acceptodds
Under review as a conference paper at ICLR 2027

DualWAM: Dual-System World Action Models for Asynchronous Global Planning and Local Refinement

Abstract

World Action Models (WAMs) jointly generate robot actions and predict future world states, transferring priors from video pretraining to robot control. However, future visual prediction is computationally expensive, so existing WAMs often rely on long action chunks to amortize inference cost across control steps, at the cost of closed-loop responsiveness. We present DualWAM, a dual-system WAM that preserves broader-horizon world-action generation while enabling high-frequency closed-loop action updates by decoupling global planning and local refinement. System 2 periodically performs high-noise bidirectional denoising over a broader world-action chunk to establish a global plan, while wrist-only System 1 extracts a temporally aligned short window from the intermediate denoising state and completes low-noise refinement using the latest wrist observations, which provide action-aligned cues about local geometry, motion, and contact during interaction. The two systems operate asynchronously along a shared denoising trajectory: each global plan is reused across multiple local updates, while System 1 repeatedly incorporates fresh interaction feedback. Across zero-shot manipulation tasks on Franka and Galbot, DualWAM improves success over the strongest evaluated baseline by 4.5 percentage points on average, while achieving a 16.6 critical-path speedup. Further studies show that role-matched egocentric and UMI data improve success by 14 percentage points, and that the decoupled design naturally supports edge–cloud deployment with substantially lower communication overhead than the baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world action models, robot manipulation, zero-shot learning, video generation, asynchronous inference

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

The paper targets the cost-versus-agility trade-off in world-action models. Predicting future visual states is expensive, so existing WAMs amortise inference over long action chunks and lose closed-loop responsiveness. DualWAM splits one rectified-flow world-action trajectory between two networks. System 2 denoises a 32-frame multi-view world-action chunk from pure noise to a handoff level τ at a low rate. System 1, a smaller wrist-only WAM, continues from that partially denoised state over 8-frame windows at each control update, conditioned on fresh wrist images. System 1 refines future wrist latents as well as actions, trains on the low-noise interval up to τtrain = 0.7, and sees time-shifted inputs in later windows. One global plan serves four local windows, and System 1 finishes the low-noise interval in a single forward pass. Validation covers 20 zero-shot real-robot tasks (ten on a Franka FR3, ten on a Galbot G1), ten trials each, against π0.5, Cosmos3-Nano-Policy, Fast-WAM and AHA-WAM, plus an ablation of the dual-system split, a synchronous-versus-asynchronous comparison with edge-cloud payload accounting, and a role-matched auxiliary-data study. DualWAM reaches 54% and 45% success, 63 ms critical-path latency, and a 14-point gain from auxiliary data.

Critical strengths
  1. Both systems are world-action models on one shared denoising trajectory, and the handoff is a state cut rather than a second policy. That is what separates this from a generic slow-planner and fast-controller pairing, and why Table 3 is the right ablation to run. If the effect holds, it is a reusable primitive for amortising world-model inference without losing closed-loop correction.

  2. The latency work is documented to a standard this literature rarely reaches. Table 2 stages 2854 →\rightarrow 63 ms with each optimisation isolated, Appendix A.1 names the mechanisms, and the final number runs on three RTX 4090s rather than a datacentre node. With the payload accounting in Table 5, the deployment claim becomes checkable rather than aspirational.

Critical weaknesses
  1. The advantage over the strongest baseline is not resolved at the precision the abstract implies. Ten tasks times ten trials give 100 rollouts per policy per embodiment, and the margins are 5 points on Franka and 4 on Galbot. On those counts a two-sided Fisher test gives p=0.57p = 0.57 and p=0.67p = 0.67, so the "4.5 percentage points on average" is the mean of two differences that are individually indistinguishable from zero. Ten trials per task is the norm here. My concern is that the margin is narrower than the protocol can resolve, and that no interval or paired test is reported although every rollout is logged and the paired analysis is computable. The per-task pattern is more encouraging than the pooled test (Figure 7, five tasks above the strongest baseline and five level), which is exactly why that analysis is the one to report. On Franka the comparison is also confounded. Section 4.1 states that the two strongest baselines are used as released DROID checkpoints, while both DualWAM systems train for a further 10k iterations at batch size 1024. Table 3's "System 2 only" carries exactly the released model's numbers (76.6 progress, 49% success, 1043 ms, identical to Figures 4 and 7), so the ablation contrasts a trained dual system with an untrained base. Galbot does control for this, which leaves its 4-point difference as the clean estimate of the architectural effect, and that is the one that fails the significance test.

  2. The role-matched data result is presented as evidence for a mechanism it does not test. Table 6 contains only matched assignments, egocentric video to System 2 and UMI demonstrations to System 1. Each helps and together they add 14 points, but without the crossed assignment the same result is equally consistent with "more heterogeneous data helps", which is weaker and already known. The corpora also differ in scale (1,800 versus 1,200 hours) and modality, so the 9- and 6-point gains are not comparable. On the reported counts the 14-point claim is itself at the edge of resolution (59/100 versus 45/100, p=0.066p = 0.066). Either add the crossed control or restate the claim as a data-volume effect.

