acceptodds
Under review as a conference paper at ICLR 2027

WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction

11 views

Abstract

Vision–language–action (VLA) policies excel at generalist robotic manipulation but generate actions without explicitly modeling their environmental consequences. World action models (WAMs) incorporate future prediction, yet most rely on costly video generation, while latent alternatives either use predicted futures only as auxiliary targets or infer latent actions from suboptimal visual transitions through costly large-scale video pretraining. We introduce **WAVE-0** (**W**orld-model-**A**ugmented **V**LA **E**xecution), a framework that bridges VLAs and WAMs through a lightweight latent world model. By connecting the semantic understanding of pretrained VLAs with explicit future prediction, WAVE-0 allows anticipated action consequences to directly guide action generation. Built on a pretrained VLA, its 257M world model translates VLM-generated latent action plans into predicted future representations that condition the action expert, unifying latent planning, future prediction, and action generation. The world model is trained directly on ground-truth robot actions with a counterfactual ranking objective, requiring neither pixel reconstruction nor large-scale video pretraining. WAVE-0 achieves success rates of 91.47% on RoboTwin 2.0 and 98.7% on LIBERO, along with 88.0%, 61.4%, and 75.0% on LIBERO-Plus, LIBERO-Pro, and LIBERO-Para, respectively. On real-world tasks, it improves the success rate of from 38.3% to 73.3%. These results demonstrate strong manipulation performance and generalization across diverse distribution shifts. Code will be released.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Vision-Language-Action Policies, Latent World Models

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_tbC5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WAVE-0 augments a pretrained π0.5\pi_{0.5} policy (Physical Intelligence et al., 2025) with a latent world model that predicts future visual representations and uses them to condition action generation. In Stage 1, the world model and action/state encoders are trained on robot demonstrations in a frozen SigLIP-based representation space (Zhai et al., 2023), using future prediction, residual regression, and a counterfactual action-ranking objective. In Stage 2, the world model is frozen, the VLM learns latent action plans aligned with the action encoder's representations, and predicted future features are compressed into conditioning tokens for the action expert (Sec. 3.2, Fig. 2). The paper evaluates the approach on LIBERO, RoboTwin 2.0, several LIBERO generalization benchmarks, and three real-world tasks, reporting improvements over π0.5\pi_{0.5} with modest additional inference latency (Tables 1–3 and 11).

References

Critical strengths
  1. An action-ranking objective with evidence of downstream value. The counterfactual ranking loss is the most interesting methodological component to me. It targets a concrete failure mode: a latent predictor may rely heavily on the current observation while making insufficient use of the action input. By keeping the observation fixed and contrasting demonstrated and mismatched actions, the objective explicitly encourages action sensitivity (Sec. 3.2, Eq. (7)). The reported ablation provides evidence at both the prediction and policy levels: at 20k training steps, the action-sensitivity gap increases from 0.0020 to 0.1648, and adding the ranking loss improves success by 2.6 and 3.0 percentage points on LIBERO-Plus and LIBERO-Pro, respectively (Sec. 4.7.2, Fig. 4(a–c)). The downstream results make this component consequential to my assessment: they connect the proposed training intervention to improved robot performance and motivate further investigation of its role in learning action-dependent predictions.

  2. Practical improvements to a pretrained VLA with modest additional inference cost. The method provides a concrete way to incorporate predicted visual consequences into an existing π0.5\pi_{0.5} policy (Physical Intelligence et al., 2025), while retaining its action-generation framework and avoiding pixel reconstruction or additional large-scale video pretraining for the world model (Sec. 3.2). The reported gains extend across standard manipulation and distribution-shift evaluations: RoboTwin success improves from 80.38% to 91.47%, and LIBERO-Pro success improves from 53.0% to 61.4% (Tables 1(b) and 2). In the real-world inference configuration, mean server latency increases from 287.1 ms to 305.9 ms, approximately 6.5% (Appendix A.2, Table 11). These results support the practical relevance of the design: the reported performance benefits include challenging generalization settings, while the added inference cost relative to the base policy is limited.

References

Critical weaknesses
  1. The contributions of the coupled design choices are insufficiently isolated. The framework combines latent action queries, action-representation alignment, multiple auxiliary losses, future-conditioning tokens, and several conditioning inputs. The paper includes useful integration and loss ablations, but these do not establish which parts of this combination are essential. In particular, the auxiliary-only world-model variant already improves RoboTwin success from 80.38% to 89.51%, while the full method reaches 91.47% (Sec. 4.7.1, Table 1(b)). The additional 1.96-point gain supports the full design, but does not isolate the value of the predicted future content from the other changes in conditioning and training. Controls using current visual features or latent action tokens in place of predicted future tokens are missing, as is an ablation of the separate proprioceptive input to the world model. Consequently, I remain uncertain whether the improvements substantiate the proposed mechanism and justify the framework's additional complexity.

