acceptodds
Withdrawn from ICLR 2027

ScopeVLA: Task-Grounded and View-Specialized Spatial Conditioning for Robotic Manipulation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Incorporating geometric information into vision-language-action (VLA) models has become an important approach to improving robotic manipulation performance. However, identifying task-relevant spatial evidence and exploiting the distinct roles of third-person and wrist views remain challenging. In this paper, we introduce ScopeVLA, a task-grounded framework for view-specialized spatial conditioning. ScopeVLA employs Global and Wrist queries to extract task-relevant spatial distributions from intermediate vision-language model (VLM) representations. These queries are supervised by view-specific workspace targets derived from end-effector trajectories. These distributions guide two complementary conditioning pathways: the Global prior introduces a confidence-aware attention bias over third-person visual tokens, while the Wrist query assignments spatially modulate multi-level features extracted from a frozen geometry model. The modulated geometric features are projected into additional keys and values for joint attention in the action expert. Both pathways operate at selected action-expert layers while preserving the original layerwise vision-language conditioning. Controlled instruction interventions show that the learned spatial distributions remain stable under task-preserving paraphrases while responding to changes in task instructions. Experiments on two simulation benchmarks and real-world manipulation tasks show that ScopeVLA improves task success over the evaluated VLA baselines.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Geometric Priors, View-Specialized Spatial Conditioning, Robotic Manipulation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AgNsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ScopeVLA, a method for task-grounded spatial conditioning in multi-view VLA models. It adds learnable Global and Wrist queries to the VLM to read out task-relevant spatial distributions over the third-person and wrist views from intermediate layers. These distributions are supervised by targets built from the end-effector trajectory, which is turned into a 3D workspace field and projected onto each view. The Global distribution is used as an attention bias on third-person visual tokens, and the Wrist distribution modulates wrist-view geometric features from a frozen DA3 model, which are then added as extra keys and values in selected layers of a VLA-Adapter action head. Experiments on CALVIN ABC→D, LIBERO-Plus, and four real-world tasks show improvements over VLA-Adapter, from 4.50 to 4.62 average length on CALVIN, from 59.1 to 67.4 on LIBERO-Plus, and from 61.25% to 80.0% on the real-world tasks.

Critical strengths
  1. The problem is relevant. Geometric features contain both task-relevant and irrelevant structure, and selecting the task-relevant part while treating the third-person and wrist views differently is a reasonable direction for geometry-enhanced VLAs.
  2. The design is practical. The supervision is built automatically from trajectories and depth without manual labels, and depth, calibration, and trajectories are only needed during training. The NULL option and the zero-initialized bias keep the original model unchanged at initialization, and the inference cost is reported.
  3. The evaluation covers both simulation and real-world tasks, and the gains over VLA-Adapter are consistent across them. The ablations cover the components, view allocation, and injection depth, and the appendix shows how the routes respond to instruction and scene changes.
Critical weaknesses
  • The novelty is limited, and the relation to prior work on task-relevant grounding is not clear. The trajectory-based supervision is essentially a heatmap of the future end-effector path on the image, and using trajectories or paths as intermediate representations or supervision has been studied in TraceVLA, HAMSTER, and PEEK, where PEEK also derives task-relevant masks automatically from trajectories. Supervising policy attention toward task-relevant regions (GuidedVLA, Gaze-Regularized VLA), instruction-aware visual feature selection (OTTER), and masking irrelevant regions (ARRO, BYOVLA) are also closely related, but none of them is discussed, and ReconVLA appears in Table 1 without any discussion. The architecture itself combines existing components such as query readout, attention bias, multiplicative gain, and extra keys and values, so the design contribution is somewhat limited.
  • The baseline set is selective, and the main claims depend on it. The OpenVLA, UniVLA, and π0 numbers in Table 2 are taken from the LIBERO-Plus paper, but stronger methods from the same table are left out, including OpenVLA-OFT (69.6), RIPT-VLA (68.4), OpenVLA-OFT_m (67.9), and π0-Fast (61.6). ScopeVLA (67.4) is below the first three, and with these baselines it is no longer first on Camera, Light, and Background, so the claims of the highest overall score and first place in five of seven categories only hold for the selected baselines. The paper refers to parameter-comparable baselines, yet 7B models such as OpenVLA, UniVLA, and Spatial Forcing are kept in the table. In addition, the real-world experiments only compare with VLA-Adapter, and no variance is reported for the CALVIN gain from 4.50 to 4.62.
  • The ablations do not test the main claim, and the analysis is not very convincing. The main claim is task grounding from trajectory supervision, but the ablations only add or remove modules and never change the supervision, for example by removing the routing loss or replacing the trajectory field with object masks or a uniform prior. In Table 4(a), the largest single gain comes from adding DA3 wrist features (46.3 to 56.5), while G-Condition adds only 4.8 points, so much of the improvement may come from the extra geometry model. The intervention study in Appendix A only evaluates the Stage I readout and does not test whether the policy actually uses the routes, and under wrong-task instructions the Wrist route changes little (KL from 0.102 to 0.146). The motivation in Section 3.2.1 also relies on visualizations of two selected attention heads.
  • Some choices in the experimental setup raise concerns. The readout layer is selected on CALVIN and the injection layers are selected with Table 4(c) on LIBERO-Plus Object, and both are evaluation benchmarks. All ablations use only the LIBERO-Plus Object subset without variance, and although the model is trained on LIBERO, no results on standard LIBERO are reported. The real-world evaluation has only 4 tasks with 20 trials per method. Training also requires calibrated depth for both views and full trajectories, which may limit the method on large datasets without calibration.
Questions for authors
  1. What is the main novelty compared with methods that use trajectories or paths, such as TraceVLA, HAMSTER, and PEEK, and with methods that supervise attention directly, such as GuidedVLA and Gaze-Regularized VLA? Can the authors compare with at least one of them under the same backbone?
  2. Why are OpenVLA-OFT, RIPT-VLA, OpenVLA-OFT_m, and π0-Fast from the LIBERO-Plus paper not included in Table 2, and what is the criterion for parameter-comparable baselines? Also, the UniVLA total in Table 2 (43.9) differs from the original paper (42.9).
  3. Can the authors ablate the supervision source, for example by removing the routing loss or using object masks instead of the trajectory field? How much does trajectory supervision add on top of the DA3 features alone?
  4. If the routes are replaced with a uniform distribution or with routes from a wrong task at inference time, how do the success rate and behavior of the policy change? Under a wrong instruction, do the routes move toward the region of the new task?
  5. Were the readout layer and injection layers selected on a separate validation split? Can the authors report variance on CALVIN and LIBERO-Plus, as well as results on standard LIBERO?
  6. How would the method work on data without calibrated depth? Can the real-world experiments include geometry-enhanced baselines such as FALCON or Spatial Forcing, with more tasks and trials?
Reviewer_nUbVOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors argue that robotic manipulation benefits from task-relevant geometric information and propose different designs for third person and wrist camera views. They use a two-stage training to locate the relevant area and inject depth information to the action head. They evaluate their approach primarily in simulation, supplemented by a real-world robot evaluation.

