acceptodds
Withdrawn from ICLR 2027

RoboAct-WorldVLA: Reinforcing Action Generation with a Latent Action-Understanding World Model

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

13 views

Abstract

Although Vision–Language–Action (VLA) models have shown strong potential in robotic manipulation, their imitation-driven nature still limits robust deployment under domain shifts. The core issue is not merely insufficient demonstrations, but the lack of inherent physical action understanding and the absence of a framework that uses such understanding to guide action generation. As a result, a VLA policy may produce actions that resemble demonstrations without knowing whether these actions move the system towards the goal or not. In this regard, we propose RoboAct-WorldVLA, a framework that reinforces action generation with a latent action-understanding world model. Given candidate actions from a flow-based generator or an off-the-shelf VLA policy, the world model predicts their future action representations in a pretrained disentangled latent space. A language-aligned action-value evaluator assesses how well the predicted future representation aligns with the prior action representations, suppressing uncertain or off-trajectory actions. The prior action representations are generated by an existing action representation encoder RoboAct-CLIP, which disentangles the inherent action dynamics from subject/object features for better generalizability. RoboAct-WorldVLA is first pretrained offline from demonstrations, failed trajectories, and counterfactual transitions. It is then fine-tuned online with environment rewards, which refine the action generator and calibrate the evaluator while preserving imitation, latent dynamics, semantic alignment, and uncertainty regularization. We evaluate RoboAct-WorldVLA on RLBench, ManiSkill3, and real-world manipulation tasks. Extensive experiments demonstrate that RoboAct-WorldVLA can consistently boost the performance of VLA backbones by 3.8–9.0 percentage points. Notably, the ablation study reveals the proposed method's active capability to determine required causal actions and evaluate current progress, guided by prior action representations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot manipulation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_C1VuOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RoboAct-WorldVLA, which couples a frozen RoboAct-CLIP action encoder with a latent world model and a language-aligned evaluator to predict and score the consequences of candidate action chunks before execution. A policy proposes K candidates, the world model predicts their latent futures, and the evaluator ranks them. Training is offline pretraining followed by online RL. Experiments on RLBench, ManiSkill3, and real robots report gains across several VLAs.

Critical strengths
  1. Constraining the representation at the input level (CLIP encoder) makes sense.
  2. The method attempts to directly adapt to existing VLA models, lowering the barrier to use.
Critical weaknesses
  1. The motivation lacks rigor and is disconnected from the paper's central robustness claim. The intro attributes heterogeneous problems (domain shift, partial task completion, unseen object configurations, execution failures) to a single cause ``actions not being understood''. This may be an overly absolute and unsupported causal link that leaves the motivation underdeveloped. Moreover, despite being framed around robustness, the paper designs no experiment to test whether the method actually improves robustness, and the writing does not establish this connection, leaving motivation, method, and evaluation disconnected.
  2. The method lacks focus and insight, and the presentation is underpolished. The approach reads as a stack of off-the-shelf components (robust encoder, robot online RL, WM-based RL, etc.) trained with online RL. Each of them is a substantive research question in its own right, yet none is examined in depth, leaving the central contribution unclear. This is compounded by poor figure legibility. Understanding the method requires extensive carefully reading of the main text and even the appendix, and an anomalous black border in Figure 2, giving a concern of insufficient writing or polish.
Questions for authors

Please see the Critical Weaknesses above.

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

The paper proposes RoboAct-WorldVLA, a generate-then-select framework that augments a VLA action generator with a latent action-understanding world model and a learned evaluator. Candidate action chunks are proposed by a flow-based policy or an off-the-shelf VLA, a world model predicts their future outcomes directly in a frozen, disentangled action-representation space (RoboAct-CLIP) plus visual features, and an evaluator scores candidates using learned value, language-aligned semantic progress, feasibility, uncertainty, and action regularization. The system is pretrained offline on demonstrations, failures, and counterfactuals, and then refined online with environment rewards; experiments on RLBench, ManiSkill3, and real robots report consistent gains over baselines and across multiple VLA backbones.