  2. The distinctive methodological contribution over prior latent world-action models is not sufficiently established. The paper itself identifies LaWAM (Chen et al., 2026) as already conditioning action generation on predicted latent visual subgoals (Sec. 2.3). Direct action supervision and the ranking objective could provide a meaningful distinction, but the latter is underspecified and its dynamics interpretation is only partially validated. Equations (6)-(7) introduce mismatched actions without explaining their construction or how distinct actions with similar visual consequences are handled. Moreover, increasing the distance between the negative prediction and the demonstrated future does not establish that the negative prediction accurately represents the alternative action's outcome. The diagnostics in Fig. 4(a–b) measure the separation encouraged by the ranking loss, and Fig. 3(b) visualizes policy attention rather than probing the world model's representations. Despite the reported performance gains, the current evidence leaves the main new methodological insight insufficiently characterized.

References

Questions for authors
  1. Clarification: definition and validity of mismatched actions. How exactly is the mismatched sequence in Eqs. (6)-(7) constructed: from another trajectory or task, by temporal permutation, or through action perturbations? How is the mask rr computed? How are distinct action sequences that produce equivalent or nearly identical visual outcomes handled? This is important because action inequality alone does not guarantee that the observed future is an inappropriate target for the alternative action.

  2. Evidence request: attribution of the future-conditioning benefit. Have you compared predicted future tokens against current-observation tokens or directly projected latent action tokens, while keeping the remaining architecture, auxiliary supervision, token budget, and training budget as comparable as possible? Such controls would help determine whether accurate future prediction is necessary for the gain over the auxiliary-only world-model baseline in Sec. 4.7.1.

  3. Evidence request: necessity of the separate state-conditioning path. Section 3.2 and Eq. (4) condition the world model on zs=Es(st)z_s = E_s(s_t), while the action encoder already receives the same proprioceptive state through Za=Ea(At,st)Z_a = E_a(A_t, s_t). Have you removed the separate zsz_s input while retaining sts_t in the action encoder? How does this affect future prediction, action sensitivity, and downstream success? This would clarify whether the additional state pathway provides information that the action representation does not already supply.

  4. Evidence request: dependence on the prediction space. Have you compared the frozen SigLIP-based prediction space (Zhai et al., 2023) described in Sec. 4.1 with an alternative such as DINO (Caron et al., 2021), using appropriate dimensional adapters and comparable training settings? In particular, how much do future prediction and policy performance depend on reusing π0.5\pi_{0.5}'s visual representation space? This would clarify how specific the reported benefit is to the selected visual representation.

  5. Evidence request: potential distribution shift in world-model conditioning. In Stage 1, the world model is trained on action-encoder outputs, including encodings of demonstrated and mismatched action sequences. In both Stage 2 training and deployment, the frozen world model instead receives planner-generated latent actions (Sec. 3.2, Eqs. (4), (6), and (9)). What evidence shows that the alignment and future-prediction objectives in Eqs. (10)–(12) adequately address this potential mismatch? Specifically, on held-out trajectories, how do future-prediction errors compare under encoded demonstrated actions versus planner-generated latents, with current observations and robot states held fixed? Does this gap increase on generalization benchmarks, and is it associated with degraded policy performance?

  6. Clarification of evidence and scope: Stage 1 scaling. Are there controlled studies varying the amount or diversity of Stage 1 training data, with Stage 2 training held fixed, and measuring both prediction quality and downstream success? Section 4.6 compares training budgets, and Fig. 4(a–b) varies optimization steps; neither establishes scaling with dataset size. Please clarify whether such results are available. I would treat a new, extensive study on larger datasets as future work; the immediate question is what evidence currently supports the data-efficiency and generality of the learned world model.

References

11 comments

Dear Reviewer tbC5,

Thank you for your thorough review and constructive feedback. We appreciate your recognition of our counterfactual action-ranking objective as a consequential methodological contribution that mitigates observation-only shortcuts and improves both Stage-1 action sensitivity and downstream OOD success, as well as WAVE-0's practical gains on a pretrained VLA with modest (~6.5%) inference overhead without pixel reconstruction or large-scale video pretraining. Below, we (1) isolate the value of predicted future tokens via matched non-predictive, direct-query, and current-feature controls, (2) ablate the separate proprioceptive state pathway in the world model, (3) establish our distinction from prior latent WAMs with a matched Video-IDM comparison and specify the construction of mismatched actions and the validity mask in Eqs. (6)–(7), (4) compare π0.5's built-in SigLIP encoder against a frozen DINOv3 visual encoder, (5) diagnose Stage 1 to Stage 2 conditioning alignment, held-out OOD prediction errors, and rollout failures, and (6) present controlled Stage-1 scaling experiments on demonstration count and task coverage.

For Weakness 1: "The contributions of the coupled design choices are insufficiently isolated. The framework combines latent action queries, action-representation alignment, multiple auxiliary losses, future-conditioning tokens, and several conditioning inputs. The paper includes useful integration and loss ablations, but these do not establish which parts of this combination are essential. In particular, the auxiliary-only world-model variant already improves RoboTwin success from 80.38% to 89.51%, while the full method reaches 91.47% (Sec. 4.7.1, Table 1(b)). The additional 1.96-point gain supports the full design, but does not isolate the value of the predicted future content from the other changes in conditioning and training. Controls using current visual features or latent action tokens in place of predicted future tokens are missing, as is an ablation of the separate proprioceptive input to the world model. Consequently, I remain uncertain whether the improvements substantiate the proposed mechanism and justify the framework's additional complexity."

