- 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
-
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.
-
The latency work is documented to a standard this literature rarely reaches. Table 2 stages 2854 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
-
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 and , 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.
-
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, ). Either add the crossed control or restate the claim as a data-volume effect.
Questions for authors
- Could you give per-task confidence intervals, or a paired test across the ten tasks, for the four headline comparisons? Pooled Fisher gives (Franka 54 vs 49), (Galbot 45 vs 41), (auxiliary 59 vs 45) and (sync 45 vs async 39).
- On Franka, were Cosmos3-Nano-Policy and 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 get 50k steps on Galbot?
- Please state the inference operating point, the handoff level (Table 7 gives only ), 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?
- 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 margin over (91 ms) reflect architecture or engineering?