acceptodds
Under review as a conference paper at ICLR 2027

DGA-VLA: A Lightweight Depth-Guided Grasp Enhancement for VLA Models without Retraining

Abstract

Vision-language-action (VLA) models provide a unified policy foundation for a wide range of robotic manipulation tasks. However, their grasping performance remains constrained by imprecise gripper-object alignment at closure, leading to empty or unstable grasps and task failures. To address this limitation, we propose DGA-VLA, a lightweight depth-guided grasp refinement framework that augments pretrained VLA policies without VLA architectural changes or policy retraining. DGA-VLA couples the VLA's global motion generation with a wrist-depth feedback controller, enabling local correction without changing the underlying policy. When the policy indicates a closure intent, the controller uses wrist-depth feedback to estimate local gripper-object offsets and apply bounded corrective motions before closure. Experiments demonstrate that DGA-VLA improves average success rates on grasping tasks by 12.0 percentage points on RoboTwin, 8.4 points on SimplerEnv, and 14.6 points on real-robot pick-and-place tasks.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Robotic Manipulation, Grasp Refinement, Depth-Guided Control, Training-Free Adaptation

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 2 reviews, above 27% of the others.
Reviewer_aqSFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DGA-VLA, which uses wrist-camera depth and visual servoing to correct grasp positions when VLA predicts gripper closing, without retraining VLA. It improves success rates by 12.0%, 8.4%, and 14.6% on RoboTwin, SimplerEnv, and real robots.

Critical strengths

Simple and practical. It can be directly added to existing VLAs without retraining.​2. Good experiments. The method is tested on multiple benchmarks and real robots.​3. Performance improvement. The real-world success rate improves from 26.7% to 41.3%.

Critical weaknesses

1.Limited novelty. The main idea is based on traditional depth-based visual servoing. ​2. Manual calibration. Each object needs an expert demonstration to get the grasp anchor, which limits generalization.​ 3. Missing baselines. There is no direct comparison with traditional visual servoing methods.​ 4. Limited generalization. Real-world experiments only include three objects, without testing unseen objects, occlusion, or depth noise.​

Questions for authors
  1. What is the main novelty compared with traditional visual servoing?

  2. Can this method work on unseen objects without grasp anchors?

  3. How does it handle occlusion or multiple nearby objects?

  4. Can the authors provide ablations on grasp anchors, depth noise, and calibration?

5.Can the authors compare with simple position correction or standard visual servoing?

2 comments

We thank the reviewer for recognizing the simplicity, practical utility, and results across simulation and physical hardware.

1. Contribution and comparison with visual servoing (Q1 and Q5)

“What is the main novelty compared with traditional visual servoing?”

Methodological insight (Q1). As discussed in the Introduction and Related Work, our contribution is the collaboration between a frozen VLA and local visual feedback. This exploits their complementary strengths: task understanding and approach motion from the VLA, and precise grasp alignment from geometric feedback. To realize this collaboration, we design depth-based localization and geometric correction around the stage-specific characteristics of manipulation tasks and shortcomings in VLA grasp precision, drawing on a visual-servoing control law rather than adopting an existing complete algorithm.

“Can the authors compare with simple position correction or standard visual servoing?”

Standard visual servoing (Q5). Accordingly, comparing with standard visual servoing requires specifying the grasp target, desired alignment, and coordination with the VLA. The control law alone leaves these choices open, and their design is a substantive part of our framework.

Simple position correction (Q5). For simple position correction, the paper already includes a comparison with PhysVLA on LIBERO. PhysVLA incorporates simple position corrections using privileged simulator states, providing a relevant baseline for this comparison.

2. Calibration and generalization (Q2)

“Manual calibration. Each object needs an expert demonstration to get the grasp anchor, which limits generalization.”

“Can this method work on unseen objects without grasp anchors?”

DGA requires a grasp anchor to define the desired alignment. This reference can be shared across object instances with the same category and grasp affordance, and an existing library entry with similar object geometry and grasp configuration provides a candidate reference for a new object. Our RoboTwin evaluation includes within-category object variants, with anchors fixed during evaluation. When a new reference is needed, automatic demonstration-based calibration adds it to the library for future reuse. This procedure can use demonstrations already collected for VLA fine-tuning, keeping the additional preparation cost low.

3. Occlusion and nearby objects (Q3)

“How does it handle occlusion or multiple nearby objects?”

Nearby objects. DGA reduces interference through depth thresholding, a local region of interest, gripper masking, and connected-component filtering, then selects the candidate nearest to the grasp anchor. Crucially, the VLA's preceding motion has already brought the gripper close to the intended object, making local geometric extraction effective even in a scene containing multiple objects, such as the kitchen scene in LIBERO-Long Task 0.