For Question 2: "Evidence request: attribution of the future-conditioning benefit. Have you compared predicted future tokens against current-observation tokens or directly projected latent action tokens, while keeping the remaining architecture, auxiliary supervision, token budget, and training budget as comparable as possible? Such controls would help determine whether accurate future prediction is necessary for the gain over the auxiliary-only world-model baseline in Sec. 4.7.1."

First, π0.5 + WM and WAVE-0 in Table 1(b) are two parallel ways of integrating our proposed Stage-1 world model into π0.5 (Fig. 1(3) vs. Fig. 1(4)) rather than an incremental stack. We designed π0.5 + WM to place the world model after the action expert for training loss supervision only, mirroring the auxiliary-supervision paradigm of VLA-JEPA and JEPA-WAM, so that we can directly compare that paradigm against WAVE-0's pre-expert future conditioning on the exact same backbone and world model. Beyond the +1.96 lead on RoboTwin 2.0, our new evaluation of π0.5 + WM on LIBERO variants conducted during the rebuttal shows clear gains from pre-expert conditioning under distribution shifts, improving LIBERO-Plus by +1.6 points from 86.4% to 88.0% and LIBERO-Pro by +2.6 points from 58.8% to 61.4%.

Second, to isolate the value of predicted future content from extra token capacity, latent action queries, and Stage-2 auxiliary losses, we conducted matched controls on the same π0.5 backbone, data, and training budget:

  • (a) π0.5 + WM + Non-predictive Tokens. Adds non-predictive conditioning tokens (8 tokens on LIBERO and 12 on RoboTwin, matching WAVE-0's conditioning-token budget) to π0.5 + WM.
  • (b) Direct Action Queries w/o WM. Projects VLM action-query states directly into matched conditioning tokens for the action expert without the world model.
  • (c) Current-Feature Conditioning. Conditions the action expert on resampled current visual features instead of predicted future features while retaining the exact same latent-action queries, resampler, and Stage-2 world-prediction and alignment losses.
MethodLIBEROLIBERO-PlusLIBERO-ProRoboTwin CleanRoboTwin RandomRoboTwin Avg.
π0.596.984.553.281.1279.6480.38
π0.5 + WM98.486.458.889.4889.5489.51
(a) π0.5 + WM + Non-predictive Tokens98.486.558.489.7889.2889.53
(b) Direct Action Queries w/o WM98.286.958.190.8490.5090.67
(c) Current-Feature Conditioning98.487.460.490.8290.0690.44
WAVE-0 (Ours)98.788.061.491.4691.4891.47

Adding non-predictive tokens in (a) brings virtually no gain over π0.5 + WM (86.5% vs. 86.4% on LIBERO-Plus and 89.53% vs. 89.51% on RoboTwin Avg.) and trails WAVE-0 by 1.5 points on LIBERO-Plus, 3.0 points on LIBERO-Pro, and 1.94 points on RoboTwin Avg. While VLM action-query routing in (b) already improves over π0.5, bypassing the world model still drops performance across every benchmark relative to WAVE-0, including −1.1 points on LIBERO-Plus, −3.3 points on LIBERO-Pro, and −0.80 points on RoboTwin Avg. Most directly, (c) Current-Feature Conditioning keeps all latent-action queries, alignment and world losses, and resampler parameters identical to WAVE-0, yet replacing predicted future features with current visual features consistently degrades performance across all benchmarks, including −1.03 points on RoboTwin Avg. (−1.42 points under randomization) and −1.0 point on LIBERO-Pro, isolating a consistent additional benefit from conditioning on predicted future content beyond query routing and auxiliary supervision alone.

For Weakness 1 (State-conditioning input) and Question 3: "Evidence request: necessity of the separate state-conditioning path. Section 3.2 and Eq. (4) condition the world model on zs=Es(st)z_s = \mathcal{E}_s(s_t), while the action encoder already receives the same proprioceptive state through Za=Ea(At,st)Z_a = \mathcal{E}_a(A_t, s_t). Have you removed the separate zsz_s input while retaining sts_t in the action encoder? How does this affect future prediction, action sensitivity, and downstream success? This would clarify whether the additional state pathway provides information that the action representation does not already supply."