Critical strengths
  1. The paper addresses how to select task-relevant spatial information into VLA policies. Its distinction between global scene context and wrist-local interaction geometry provides a clear rationale for the asymmetric conditioning architecture.
  2. The method constructs grounding targets from demonstration trajectories and calibrated depth observations, avoiding manual object masks or region labels. These supervisory inputs are not required at inference.
  3. Consistent improvements across the evaluated settings. ScopeVLA improves over VLA-Adapter on CALVIN, LIBERO-Plus overall, and all four real-world tasks.
Critical weaknesses

1.Limited validation across VLA backbones and against stronger baselines. The method is evaluated only within the VLA-Adapter framework, leaving its applicability to other VLA architectures unverified. Although ScopeVLA improves over its chosen baseline, the comparisons do not establish whether these gains remain meaningful relative to stronger contemporary policies, such as pi_{0.5}. Evaluating the proposed spatial-conditioning mechanisms on a stronger backbone, together with its unmodified counterpart under the same training and evaluation protocol, would help determine whether the approach offers broadly useful improvements rather than primarily compensating for limitations of the selected baseline.

  1. The contribution of geometric knowledge is not sufficiently isolated. Table 4(a) includes a Global-only baseline without DA3, but no corresponding baseline that applies Wrist-region guidance to the existing wrist visual features without DA3. A matched non-geometric feature baseline is also missing (such as dino, etc). Consequently, the results do not clearly distinguish the benefits of task-region selection, additional visual representations, and geometry-specific knowledge.

  2. The trajectory-derived supervision has unverified assumptions and robustness. The target construction relies on an estimated interaction anchor, a fixed gripper-envelope radius, kernel bandwidths, and top-20% patch aggregation, without any sensitivity analysis. Moreover, a single workspace field constructed from the complete demonstration does not explicitly distinguish successive interaction stages. Its suitability for multi-stage tasks remains insufficiently established. For example, a trajectory may completely takes up the entire workspace in some complex tasks.

4.The central design choices are only partially validated. The ablations do not directly establish the necessity of two-stage training, the asymmetric query-visibility mask, the NULL option, or the query-specialization regularizer. Furthermore, most design ablations are restricted to LIBERO-Plus Object. In Table 4(b), the 59.7% versus 59.5% comparison supports no demonstrated benefit from adding Global geometry, rather than a clear disadvantage of dual-view geometry.