Occlusion. The same collaboration helps with occlusion. When nearby objects obstruct a direct grasp, the VLA is responsible for selecting a graspable approach direction through its preceding motion. This brings the gripper and wrist camera close to the intended grasp region, where DGA uses local depth feedback to refine alignment before closure.

4. Grasp-anchor ablation and depth-noise evidence (Q4)

“Can the authors provide ablations on grasp anchors, depth noise, and calibration?”

Grasp-anchor calibration. The grasp anchor defines the desired state for alignment. We therefore examine its construction by replacing demonstration-derived anchors with a default reference at the gripper center on the same eight RoboTwin tasks reported in the paper, retaining the other settings. Average task success drops to 23.4%, compared with 42.3% for the vanilla VLA and 54.3% for calibrated DGA, demonstrating the importance of calibrating an appropriate grasp reference.

Depth noise. Simulation uses depth from the simulator, while the real-robot experiments use measurements from the wrist-mounted stereo depth camera. As shown in the paper's Figure 3(d) and described in Section 4.3, these observations contain invalid returns and local measurement noise. DGA still improves real-robot success from 26.7% to 41.3%, providing direct evidence of effectiveness with actual sensor depth.

Reviewer_aqSFOpenReview ↗

Rebuttal response

Thank you authors for providing feedbacks, my concerns are fully resolved. I have updated my scores.

Reviewer_GrWqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a grasp assistance module for VLAs using wrist-depth feedback and object-category-specific grasp references. The module does not require retraining the VLA and instead acts as a wrapper, temporarily taking control when the VLA indicates gripper-closure intent. It uses expert demonstrations to build a library of reference object positions and depths, indexed by task and object category, and applies local corrections to align the gripper before returning control to the VLA. The paper evaluates the system on three simulation benchmarks and real-world pick-and-place tasks, reporting varying levels of improvement.

Critical strengths

None

Critical weaknesses

The method offers limited methodological novelty while introducing restrictive priors that undermines the VLA’s broader capabilities. It is unclear what substantial new insight the method provides beyond a task-calibrated visual-servoing wrapper. The correction depends on a manually configured library of task- and object-category entries calibrated from expert demonstrations, requiring new entries for additional categories and leaving behavior without a matching entry unspecified. The wrapper intervenes based on gripper-closure intent without establishing that its alignment reference is appropriate for the current interaction. For example, grasping a mug by its handle rather than its body, or closing the gripper to push something, can require substantially different behavior. The paper does not establish that the benefit on selected grasping tasks justifies imposing these priors on a general-purpose policy.

The experiments do not establish a compelling or broadly applicable benefit. The reported improvements are uneven, with particularly small aggregate gains on LIBERO, while the other evaluations concentrate on selected grasping and pick-and-place tasks. The evaluation does not adequately test interactions where the imposed grasp prior may be inappropriate, or new object categories without calibrated references. Consequently, it does not establish the cost of these interventions to the policy’s existing capabilities. In SimplerEnv, DGA additionally receives a wrist-depth view unavailable to the baseline policy, so the comparison does not isolate the proposed correction mechanism from the benefit of additional sensing. These limitations leave the experimental support substantially narrower than the proposed use as a general VLA assistance module.

Questions for authors

I have no questions whose answers would change my assessment.

3 comments

Thank you for the review. We address the two main concerns separately below.

1. Insight and generality

“The method offers limited methodological novelty while introducing restrictive priors that undermines the VLA’s broader capabilities.”

Insight and selective assistance. DGA preserves the VLA's broader capabilities by enabling assistance selectively for grasping interactions identified from the task instruction. Other interactions remain under the VLA's original control. Our insight is that lightweight geometric feedback can compensate for shortcomings in VLA grasp precision at very low additional cost, combining the policy's task-level capabilities with precise local feedback.

“The correction depends on a manually configured library of task- and object-category entries calibrated from expert demonstrations, requiring new entries for additional categories and leaving behavior without a matching entry unspecified.”

Calibration and reference selection. The grasp references are calibrated automatically from successful demonstrations and reused across instances with the same category and grasp configuration. When an exact category entry is unavailable, we select the closest available category with compatible grasp geometry. DGA can reuse a small subset of the demonstrations already collected for VLA fine-tuning, making calibration a low-cost addition to the deployment workflow.

“For example, grasping a mug by its handle rather than its body, or closing the gripper to push something, can require substantially different behavior.”