Under identical training settings on LIBERO, we remove the separate state token zsz_s from the world model while retaining the proprioceptive state in the action encoder. Across sampled demonstration segments from all 40 LIBERO tasks, Stage-1 future prediction error and temporal-roll counterfactual action sensitivity compare as follows:

MethodFuture Cosine Error ↓Mismatched-Action ErrorAction Sensitivity
WAVE-0 w/o zsz_s7.91×10−47.91\times 10^{-4}0.06410.0633
WAVE-0 (Ours)7.68×10−47.68\times 10^{-4}0.05740.0566

Downstream policy success rates (%) across the three LIBERO benchmarks are:

MethodLIBEROLIBERO-PlusLIBERO-Pro
WAVE-0 w/o zsz_s98.587.259.2
WAVE-0 (Ours)98.788.061.4

Architecturally, the two pathways play distinct roles. In Stage 1, because the action encoder receives the proprioceptive state to compute relative action displacements and directly feeds the world model, removing zsz_s changes offline metrics only slightly, modestly increasing future prediction error from 7.68×10−47.68\times 10^{-4} to 7.91×10−47.91\times 10^{-4} while routing all state-action conditioning through ZaZ_a (0.0566 vs. 0.0633 sensitivity). More importantly, in Stage 2 the action encoder only provides alignment targets (Eq. 10) rather than feeding the world model (Eq. 9), and at deployment it is not used at all. Instead, the frozen world model receives VLM-generated latent action plans, while π0.5's VLM only observes coarsely discretized text tokens of proprioception rather than continuous joint states. Consequently, zsz_s serves as the only direct continuous proprioceptive input to the world model during Stage-2 future prediction and online deployment. Consistent with this design, removing zsz_s drops LIBERO by 0.2 points and causes larger drops under distribution shift on LIBERO-Plus (−0.8 points) and LIBERO-Pro (−2.2 points), where perturbed initial robot states, camera viewpoints, and object layouts require accurate continuous proprioceptive grounding.

Reviewer_g1GoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

WAVE-0 is a framework combining vision–language–action (VLA) policies with lightweight world action models (WAMs) for robotic manipulation. It predicts future consequences to guide actions without costly video generation or large-scale pretraining. Built on a pretrained VLA, WAVE-0 uses a 257M world model trained on robot actions with counterfactual ranking, achieving high success rates across benchmarks and improving real-world task performance significantly.

Critical strengths
  • The experiments are relatively sufficient, providing results on multiple datasets such as LIBERO, RoboTwin 2.0, LIBERO-Plus, and LIBERO-Pro.
  • The ablation study is relatively thorough, the writing is smooth, and the real-world experiments are relatively sufficient.
Critical weaknesses
  • Recently, many papers have claimed to combine world models, vision-language models, and action models to improve task success rates. This article is one such work. Although the introduction and related work sections compare these methods, a technical discussion is needed to clarify the fundamental differences between the proposed method in this paper and those approaches.
  • In terms of experiments: First, it is necessary to explain from the experimental results which design serves as the most critical performance-enhancing module in this paper. Second, the results on Robotwin2.0 are not very competitive; please elaborate on the main issues encountered with this dataset.
Questions for authors
  • Are the experimental results of other works compared on LIBERO-Pro re-implemented by the authors? Many papers do not have official result reports.
  • Lacks real device demo videos; although there are some screenshots, they are not intuitive.
7 comments

Dear Reviewer g1Go,

Thank you for your positive assessment and valuable suggestions. We appreciate your recognition of WAVE-0's lightweight 257M action-grounded world model that guides VLA action generation without costly video generation or large-scale pretraining, as well as our clear presentation, thorough ablations, and comprehensive evaluations across simulation and real-world benchmarks. Below, we (1) clarify WAVE-0's technical distinctions from recent WAMs, (2) identify the key performance-enhancing modules with new matched ablations and analyze efficiency and task-level bottlenecks on RoboTwin 2.0, and (3) detail the LIBERO-Pro baseline protocol and provide real-robot rollout and side-by-side comparison videos in the Supplementary Material.

For Weakness 1: "Recently, many papers have claimed to combine world models, vision-language models, and action models to improve task success rates. This article is one such work. Although the introduction and related work sections compare these methods, a technical discussion is needed to clarify the fundamental differences between the proposed method in this paper and those approaches."

As shown in Fig. 1 and discussed in Secs. 2.2 to 2.3, WAVE-0 differs from existing approaches in three technical aspects:

  1. Visual generation-based WAMs, including DreamZero, Motus, LingBot-VA, Fast-WAM, GigaWorld-Policy, AHA-WAM, BagelVLA, and ImageWAM, build on pretrained video or image generation backbones. Methods like Fast-WAM, AHA-WAM, and ImageWAM reduce inference cost by omitting future generation, reusing predictive features, or avoiding pixel decoding, but their training still relies on visual generative objectives. WAVE-0 instead learns action-conditioned dynamics directly in the frozen visual representation space of pretrained π0.5, using its SigLIP encoder and visual projector followed by pooling to a compact 4×4 grid per camera. Trained on demonstrated robot action sequences, an AdaLN-Zero Transformer predicts visual feature residuals without pixel reconstruction or additional image or video generative pretraining. The world model module only contains approximately 257M parameters.
  2. Among latent WAMs, VLA-JEPA learns from action-free videos, while LaWAM infers latent actions indirectly from current and future visual pairs via an inverse dynamics model over 4,500 hours of video pretraining and 200k policy-integration steps. Because inverse dynamics summarizes visual changes after a transition occurs, its latent actions can omit non-visual control details and require costly multi-stage pretraining. In contrast, WAVE-0 trains its Stage-1 world model directly on ground-truth robot action chunks and proprioceptive states from action-labeled demonstrations, using a counterfactual action-ranking loss to prevent observation-only shortcuts and enforce sensitivity to distinct action sequences without external video pretraining.
  3. In terms of policy integration, VLA-JEPA and JEPA-WAM use latent prediction only as an auxiliary training objective and leave inference as a direct observation-to-action mapping, whereas LaWAM and WAVE-0 condition action generation on predicted future representations at inference. Specifically, WAVE-0 integrates the frozen world model between the VLM and the action expert to form a unified pipeline of latent planning, future prediction, and action generation. VLM-generated latent action plans are aligned with the Stage-1 action representation space and translated by the world model into predicted future visual representations, which a lightweight resampler compresses into conditioning tokens for the action expert. Because latent-action query tokens are excluded from the VLM context passed to the expert, the planning branch guides action generation only through its predicted visual consequences during both training and inference.