Questions for authors
  1. Could you give per-task confidence intervals, or a paired test across the ten tasks, for the four headline comparisons? Pooled Fisher gives p=0.57p = 0.57 (Franka 54 vs 49), 0.670.67 (Galbot 45 vs 41), 0.0660.066 (auxiliary 59 vs 45) and 0.470.47 (sync 45 vs async 39).
  2. On Franka, were Cosmos3-Nano-Policy and π0.5\pi_{0.5} retrained with DualWAM's 10k budget, or evaluated as released? Was "System 2 only" trained at all, since it reports that checkpoint's 76.6 / 49 / 1043 exactly? And why does π0.5\pi_{0.5} get 50k steps on Galbot?
  3. Please state the inference operating point, the handoff level τ\tau (Table 7 gives only τtrain=0.7\tau_\mathrm{train} = 0.7), its sensitivity, the control frequency, System 1's replanning period, and the end-to-end observation-to-action latency. Was System 1's single-step refinement tested against multi-step on success?
  4. Figure 4 gives AHA-WAM 238 ms and Fast-WAM 260 ms, while their own papers report about 41 and 190 ms, and Section 4.1 says only Cosmos3 got your optimisations. Does the 1.4×1.4\times margin over π0.5\pi_{0.5} (91 ms) reflect architecture or engineering?
Reviewer_FQtyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DualWAM, a world action model (WAM) that consists of two models running at different rates. WAMs are slow because they also predict future video, so existing WAMs usually execute long action chunks and cannot respond quickly to new observations. The key idea is to divide one denoising process between two WAMs. System 2 runs about every two seconds. It uses both the global and wrist views and denoises a long chunk of future video and actions only up to an intermediate noise level, which serves as a coarse plan. System 1 is a smaller model that uses only the wrist camera. At every control step, it takes the part of the plan that corresponds to the current time and completes the denoising in a single step, conditioned on the latest wrist image. Because the plan can become misaligned with the actual execution, System 1 is trained with temporally shifted plan segments so that it follows the current observation when the two do not match. In zero-shot experiments on Franka and Galbot robots, DualWAM achieves the highest success rates among the compared methods and reduces latency to 63 ms, about 16 times faster than using System 2 alone.

Critical strengths
  • The paper addresses an important problem, since latency has been a major bottleneck when deploying WAMs in real-world settings.
  • The paper performs extensive ablation studies to verify each component.
  • The paper conducts fairly extensive real-world experiments to verify the effectiveness of the method.
Critical weaknesses
  • In Table 6, all variants are DualWAM models trained with different data, and each one is compared only against robot-only training. It would help to also train a single model, such as Cosmos3-Nano-Policy, on the same egocentric and UMI data. Another useful comparison is to give the data to the other system, with egocentric videos going to System 1 and UMI data to System 2. Without these experiments, it is hard to tell whether the 14-point gain comes from assigning each data source to the matching system or simply from adding about 3,000 hours of extra data.
  • Some important hyperparameters are not reported. Table 7 uses τtrain\tau_\text{train}=0.7, but the paper does not say which τ\tau is used at inference. The video frame rate and the control frequency are also missing, so it is not clear how many seconds H1H_1 = 8 frames and H2H_2 = 32 frames correspond to, or how often System 1 updates in practice.
  • There is no ablation on τ\tau, H1H_1, H2H_2, or the System 2 refresh interval. These choices decide how much System 2 fixes the plan and how much System 1 can still correct, so it would help to show how sensitive the results are to them. For example, success rate and latency could be reported for a few different values of τ\tau.
Questions for authors
  • Is there a reason why some ablations (e.g., the architecture ablation) are conducted on Franka, while others (e.g., asynchronous execution) are conducted on Galbot?
Reviewer_mF4vOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

DualWAM combines low-frequency global world-action planning with high-frequency wrist-based local refinement to enable responsive and efficient manipulation.

Critical strengths

The decomposition is intuitive and directly addresses the tension between long-horizon coherence and responsive closed-loop control.

The evaluation is broad and practically relevant, covering multi tasks across real robot platforms, together with latency, auxiliary-data, and edge–cloud deployment studies.

Critical weaknesses

Could the authors clarify how System 1 is trained to handle the actual intermediate trajectories produced by System 2? System 1 appears to be trained on ground-truth trajectories with synthetic noise, but receives imperfect System 2 outputs at inference time. This mismatch may affect whether System 1 can reliably refine System 2’s plans, particularly because the two systems are initialized from different models. An evaluation using actual System 2 outputs during System 1 training and testing would help address this concern.

How does System 1 switch to a newly generated System 2 plan while maintaining temporal alignment and action continuity? The window index is updated or how discontinuities between the old and new plans

Questions for authors

See the Weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.