Grasp affordances. The VLA determines the intended grasp region through its preceding motion, and DGA uses a reference matched to the corresponding object category and grasp affordance. This preserves the division of labor described in Section 3.3: the VLA supplies task-level affordance selection, and DGA supplies local geometric refinement.

Other interactions. Non-grasp closure commands remain under VLA control. The same geometric-feedback principle could improve precision in other manipulation stages through adjustments to the control-arbitration conditions. We focus on grasping because improving this fundamental capability is an important step toward practical deployment and broader robotic manipulation.

2. Experimental evidence

“The evaluation does not adequately test interactions where the imposed grasp prior may be inappropriate, or new object categories without calibrated references. Consequently, it does not establish the cost of these interventions to the policy’s existing capabilities.”

Grasp reliability. As clarified above, DGA intervenes selectively during grasping and preserves the VLA's control of other interactions. Our experiments focus on the resulting improvement in grasp reliability: for example, average first-grasp success on RoboTwin increases from 68.3% to 79.5%. This addresses a basic capability on which many manipulation tasks depend.

“In SimplerEnv, DGA additionally receives a wrist-depth view unavailable to the baseline policy, so the comparison does not isolate the proposed correction mechanism from the benefit of additional sensing.”

Additional sensing in SimplerEnv. In SimplerEnv, wrist depth is an additional input to the assistance module, while the VLA's observation inputs and parameters remain unchanged. The comparison with the vanilla policy measures the overall benefit of this augmentation. To examine the contribution of the correction mechanism, Table 3 compares alternatives with the same wrist-depth access and triggering conditions: DGCS delays closure, DGCR immediately replans, and DGA-1Step applies one geometric correction.

Thank you for the response. It does not change my assessment.

My main concern stands. A VLA is a general-purpose policy, and grasping is only one of the things it does. The method adds a hand-calibrated grasp prior on top of it, and the evidence is improvement on grasping tasks that were selected and calibrated for that prior. This does not show that the wrapper is a net benefit to the policy. The paper claims the policy's general behavior is preserved, but this is never measured.

  1. "When an exact category entry is unavailable, we select the closest available category with compatible grasp geometry." Sec. 3.5 only describes reuse for the same object category and task instruction, and every evaluated task had its own calibration. How is the closest category chosen, and by whom? Was this used in any reported experiment? Regardless of the answer, I see no reason to believe that a grasp geometry chosen per category is the right one for a given interaction. Fixing it in advance defeats the purpose of using a VLA.
  2. "Non-grasp closure commands remain under VLA control." Sec. 3.2 and Algorithm 1 trigger on any upcoming close command. The response says grasps are "identified from the task instruction," but this is not specified anywhere. If it is a per-task switch, it is a hand-set rule and not general, and it cannot tell a grasp from a non-grasp closure within the same task. How does the system know when the VLA is closing the gripper for something other than a grasp?

I am keeping my rating.

Thank you for following up and clarifying your concerns about how grasp assistance affects the broader policy.

1. Generality and grasp performance

A VLA is a general-purpose policy, and grasping is only one of the things it does.

We agree that a calibrated grasp controller has a narrower scope than the VLA it assists. We see a trade-off between generality and task-specific performance. As VLAs become more general, improving their performance on recurring tasks such as grasping remains an important direction. As the title indicates, our focus is grasp enhancement. Grasping is a recurring step in manipulation, and grasp reliability matters in real deployments. Although VLAs have shown an ability to generalize, we observed many failed grasps in which the gripper reached the target object but was misaligned at closure. Our method offers a simple, efficient way to improve grasp performance in those settings at low adaptation cost, while retaining the VLA's ability to handle the broader task.

The paper claims the policy's general behavior is preserved, but this is never measured.

By “preserving the general behavior,” we mean that the VLA still determines how to approach the object and how to proceed after local correction. Across successful episodes on the eight RoboTwin tasks in Table 1, DGA supplied an average of 5.4 of 183.7 control steps (2.9%). The VLA thus retained control for most of each episode, with DGA refining only the local grasp motion.

“the evidence is improvement on grasping tasks that were selected and calibrated for that prior.”

We have added Appendix E to report results for all 50 RoboTwin tasks. For π₀.₅, mean task success increased from 44.38% to 48.46% in clean scenes and from 47.90% to 52.82% in randomized scenes.

For LIBERO, we evaluated all 40 tasks without filtering for grasping tasks. Alongside pick-and-place tasks, the benchmark includes a small number of tasks involving other interactions. On “turn on the stove” and “push the plate to the front of the stove,” π₀.₅ completed all 50 trials of each task both with and without DGA. On “turn on the stove and put the moka pot on it,” the policy completed the full instruction in 45 of 50 trials without DGA and 50 of 50 with DGA.