The table below summarizes these differences:

MethodCompact Latent SpaceGT-Action Forward WMInference Future Cond.No Extra Embodied Pretraining
Fast-WAM✗✗✗✓
GigaWorld-Policy✗✓✗✗
AHA-WAM / BagelVLA / ImageWAM✗✗✓✓
DreamZero / Motus / LingBot-VA✗✗✓✗
VLA-JEPA✓✗✗✗
JEPA-WAM✓✗✗✓
LaWAM✓✗✓✗
WAVE-0 (Ours)✓✓✓✓

For Weakness 2: "In terms of experiments: First, it is necessary to explain from the experimental results which design serves as the most critical performance-enhancing module in this paper. Second, the results on Robotwin2.0 are not very competitive; please elaborate on the main issues encountered with this dataset."

(1) Most critical module. Our ablations show that coupling an action-sensitive Stage-1 latent world model with aligned pre-expert future conditioning is key to WAVE-0's performance, as removing either the action-grounding objectives or the pre-expert integration degrades results.

  • Stage-2 world and alignment supervision in Fig. 4(d). Keeping the frozen world model in the pathway while training Stage 2 with the flow loss alone gives 84.6% on LIBERO-Plus, matching base π0.5 at 84.5%. Supervising latent plans with world consistency and action alignment raises success to 88.0% on LIBERO-Plus and from 58.0% to 61.4% on LIBERO-Pro.
  • Stage-1 counterfactual ranking in Fig. 4(a to c). Removing ranking collapses the action sensitivity gap from 0.1648 to 0.0020 and drops LIBERO-Plus and LIBERO-Pro by 2.6 and 3.0 points.
  • Ground-truth action grounding vs. video IDM. In a new matched comparison conducted during the rebuttal on the same π0.5 backbone and training budget, replacing our Stage-1 ground-truth action encoder with a LaWAM-style visual inverse-dynamics encoder drops success from 98.7% to 98.3% on LIBERO, 88.0% to 86.2% on LIBERO-Plus, 61.4% to 60.0% on LIBERO-Pro, and 91.47% to 90.15% on RoboTwin 2.0 Avg. (including 91.48% to 89.56% under randomization), confirming the effectiveness of grounding latent dynamics directly in executable robot actions rather than inferring latent actions indirectly from visual transitions.
  • Pre-expert conditioning vs. post-expert supervision. Comparing the two parallel ways of integrating our Stage-1 world model into π0.5 in Fig. 1(3, 4), placing the world model before the action expert to directly condition action generation in WAVE-0 outperforms the parallel post-expert variant π0.5 + WM at 91.47% vs. 89.51% on RoboTwin 2.0 in Table 1b, as well as on our newly evaluated LIBERO benchmarks during the rebuttal, improving LIBERO from 98.4% to 98.7%, LIBERO-Plus from 86.4% to 88.0% (+1.6 points), and LIBERO-Pro from 58.8% to 61.4% (+2.6 points).

(2) Performance and main issues on RoboTwin 2.0. First, our base VLA model π0.5 only achieves 80.38% on RoboTwin 2.0, and simply introducing our lightweight 257M latent world model boosts it by +11.09 points to 91.47%, ranking first among latent WAMs ahead of LaWAM at 91.22%. Second, competing RoboTwin WAMs either curate custom data recipes for large-scale embodied pretraining, such as LingBot-VA on 1.4T tokens and LaWAM on 4,500h of videos, or follow a visual-generation WAM paradigm (Fast-WAM, AHA-WAM, and ImageWAM) whose architecture differs substantially from VLAs by coupling video or image generation backbones with custom DiT action models, totaling 6.0B to 8.6B parameters. In contrast, WAVE-0 uses no extra pretraining data and no action-policy redesign to demonstrate an effective framework on top of an existing VLA, offering clear advantages in model size at 4.3B and inference latency at 208.7 ms on the same NVIDIA RTX PRO 6000 Blackwell GPU, while leaving joint large-scale world-policy pretraining for future work.

MethodRoboTwin SR (%)ParamsLatency (ms) ↓
ImageWAM93.388.6B234.4
AHA-WAM92.807.2B443.4
LingBot-VA92.205.3B7,366.3
Fast-WAM91.836.0B369.2
WAVE-0 (Ours)91.474.3B208.7

Per-task results in Appendix Table 13 further show that WAVE-0 already reaches 95% to 100% on most manipulation and stacking tasks, and the remaining gap mainly comes from precise-alignment and articulated tasks.

