acceptodds
Under review as a conference paper at ICLR 2027

Rolebind: training vision-language-action policies with task-conditioned role–instance supervision

Abstract

Vision-Language-Action (VLA) models have demonstrated strong capabilities in general-purpose robotic manipulation by jointly modeling visual observations, language instructions, and robotic actions. However, existing VLA models primarily rely on action supervision for policy learning, and there is a lack of explicit constraints on the correspondence between task semantics and physical entities. Robotic manipulation requires not only the identification of relevant entities but also an understanding of their roles in the current task. To address this issue, we propose RoleBind—a training method for implementing the binding of task roles to object instances within existing VLA policies. In addition to the original action learning objective, RoleBind introduces role-instance binding supervision, which explicitly aligns task roles (such as manipulated objects and target objects) with their corresponding visual instances in a shared representation space, thereby enhancing the VLA’s representation of task-relevant entities and their role relationships. Role labels and instance region information are used only during the training phase. We applied RoleBind to different action modeling methods on LIBERO and further evaluated its generalization ability in out-of-distribution scenarios. Experimental results show that RoleBind increases the average success rate of StarVLA-OFT from 98.0% to 98.55%, an improvement of 0.55 percentage points, and yields a further performance improvement of 0.8 percentage points on StarVLA-. These results demonstrate that RoleBind can improve task success rates without altering the original inference process, and also provide a viable approach for explicitly linking task semantics to physical entities in VLA.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models,Robot Manipulation,Task-Role Binding,Visual Grounding

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_LEZFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes RoleBind, a training-time auxiliary supervision that explicitly binds task roles (manipulated / target object) to their corresponding visual instances via a contrastive loss, while leaving the base VLA's architecture and inference unchanged. On LIBERO it raises StarVLA-OFT from 98.00% to 98.55% and StarVLA-π from 97.50% to 98.30%. The method requires role labels and instance masks only during training.

Critical strengths
  1. By directly supervising the correspondence between task semantics and physical entities, RoleBind improves VLA accuracy across two different VLA models.

  2. The additional supervision is used only during training and has zero impact on inference: the binding branch, projectors, and masks are all removed at deployment, with no segmentation model and no change to the input–output interface.

Critical weaknesses
  1. Lack of real-world experiments, there is a huge sim to real gap between the real world and simulation, and experiments must be conducted on real robots to verify the effectiveness of your method.

  2. The paper compares only against its own base model and lacks comparisons with related auxiliary-supervision baselines under matched training budgets. And there are many formatting issues with the article, e.g., in Tables 2 and 3 the success-rate values are not aligned in decimal places (98 vs 98.55, 75 vs 75.83); please correct this in the final version.

Questions for authors
  1. Only two static roles (manipulated / target object) are modeled, and evaluation is limited to relatively simple tabletop tasks. In long-horizon, multi-object datasets such as BEHAVIOR-1K, activities like "BakingSugarCookies" involve ingredients, a mixer, and an oven whose roles change across stages. How can your method handle this problem?

  2. On real-robot data, masks and instance-level information needs manual annotation and additional detection/segmentation, which cannot be obtained by simulator replay. How do you consider this additional cost?

Reviewer_vxi7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RoleBind, a training-time supervision method to improve task-conditioned visual grounding in VLA models. It introduces an auxiliary contrastive loss to align task-specific semantic roles with their corresponding visual object instances. However, experiments on LIBERO and LIBERO-Plus show only marginal improvements over existing VLA baselines, despite introducing no additional inference overhead.

Critical strengths
  1. The paper addresses task-conditioned visual grounding in VLA models by explicitly aligning task-specific object roles with their corresponding visual instances.

  2. The proposed method is simple and can be easily incorporated into existing VLAs without introducing additional inference overhead.

Critical weaknesses
  1. The performance gain is marginal. The overall improvements on LIBERO and LIBERO-Plus are less than 1%. Further evaluation on more challenging benchmarks is needed to better demonstrate the effectiveness of the proposed method.

  2. The technical contribution is limited. The proposed method mainly introduces an auxiliary contrastive loss to align task-specific role representations with corresponding visual object instances. The novelty beyond existing visual grounding approaches remains limited.

  3. No comparisons with existing baselines for visual grounding. This paper only reports the performance against the original VLA baselines. Comparisons with existing visual grounding or auxiliary supervision methods is needed.

Overall, I do not think this paper meets the standards of ICLR, given its limited technical novelty and insufficient experimental validation.

Questions for authors

Please refer to the weaknesses.

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

RoleBind supplements VLA action learning with a training-only contrastive objective aligning manipulated-object and target-object descriptions with visual features pooled from simulator instance masks. The auxiliary branch is removed at inference. Using a shared Qwen3-VL backbone with regression and flow-matching action heads, it reports average improvements of 0.55/0.8 percentage points on LIBERO and 0.83/0.96 points on LIBERO-Plus, respectively.

Critical strengths
  • Explicit supervision connecting task roles to physical instances addresses a relevant gap between action imitation and language grounding. Combining role descriptions with mask-pooled visual features provides a simple auxiliary objective, while removing the branch at inference preserves the original policy interface and action-generation process. This is practically appealing when training-time annotations are available.
  • Compatibility with both regression and flow-matching action heads is tested under a shared backbone, with matched base checkpoints, data, and training schedules. Positive aggregate results on both LIBERO and LIBERO-Plus provide initial evidence beyond one decoder or evaluation condition. The swapped-binding comparison also usefully tests sensitivity to supervision correctness, although it does not fully isolate role semantics.
Critical weaknesses
  • The central distinction between task-role binding and object-name grounding is not isolated. In Sec. 3.2, each description contains both a role and an object name, so name-to-region matching could explain the benefit. Table 5 compares correct and swapped binding (98.55% versus 97.85%), but corrupted correspondence is not a clean object-grounding control; the exact swapped descriptions are also unspecified. Please compare against object-name-only alignment using the same masks, projectors, loss, and training budget, and clarify the swapped text-region pairs.
  • The reliability of the small performance differences remains unclear. Tables 2-3 report gains of 0.55/0.8 percentage points on LIBERO and 0.83/0.96 points on LIBERO-Plus, without uncertainty estimates or independently trained seed results. Sec. 4.1 specifies 50 trials per task and shared initial states, which is helpful, but evaluation repetitions do not establish training stability. Please report independent-seed comparisons with paired, task-aware uncertainty and success counts, and specify the checkpoint-selection criterion and evaluation split. These details are consequential near saturation, where a few changed outcomes or checkpoint selection can materially affect the reported gain.
Questions for authors
  • Does RoleBind outperform an otherwise matched object-name-only alignment control, and exactly which text descriptions are paired with which regions in the swapped variant, including whether object-name-to-region correctness is preserved?
  • Do the improvements persist across independent training seeds, with paired uncertainty estimates and success counts, and what checkpoint-selection criterion and evaluation split produced the reported results?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.