5.The real-world evaluation is limited to simple pick-and-place tasks, leaving the method’s applicability to more complex manipulation uncertain. Evaluation on more challenging manipulation tasks is needed.

Questions for authors

Please see the critical weaknesses.

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

This paper proposes ScopeVLA, a task-grounded framework for view-specialized spatial conditioning in VLA models. Learnable Global (G) and Wrist (W) queries read intermediate VLM representations and produce view-specific task-region distributions, supervised by end-effector-trajectory-derived workspace targets without manual masks. The Global distribution provides a confidence-aware attention bias over third-person visual tokens via Bridge Attention, while Wrist query responsibilities drive W-RSFM to modulate multi-level geometric features from a frozen DA3-BASE expert, injected as additional keys/values at selected action-head layers. Training is two-stage: route learning then joint policy training with LoRA. Experiments on CALVIN ABC→D, LIBERO-Plus zero-shot, and four real-world tasks with distractor objects compare against VLA-Adapter and several VLA/geometry-enhanced baselines, with ablations on components, view allocation, and injection depth.

Critical strengths
  1. The trajectory-grounded dual-view grounding is well motivated and practical. Using end-effector trajectories to define a continuous workspace field avoids manual region masks while still providing dense, view-specific supervision. The NULL option and query competition are sensible design choices that allow the route to represent "how much relevant evidence is visible," not just a normalized spatial prediction.

  2. The role-asymmetric integration is a clean architectural idea. Global grounding as an attention bias on third-person tokens, and Wrist grounding as spatial modulation of dense geometric features, reflects the complementary roles of the two views. The two-stage training with disjoint parameter updates is a reasonable way to stabilize route learning.

  3. The analysis is broader than success rates. Controlled instruction/scene interventions, view-allocation ablations, injection-depth ablations, and inference-efficiency measurements give useful insight. The appendix is detailed (attention mask, target construction, hyperparameters, latency), which supports reproducibility.

Critical weaknesses

1. Baseline coverage is selective.
Table 1 (CALVIN) compares against VLA-Adapter, π0, OpenVLA, UniVLA, ReconVLA, Spatial Forcing, FALCON, and DreamVLA, but omits geometry-enhanced methods such as PointVLA, GeoVLA, BridgeVLA, GLaD, ROCKET, and VEGA, several of which are cited in related work. Table 2 (LIBERO-Plus) includes Spatial Forcing and DreamVLA but not the others. Since the paper's main claim is about geometric integration, the most relevant comparisons are missing. If those methods do not have public CALVIN/LIBERO-Plus results, the paper should say so explicitly rather than leave the omission unexplained.

2. The wrist geometry expert and target construction raise practical questions.
DA3-BASE is a monocular depth/geometry model. How robust is it on wrist-view images with close-up objects and potentially poor lighting? The trajectory-derived targets require calibrated depth and camera poses at training time; the paper says these are not needed at inference, but the quality of training targets depends on calibration accuracy. An analysis of target quality (e.g., sensitivity to depth noise or calibration error) would be valuable. Also, during manipulation the wrist camera moves with the end-effector—how does this affect back-projection and the workspace field evaluation?

3. The computational overhead is non-trivial and under-discussed.
Latency increases from 31.48 ms to 52.67 ms (+67%), FLOPs from 1.496T to 1.799T, and query rate drops from 31.76 Hz to 18.99 Hz. For an 8-step action chunk, 18.99 Hz still yields a reasonable effective action rate, but the paper does not discuss whether this overhead is acceptable for contact-rich or reactive tasks, nor whether the DA3 expert can be distilled or cached. The conclusion should mention this trade-off.

Questions for authors
  1. How were the injection blocks {0,6,11,12}\{0,6,11,12\} chosen? Can you report a search over a larger set of block combinations, or at least justify blocks 0, 6, and 11 individually?

  2. Why are PointVLA, GeoVLA, BridgeVLA, GLaD, ROCKET, and VEGA not compared on CALVIN or LIBERO-Plus? If their results are unavailable under the same protocol, please state this explicitly.

  3. How robust is the trajectory-derived target construction to depth noise and calibration error? How is the interaction anchor k⋆k^\star estimated, and how sensitive are results to it?

  4. How does the moving wrist camera affect back-projection and workspace-field evaluation during manipulation? Is the DA3 expert run online at every step, and can it be cached or distilled?

  5. The latency increases by 67% (31.48→52.67 ms). Is this acceptable for contact-rich or reactive tasks? Can you discuss the trade-off and possible efficiency improvements?

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

Related papers

Discussion (0)

Sign in to comment.