Taken together, these results indicate that the wrapper is a net benefit to the policy in the evaluated settings.

2. Reference selection

How is the closest category chosen, and by whom? Was this used in any reported experiment?

We did not use closest-category matching in the reported experiments. Our earlier reply described a possible deployment option.

“I see no reason to believe that a grasp geometry chosen per category is the right one for a given interaction.”

In settings such as factories, a limited set of interactions may recur. When the same grasp is needed again, the robot can reuse a reference obtained from demonstrations without retraining the VLA. Our experiments show that reusing the same calibrated grasp geometry across trials of a given task can improve task success.

3. Closure intent

How does the system know when the VLA is closing the gripper for something other than a grasp?

Section 3.2 and Algorithm 1 describe how a close command triggers local correction once grasp assistance is enabled. In the evaluation, we used keywords in the task instruction to distinguish interaction types. For example, RoboTwin instructions for picking up two bottles use verbs such as “pick up” and “grasp,” while instructions for ringing a bell use “click” and “press.” When a task involved both grasp and non-grasp closures, we used the closure order recorded during calibration to determine which closures were eligible for assistance.

Comments on the paper

1 comment

We thank the reviewer for recognizing the practical value of DGA-VLA and the evidence from our paired real-robot evaluation.

1. Methodological insight

“Although this integration is practically useful, the paper does not sufficiently establish what new methodological insight it offers beyond adapting familiar control techniques to frozen VLA execution.”

Our key insight is to exploit the complementary strengths of learned policies and geometric feedback through a hybrid control framework. The VLA provides task understanding and approach motion, while local feedback supplies the precision needed at grasp closure. This collaboration offers a lightweight way to improve grasp execution while retaining the VLA's broader capabilities. Guided by the stage-specific characteristics of manipulation tasks and shortcomings in VLA grasp precision, we design the depth-based localization and geometric correction components specifically for this collaboration, drawing on the principle of visual servoing.

2. Further clarification of calibration (Q1 and Q2)

“How are anchors selected across successive grasps? Does this require predefined object identities or task-stage information, and are calibration and evaluation configurations disjoint?”

Anchor selection and calibration/evaluation separation (Q1). We select references using the object category in the task instruction and the grasp index for successive grasps, tracked from the VLA's gripper-closing intent. Calibration and evaluation use non-overlapping initial scene configurations. In LIBERO and SimplerEnv, calibration uses a successful baseline-policy trajectory collected separately from evaluation. RoboTwin uses ten successful expert demonstrations per task from the official repository.

“Do fixed anchors remain effective for new object instances or orientations without recalibration?”

Reference reuse (Q2). Calibration parameters are reused by object category and grasp affordance, based on shared local grasp geometry. DGA uses depth-derived geometry, allowing instances with different colors or textures to share a reference when their grasp configurations are similar. In our evaluation, LIBERO uses initial states with varying object positions and orientations, while RoboTwin includes randomized within-category object variants, positions, and orientations; anchors remain fixed across evaluation episodes. The VLA accommodates orientation changes through its approach to the intended grasp region, while DGA performs final local alignment in the resulting wrist view. Objects with substantially different shapes or grasp configurations can be handled by adding corresponding references to the library.

“Consequently, it remains unclear how much of the reported benefit transfers beyond the calibrated settings and how much task-specific preparation deployment requires.”

Calibration cost in deployment. VLA deployment often involves collecting demonstrations for task-specific fine-tuning. DGA can reuse a small subset of these data to calibrate grasp references through lightweight geometric processing. This makes calibration a small addition to the existing adaptation workflow and offers a low-cost way to improve grasp precision compared with methods that require additional fine-tuning or training.

3. LIBERO pairing and uncertainty estimates (Q3)

“Are LIBERO trials paired across methods, and can you provide uncertainty estimates for the small reported gains?”

Yes, LIBERO trials are paired across methods, with identical initial states for corresponding episodes. Using the original evaluation records for π0.5\pi_{0.5} and OpenVLA, we computed two-sided 90% confidence intervals for the average success-rate improvements across the four suites. Each model has 40 tasks with 50 paired trials per task (2,000 pairs in total):

ModelBaselineDGAImprovement (pp)90% paired bootstrap CI (pp)
π0.5\pi_{0.5}96.65%97.40%+0.75[0.10, 1.40]
OpenVLA75.75%76.90%+1.15[0.20, 2.10]

Confidence intervals were computed using a task-stratified paired percentile bootstrap, preserving baseline–DGA pairing and equal task weights.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.