Critical strengths
  1. Predicting candidate-conditioned consequences directly in a disentangled, action-centric latent space (RoboAct-CLIP) instead of pixel space is a clear and well-motivated design that reduces unnecessary appearance modeling while retaining task-relevant dynamics. The framework can be integrated with different action generators.

  2. Ablations isolate key components: world model and predicted-future signals versus evaluator-only critic, role of online fine-tuning, importance of RoboAct-CLIP alignment, uncertainty penalty, and candidate budget trade-offs. Diagnostic analyses (cosine predictive fidelity versus copy-current baselines, evaluator accuracy, case study showing correct and wrong placement ranked appropriately) support the claimed mechanisms. I appreciate these designs.

Critical weaknesses
  1. I found the method section unnecessarily difficult to follow. The definition of semantic progress, uncertainty estimation, evaluator training and policy optimization are all deferred to Appendix A. Equations (2)-(3) therefore appear before enough information is given to understand how the scores and reward are obtained. These choices determine how actions are selected and reinforced, and the essential definitions should be in the main text. Reading the appendix does not fully resolve this: the computation and supervision of sfeass_\text{feas}, the construction of counterfactual training examples, and their good/bad labels are still not specified concretely.

  2. The experiments do not yet make the source and scope of the improvement clear. The VLA policies are also fine-tuned, but Tables 2-3 do not establish that the baselines receive comparable adaptation and online interaction budgets. Table 4 is helpful, but a comparison with an established chunk critic such as VGAS[1] or test-time verification methods with rejection sampling such as TACO[2] would test the benefit of future prediction more directly. Figure 3 offers limited further support: Appendix C uses constructed waypoint candidates and a diagnostic JlogJ_\text{log}, rather than the full selection score in Eq. (2). Absolute inference latency is also missing, making the cost of candidate evaluation difficult to assess. Finally, the claims about domain shifts and failure recovery deserve more targeted evaluation. Results on LIBERO or RoboTwin 2.0, including instruction changes or perturbed initial states, would be more convincing than additional nominal-task averages.

Reference

[1] Xu, Changhua, et al. "VGAS: value-guided action-chunk selection for few-shot vision-language-action adaptation." arXiv preprint arXiv:2602.07399 (2026).

[2] Yang, Siyuan, et al. "Steering vision-language-action models as anti-exploration: A test-time scaling approach." arXiv preprint arXiv:2512.02834 (2025).

Questions for authors
  1. How are counterfactual training candidates and their labels obtained? Please give the actual definition and training target of sfeass_\text{feas}, and specify which experiments use the semantic reward term in Eq. (14).

  2. Can the gains from extra policy fine-tuning and candidate selection be separated using the same adapted policy? Please report the online interaction budgets, ManiSkill/real-robot evaluation counts and uncertainty, and absolute control-step latency.

  3. The motivation appears closely related to PRTS[3], which learns goal-reachability-aware representations through contrastive RL during VLA pretraining. How do the authors view the trade-off between that approach and keeping a world model and evaluator active at inference? What evidence shows when explicit future prediction is preferable to directly learning a goal-conditioned value representation?

Reference

[3] Zhang, Yang, et al. "PRTS: A Primitive Reasoning and Tasking System via Contrastive Representations." arXiv preprint arXiv:2604.27472 (2026).

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

This paper proposes a framework that enhances action generation through a latent action-understanding world model. The framework predicts the consequences of candidate actions generated by a VLA policy, evaluates their estimated quality, and selects a candidate for execution. Experiments on simulation benchmarks and real-world manipulation tasks demonstrate improvements in task success rates.

Critical strengths
  1. The system design is relatively comprehensive, integrating a latent world model, an action-value evaluator, and a pretrained representation encoder. The framework supports closed-loop execution and combines offline training with online policy and evaluator updates.

  2. The evaluation covers two simulation benchmarks and two real-world robotic platforms, providing evidence of the framework’s applicability across multiple settings.

Critical weaknesses
  1. Insufficient evidence for robustness under domain shifts. While the reported success rates demonstrate performance improvements, the paper motivates its approach partly through the limitations of VLA policies under domain shifts. The current experiments do not clearly isolate such shifts or establish how much the proposed method improves robustness to them. Evaluation across different platforms alone does not establish robustness under a controlled distribution shift.