  • Coarse spatial pooling. Pooling each camera to a 4×4 latent grid and 2×2 expert tokens keeps latency overhead minimal (+18.8 ms in Table 11), but may discard fine local geometry needed for thin-structure alignment such as Hanging Mug at 48.0% and Move Stapler Pad at 53.5%.
  • Single-endpoint prediction. Predicting only a single future endpoint 31 steps ahead may provide limited guidance on intermediate contact states during constrained articulated motions such as Open Microwave at 66.0% and Turn Switch at 73.5%.
Reviewer_36i2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WAVE-0 adds a latent world model with 257M parameters, including its encoders, to π0.5.

The world model is first trained with ground-truth actions to predict future features in π0.5's frozen SigLIP space, using a cosine loss, a residual loss and a ranking loss that asks the prediction under the true action to be closer to the real future than the prediction under a mismatched action. It is then frozen. Latent action queries appended to the VLM input produce an action plan, the world model turns the plan into a predicted future, and 12 resampled tokens of that future condition the action expert. The paper reports 91.47% on RoboTwin 2.0 against 80.38% for π0.5, 98.7% on LIBERO, gains on LIBERO-Plus and LIBERO-Pro, and 73.3% against 38.3% for π0.5 on three real-robot tasks.

Critical strengths
  1. The design is simple and cheap to add to an existing VLA. The world model predicts the future in the feature space of a frozen copy of π0.5's SigLIP encoder and projector, so no pixel decoding is needed, and in the real-robot setup the full system adds 18.8 ms of mean server-side latency to π0.5 according to Table 11.

  2. The counterfactual ranking loss addresses a real weakness of latent world models. In Fig. 4(a), without the loss the prediction errors under the true action and under a mismatched action differ by only 0.0020 on average after 20k Stage-1 steps, against 0.1648 with it. In Fig. 4(c), adding the loss also improves LIBERO-Plus and LIBERO-Pro by 2.6 and 3.0 points.

Critical weaknesses
  1. The contribution is argued against the wrong baseline. Most headline results are gains over π0.5, a policy that does no world modeling: +11.09 points on RoboTwin 2.0, +1.8 on LIBERO and +35.0 on the real robot. That predicting futures helps a robot policy has already been reported by prior world action models, including LaWAM, VLA-JEPA and JEPA-WAM in the paper's own tables, and Seer, VPP and DreamVLA. What the paper needs to show is what this particular technique adds over them. In the paper's own Table 1b, most of the RoboTwin gain, 9.13 of the 11.09 points, already appears when the Stage-1 world model is used only as an auxiliary loss, which is the paradigm the paper attributes to prior work; explicit conditioning adds the remaining 1.96. Against other world action models the evidence is weak. On RoboTwin 2.0, WAVE-0 reaches 91.47%, below ImageWAM at 93.38%, AHA-WAM at 92.80%, LingBot-VA at 92.20% and Fast-WAM at 91.83%, and above LaWAM by 0.25 points, Motus at 87.84% and GigaWorld-Policy at 85.70%. On LIBERO it is 0.1 points above LaWAM. Section 4.3 discusses only π0.5 and LaWAM.

  2. Where WAVE-0 does beat the strongest world-model baseline by a clear margin, the comparison does not isolate the technique. This happens on LIBERO-Plus, 88.0 against 83.1 for ImageWAM, and on LIBERO-Pro, 61.4 against 52.7 for LaWAM. But on LIBERO-Plus, π0.5 alone already scores 84.5, above every world-model baseline in Table 2, and on LIBERO-Pro its reported 53.0 is level with the best of them. Here and below I use the reported 53.0, although π0.5's 20 entries in Table 12 average 48.2, as noted in question 3. Beating these baselines therefore does not show that this technique is better than theirs. On these benchmarks the evidence for the technique is again its gain over π0.5, 3.5 and 8.4 points, which does not separate it from other latent world modeling on the same backbone.

The comparisons that isolate WAVE-0's design choices are few and mixed. In Table 1b, explicit conditioning adds 1.96 points over the auxiliary-loss variant, on RoboTwin only and with no seed count or variance reported. In Fig. 4(c), the ranking loss adds 2.6 and 3.0 points on LIBERO-Plus and LIBERO-Pro. In Fig. 4(d), the conditioning path trained with the flow loss alone scores 84.6 on LIBERO-Plus and 58.0 on LIBERO-Pro, against 84.5 and 53.0 for π0.5 in Table 2. Whether the Stage-1 world model, trained on ground-truth actions, is better than an existing latent world model in the same position is not tested. The only support for this choice is cross-method comparisons, mainly with LaWAM, which in the real-robot setup of App. A.2 also differs in model size, action chunk length and state input. The real-robot comparisons are not matched either: according to App. A.2, WAVE-0's world model is trained on all 600 demonstrations and each baseline on the 200 of its own task, although Sec. 4.5 says all methods use the same data. Table 4, in turn, counts LaWAM's large shared pretraining in its real-robot budget, so the data differs in both directions.

