performance.

[1] Xu Pan et al. 2026. SA-VLA: Spatially-Aware Flow-Matching for Vision-Language-Action Reinforcement Learning. arXiv:2602.00743.

AuthorsOpenReview ↗

Response to Reviewer 31bH (3/3)

[Q2] 2.1 How well do selection scores correlate with fully trained policy success across candidate mappings?

Response: We extend the original comparison across all five three-branch mappings with completed fine-tuning—two high-ranked ActSelect candidates and three random controls. Specifically, higher LCB scores are generally associated with higher average LIBERO success, with a Spearman rank correlation of 0.90. The selected mapping ranks first by both measures, with a small reversal among the lowest-scoring random controls indicating a positive overall trend rather than strict monotonicity.

Evidence in the original submission. The reason for expecting an association was discussed in Sec. 3.2 of the original manuscript: ActSelect scores source–target pairs by their normalized first-order contribution to reducing the native action loss. The matched target-layer ablation in Sec. 4.2 provided initial downstream evidence, with 98.1% average success for the selected mapping versus 97.3%–97.5% for random targets.

Additional experiment and analysis. We have expanded Appendix B.3 (Table 10) to compare LCB scores with success rates after full fine-tuning across all five three-branch mappings, covering two high-ranked ActSelect candidates and the three random controls in Table 2. We compute LCB scores for the existing random controls using the same ActSelect procedure, with VGGT source layers and the training protocol held fixed across all configurations.

MappingNormalized LCBAverage success rate (%)
(1, 7, 9)1.851998.1
(0, 8, 9)1.773397.8
Random layer selection 3 (1, 7, 16)1.339997.5
Random layer selection 2 (5, 9, 13)1.216697.3
Random layer selection 1 (0, 3, 6)1.002597.4

Each configuration is evaluated on 500 rollouts per suite across the four LIBERO suites.

The fourth-ranked ActSelect candidate (0, 8, 9) lies between the selected mapping and the random controls in both LCB score and average success rate. The small reversal between the two lowest-scoring random controls indicates an overall positive trend rather than strict monotonicity.

This trend is consistent with the scoring mechanism in Sec. 3.2: ActSelect aggregates normalized first-order action-loss utilities across probes and applies an LCB correction for uncertainty across calibration evaluations. The observed association supports the downstream relevance of this action-based selection criterion.

[Q2] 2.2 Would alternative source-layer choices change the conclusion?

Response: The tested alternative source-layer choices do not change our overall conclusion. We extend the original source-selection discussion with additional comparisons. Specifically, both tested alternatives remain above the π₀.₅ baseline of 96.9%, while CKA-selected sources achieve higher success on every suite: 98.1% on average, compared with 97.4% for uniform selection and 97.2% for random selection.

Evidence in the original submission. This point was partially discussed in Secs. 3.1–3.2 of the original manuscript: VGGT features span localized structure and broader context, motivating complementary source selection by representational similarity. Appendix A.1 (Figure 9) recovered the same stages—0–13, 14–17, and 18–23—across three independently sampled calibration sets, supporting stable coverage of distinct representations.

Additional experiment. Under the same training settings and VLA mapping (1, 7, 9), our CKA-selected sources (8, 16, 21) achieve 98.1% average LIBERO success, compared with 97.4% for uniform selection and 97.2% for random selection, with higher success on every suite. These results support source selection based on the GFM’s representation structure, and the stage-coverage difference provides a mechanism consistent with the observed gains. We have added these results and analysis to Appendix A.1 (Table 5).

Source selectionSource layersAverage LIBERO success (%)
Uniform(6, 12, 18)97.4
Random(4, 10, 12)97.2
CKA-based(8, 16, 21)98.1

Analysis. For the tested alternatives, uniform sources (6, 12, 18) omit the middle stage, while random sources (4, 10, 12) lie exclusively within the first stage. Equal spacing in layer indices therefore does not ensure coverage of distinct representational stages.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.