  2. Limited validation of the claimed physical action understanding. The paper argues that the latent world model addresses the lack of inherent physical action understanding. However, the proposed latent representations and learned evaluator remain difficult to interpret. The feature-prediction diagnostics and component ablations provide useful evidence, but they do not sufficiently establish that the gains arise from improved physical action understanding. More targeted tests of action preconditions, physical constraints, and consequences would strengthen this claim. In addition, the ManiSkill evaluation includes only OpenVLA, Dita, and CogACT. Including an additional strong baseline such as π₀.₅ would help assess whether the improvements extend to more capable policies.

  3. The performance–latency trade-off requires stronger justification. Table 4 indicates that the full system incurs approximately twice the latency of direct policy execution. Although the success-rate gains are meaningful, the current evaluation does not sufficiently demonstrate when this additional computational cost is justified. The real-world tasks appear relatively limited in complexity. More challenging long-horizon tasks, or tasks involving execution disturbances and recovery from intermediate failures, would better demonstrate the practical value of predicting and evaluating action consequences.

  4. Insufficient implementation details for the policy used in Table 1. The proposed method is listed only as “RoboAct-WorldVLA.” Although the text identifies its action generator as a flow-based policy and provides training objectives, the generator’s architecture, model size, initialization, and detailed training configuration are not sufficiently specified. This makes it difficult to assess reproducibility and attribute the reported performance gains. Please provide these implementation details, together with results for the same generator without the proposed world model and evaluator across all ten tasks.

  5. A dedicated domain-shift evaluation would strengthen the central motivation. Consider evaluating on LIBERO-Plus or using a clean-to-randomized evaluation protocol in RoboTwin, with the training and evaluation distributions explicitly specified. Such experiments would more directly test whether the proposed framework improves robustness to changes in object appearance, scene configuration, or other environmental factors.

Questions for authors
  1. Insufficient evidence for robustness under domain shifts. While the reported success rates demonstrate performance improvements, the paper motivates its approach partly through the limitations of VLA policies under domain shifts. The current experiments do not clearly isolate such shifts or establish how much the proposed method improves robustness to them. Evaluation across different platforms alone does not establish robustness under a controlled distribution shift.

  2. Limited validation of the claimed physical action understanding. The paper argues that the latent world model addresses the lack of inherent physical action understanding. However, the proposed latent representations and learned evaluator remain difficult to interpret. The feature-prediction diagnostics and component ablations provide useful evidence, but they do not sufficiently establish that the gains arise from improved physical action understanding. More targeted tests of action preconditions, physical constraints, and consequences would strengthen this claim. In addition, the ManiSkill evaluation includes only OpenVLA, Dita, and CogACT. Including an additional strong baseline such as π₀.₅ would help assess whether the improvements extend to more capable policies.

  3. The performance–latency trade-off requires stronger justification. Table 4 indicates that the full system incurs approximately twice the latency of direct policy execution. Although the success-rate gains are meaningful, the current evaluation does not sufficiently demonstrate when this additional computational cost is justified. The real-world tasks appear relatively limited in complexity. More challenging long-horizon tasks, or tasks involving execution disturbances and recovery from intermediate failures, would better demonstrate the practical value of predicting and evaluating action consequences.

  4. Insufficient implementation details for the policy used in Table 1. The proposed method is listed only as “RoboAct-WorldVLA.” Although the text identifies its action generator as a flow-based policy and provides training objectives, the generator’s architecture, model size, initialization, and detailed training configuration are not sufficiently specified. This makes it difficult to assess reproducibility and attribute the reported performance gains. Please provide these implementation details, together with results for the same generator without the proposed world model and evaluator across all ten tasks.

  5. A dedicated domain-shift evaluation would strengthen the central motivation. Consider evaluating on LIBERO-Plus or using a clean-to-randomized evaluation protocol in RoboTwin, with the training and evaluation distributions explicitly specified. Such experiments would more directly test whether the proposed framework improves robustness to changes in object appearance, scene configuration, or other environmental factors.

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

Related papers

Discussion (0)

Sign in to comment.