Questions for authors
  1. Can you add matched comparisons on the same π0.5 backbone, data and training budget, on RoboTwin and LIBERO with averaging over seeds: (a) the auxiliary-loss variant, π0.5 + WM, with 12 extra non-predictive conditioning tokens and the same training steps as WAVE-0, (b) a world model trained on latent actions inferred from video instead of ground-truth actions, (c) the action queries passed to the action expert without the world model, and (d) conditioning on StS_t or on the ground-truth St+hS_{t+h}? These are the comparisons that would show what the technique adds over existing latent world modeling.

  2. On RoboTwin 2.0, ImageWAM, AHA-WAM, LingBot-VA and Fast-WAM all score above WAVE-0. If the argument for WAVE-0 is cost, can you report inference latency for AHA-WAM, LingBot-VA and Fast-WAM, and parameters and training compute for all four? In the real-robot setup, Table 11 gives 582.6 ms for ImageWAM. LaWAM is faster than WAVE-0 at 223.9 ms against 305.9 ms, but it is also smaller, with 2.3B parameters against 4.3B, and returns 18 actions per request instead of 32.

  3. Are the world-model baselines in Tables 1 and 2 reproduced or taken from papers, and how many evaluation episodes are used for LIBERO and its variants? Which LIBERO-Pro average for π0.5 is correct: the reported 53.0, or 48.2, the mean of its 20 entries in Table 12?

  4. On the real robot, the comparison that bears on the contribution is the one with LaWAM and ImageWAM. How does WAVE-0 perform if Stage 1 uses only the 200 demonstrations of the target task, as the baselines do? Are all methods using the same initial object placements and the same success criterion?

I would raise my score if the questions above are answered properly.

Also a quick suggestion on citation and differentiation: the introduction separates WAVE-0 from VLA-JEPA and JEPA-WAM by its explicit conditioning on predicted latent futures, but Seer, VPP and DreamVLA, which are not cited, already condition action generation on predicted futures. The positioning should say how WAVE-0 differs from them.

12 comments

Dear Reviewer 36i2,

Thank you for your careful reading and constructive feedback. We appreciate your recognition that WAVE-0's design is simple and low-cost to add to an existing VLA by predicting futures in π0.5's frozen SigLIP feature space without pixel decoding (+18.8 ms latency in Table 11), and that our counterfactual action-ranking objective addresses a real weakness of latent world models with clear gains in both Stage-1 action sensitivity and OOD policy success. Below, we (1) clarify WAVE-0's positioning and technical differences from prior generative and latent WAMs including Seer, VPP, and DreamVLA (which we will cite and discuss in the revised manuscript), (2) provide matched comparisons on the same π0.5 backbone and training budget, (3) report unified latency, model size, and training compute on RoboTwin 2.0, (4) clarify baseline sources, episode counts, and the 53.2% π0.5 score on LIBERO-Pro, and (5) present the matched 200-demonstration single-task Stage-1 real-robot experiment and evaluation protocol.

For Weakness 1: "The contribution is argued against the wrong baseline. Most headline results are gains over π0.5, a policy that does no world modeling: +11.09 points on RoboTwin 2.0, +1.8 on LIBERO and +35.0 on the real robot. That predicting futures helps a robot policy has already been reported by prior world action models, including LaWAM, VLA-JEPA and JEPA-WAM in the paper's own tables, and Seer, VPP and DreamVLA. What the paper needs to show is what this particular technique adds over them."

For Citation Suggestion: "Also a quick suggestion on citation and differentiation: the introduction separates WAVE-0 from VLA-JEPA and JEPA-WAM by its explicit conditioning on predicted latent futures, but Seer, VPP and DreamVLA, which are not cited, already condition action generation on predicted futures. The positioning should say how WAVE-0 differs from them."

π0.5 serves as the base-policy reference because WAVE-0 directly extends its architecture to bridge pretrained VLAs and WAMs. We agree that gains over π0.5 alone do not establish what WAVE-0 adds over existing WAMs. Our contribution is not the general observation that future prediction benefits robot policies, but a specific way to learn action-conditioned latent dynamics directly in a pretrained VLA's feature space from demonstrated robot actions and use predicted action consequences to condition action generation, without pixel reconstruction or additional large-scale video pretraining. We clarify the technical differences below and support them with WAM comparisons and matched-backbone ablations in Questions 1 and 2:

1. Visual generation-based methods. Approaches built on image or video generation backbones or dense visual prediction targets include ImageWAM, AHA-WAM, LingBot-VA, Fast-WAM, Seer, VPP, and DreamVLA. Modeling dense visual futures can incur substantial computation, particularly when iterative visual denoising is required at inference. Recent methods address this cost in different ways. LingBot-VA jointly models video and actions through autoregressive diffusion, with closed-loop observation updates and asynchronous execution. Fast-WAM retains video prediction during training but removes future prediction at inference. AHA-WAM uses a low-frequency video planner whose context is reused by a high-frequency action expert through asynchronous, observation-guided routing. ImageWAM conditions a flow-matching action expert on KV caches from image-editing denoising, without decoding future images at inference. Seer jointly learns visual foresight and inverse dynamics, allowing action tokens to attend to predictive visual tokens supervised by future RGB reconstruction. VPP conditions a diffusion policy on internal predictive features extracted from a video diffusion model using a single denoising step. DreamVLA predicts dynamic regions, depth, and semantic features, and uses the resulting world embeddings for action generation while skipping visual decoding at inference. These works demonstrate the value of predictive representations for control, while still relying on visual generation backbones or dense reconstruction targets. Specifically, Seer, VPP, and DreamVLA forecast future visual or multimodal targets directly from current observations and instructions before inferring actions, whereas WAVE-0 predicts the consequences of VLM-generated latent action plans in π0.5's frozen feature space.

2. Latent world action models. Rather than relying on generative backbones or dense reconstruction, latent WAMs predict future representations in a compact embedding space. However, VLA-JEPA and JEPA-WAM use latent prediction only as an auxiliary training objective to shape representations, without conditioning action generation on predicted futures at inference. LaWAM is the closest comparison to WAVE-0. Both methods use a predicted latent action to forecast future visual features that guide action generation. They differ in how they obtain the action representation used to train the world model. LaWAM infers latent actions from pairs of current and future visual features via inverse dynamics. WAVE-0 encodes the robot action sequences recorded in the demonstrations and trains the world model to predict their visual consequences in π0.5’s frozen feature space. A counterfactual ranking loss encourages these predictions to depend on the action sequence, and its contribution is supported by the controlled ablation in Sec. 4.7.2, Fig. 4. In Stage 2, the VLM learns latent action plans aligned with these action-encoder representations, which pass through the frozen world model to produce future visual features that condition the action expert. Their additional training requirements also differ. LaWAM pretrains its world model on approximately 4,500 hours of robot and human videos, followed by 200k steps of policy-integration pretraining before benchmark-specific training. WAVE-0 reuses pretrained π0.5 and trains its added world model and policy directly on action-labeled robot demonstrations, without these additional large-scale pretraining stages. Despite using a fraction of the training budget (e.g., 670 vs. 1,280 H100 GPU hours for RoboTwin 2.0 training alone, without the 307.2M-sample video pretraining), WAVE-0 still outperforms LaWAM on both LIBERO (98.7% vs. 98.6%) and RoboTwin 2.0 (91.47% vs. 91.22%).

The table below summarizes the key differences across these methods:

MethodCompact Latent SpaceGT-Action Forward WMInference Future Cond.No Extra Embodied Pretraining
Fast-WAM✗✗✗✓
AHA-WAM / ImageWAM / DreamVLA✗✗✓✓
LingBot-VA / Seer / VPP✗✗✓✗
VLA-JEPA✓✗✗✗
JEPA-WAM✓✗✗✓
LaWAM✓✗✓✗
WAVE-0 (Ours)✓✓✓✓

Comments on the paper

1 comment
AuthorsOpenReview ↗

Changes in the revised version of the paper

Dear Area Chair and Reviewers,

We sincerely appreciate your insightful and constructive feedback on our paper. Across the rebuttal, we conducted additional experiments and clarifications addressing all reviewer concerns. We have now uploaded the revised manuscript and updated Supplementary Material, with newly added content in the manuscript highlighted in blue for your convenience:

In Sec. 1 and Sec. 2.2, we added citations and discussion differentiating WAVE-0 from prior predictive policies including VPP, Seer, and DreamVLA.

In Sec. 3.2, we added the exact construction of mismatched action sequences, the definition of the validity mask rr, and the role of the counterfactual ranking objective in Eqs. (6)–(7).

In Sec. 4.4, Table 2, and Appendix A.4 (Table 18, formerly Table 12), we corrected the π0.5 scores on LIBERO-Pro.

In Sec. 4.6, we added a comparison of model size and inference latency against leading visual generation-based WAMs on RoboTwin 2.0 (Table 5).

In Sec. 4.7.1 and Sec. 4.7.2, we consolidated π0.5 + WM into Table 6, added its evaluation across LIBERO and its variants, and included matched-backbone ablations on non-predictive conditioning tokens, a Video IDM world model, direct action queries without the world model, and current-feature conditioning.

In Appendix A.1, we added evaluation episode counts for LIBERO and its variants alongside the sources of baseline results in Tables 1 and 2.

In Appendix A.3.1, we added Stage-1 diagnostics and downstream ablation results on the separate proprioceptive state pathway zsz_s (Tables 14–15).

In Appendix A.3.2, we added a comparison of visual prediction spaces replacing the frozen SigLIP encoder with a frozen DINOv3 encoder across LIBERO and RoboTwin 2.0 (Table 16).

In Appendix A.3.3, we added controlled Stage-1 data scaling and task-coverage experiments measuring both future prediction quality and downstream LIBERO-Plus success (Table 17).

In the Supplementary Material, we included individual real-robot rollout videos of WAVE-0 as well as side-by-side comparison videos across all evaluated methods on the three real-world tasks.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.