{"id":"aab15e70-ec56-4ed9-86ce-a019563512b2","arxiv_id":"2608.04996","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Adding motion, depth, and semantic supervision to a world action model's future prediction during training, then removing it at inference, improves robot manipulation robustness under visual perturbations.","lead":"DreamWAM trains a robot policy to imagine not only what future video frames look like, but also how objects move, where they are in space, and what they are, using these extra signals only during training. The policy then runs on plain RGB input at deployment and gets more robust when lighting, backgrounds, or object layouts change, outperforming matched RGB-only baselines on LIBERO, LIBERO-Plus, and a real robot.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The motion view, which the ablation identifies as the most consequential supervision signal, depends on a color-coded flow pipeline whose fidelity is never checked, so the central \"form matters\" claim could rest on an unvalidated proxy.","rationale":"The reader's weakest assumption matches my concern, so I agree. The central claim is that representing future states in action-relevant forms, namely motion, geometry, and semantics, is what improves robustness. The ablation isolates motion as the most important view, so the fidelity of the motion target is a necessary condition for the mechanism. I considered two alternative concerns: (i) lack of confidence intervals, and (ii) model selection on LIBERO. Both are real but less load-bearing: the LIBERO-Plus effect is large and consistent across seven shifts and two inference modes, so binomial sampling error is unlikely to explain it; the routing selection is a standard validation procedure. The motion pipeline concern is more fundamental because it questions whether the measured effect is due to the claimed form at all. A second related concern is the train/test mismatch from removing residual branches (Eq. 3) at inference, but that is common in auxiliary-loss designs and the matched baselines share the deployment interface. The concrete fidelity check is cheap and decisive: if the VAE latent preserves flow direction and magnitude, the concern is resolved; if not, the paper's interpretation needs substantial revision. Therefore the reader's CONDITIONAL verdict remains appropriate, with the motion-fidelity check added as a condition.","tokens_in":15655,"tokens_out":11248,"duration_ms":140825,"concrete_test":"Decode the precomputed z_flow latents (Eq. 5) with the Wan2.2 VAE decoder and run RAFT on the decoded color-coded flow videos; compute average endpoint error relative to the original RAFT fields. Also train a linear probe on z_flow to predict per-pixel flow magnitude and direction. If the endpoint error is comparable to the flow magnitude or the probe cannot recover direction and magnitude, the motion supervision is not preserving action-relevant motion, and the motion-view explanation in Table 3 should be revisited.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's ablation (Table 3) identifies motion as the most consequential single view, yet the motion target is constructed by converting RAFT optical flow into color-coded RGB visualizations and encoding them with the Wan2.2 video VAE (Appendix A.1, Eq. 5). The VAE was trained to reconstruct natural RGB video; color-coded flow is far out of distribution. The paper reports no check that the VAE latent preserves flow direction or magnitude, nor even that decoded flow videos resemble the originals. If the VAE treats the color-code as texture rather than motion, the 'motion' objective is an appearance-based denoising task, not a temporal-dynamics task. That would break the mechanism claimed: the improvement attributed to motion would actually be an uninterpretable auxiliary appearance target. The D/D/D routing control shows that the route matters, but it does not establish that the content being routed is motion. Since the central claim is about the form of the future representation, this is load-bearing: the main novelty rests on the motion view being motion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DreamWAM, a world action model that supplements RGB future prediction with training-time supervision from motion, geometry, and semantic views. RGB and optical-flow latents are jointly denoised, while depth and DINOv2 features enter through gated residual branches; at inference all beyond-RGB pathways are disabled, so deployment is RGB-only. The authors report consistent gains over matched Fast-WAM baselines on LIBERO, larger gains on the unseen LIBERO-Plus perturbations, and improvements on a real-robot evaluation under visual perturbations. They also present ablations on which views help and on how each view should be injected (full denoising vs. gated residual).","tokens_in":15754,"tokens_out":4316,"duration_ms":57463,"significance":"If the central claim holds, the paper makes a useful contribution: it identifies the representation of the predicted future, not just the act of prediction, as a design variable in world action models, and it provides a concrete architecture that uses structured future supervision while retaining an RGB-only deployment interface. The evaluation is well controlled relative to many prior WAM papers: the Fast-WAM-Joint baseline shares backbone, data, and training protocol, gains appear across all seven LIBERO-Plus perturbation dimensions, and the real-robot experiments test true distribution shifts. The public release of code and models is a further strength. However, the load-bearing mechanism—that the motion view is actually providing motion information—is not validated, and the empirical claims lack uncertainty quantification. The paper is therefore promising but needs additional evidence before the central claim can be accepted.","major_comments":[{"comment":"The motion view is identified in the ablation as the most consequential single view, yet the paper never validates that the Wan2.2 VAE encoding of color-coded RAFT flow preserves the flow direction or magnitude. The VAE is trained for natural RGB video, and color-coded flow visualizations are far out of distribution; Appendix A.1 explicitly states that z_flow is the VAE encoding of a visualized motion sequence rather than raw metric flow. If the VAE treats the color code as texture, the \"motion\" objective is not a temporal-dynamics target and the central claim that the form of the future representation matters would not be supported by the motion ablation. I request a concrete fidelity check: decode z_flow and compare it to the original flow visualization or to raw flow fields using an average endpoint error or correlation metric, and also report a control where the motion target is replaced by an appearance-matched but temporally meaningless target (or where a different flow color map is used). Without such evidence, the motion attribution in Table 3 is not established.","section":"Appendix A.1, Eq. (5); Table 3"},{"comment":"All simulation results average only two random seeds, and no confidence intervals, standard deviations, or significance tests are reported. This is particularly important for the LIBERO results, where the claimed gains are small (97.30 to 98.40 and 98.00 to 98.90) and the per-task rollout count is only 50 per seed. The LIBERO-Plus gains are larger, but a formal comparison (e.g., bootstrap confidence intervals over tasks or per-seed results) is needed to rule out seed noise. The real-robot results are based on 30 trials per setting and are also reported as point estimates; error bars or confidence intervals should be added.","section":"Experimental Setup; Tables 1 and 2; Figure 3"},{"comment":"The claim that Fast-WAM-Joint and DreamWAM are \"matched\" is not exact: DreamWAM adds 48 flow input channels to the VideoDiT input projection and includes residual branches and prediction heads during training, so the parameter count and the training-time update dynamics differ from the baseline. Although the flow channels are zeroed at inference, the corresponding projection weights are trained, so gains could in principle come from extra capacity rather than from the structured form of the supervision. Please report parameter counts for the matched pairs, and consider an additional control that uses the same extra channels with a non-informative target (e.g., random noise or a static flow-like signal) to isolate the effect of the motion content.","section":"Experimental Setup; Eqs. (8)-(10)"}],"minor_comments":[{"comment":"The label \"DA3geometrylatent\" should read \"DA3 geometry latent\"; also clarify whether the duplicated first transition in the flow video is the first flow field or the first transition, as the current sentence is slightly ambiguous.","section":"Appendix A.1"},{"comment":"The method name \"LingBot-VA (2026b)\" is inconsistent with the reference list, which cites \"A Pragmatic VLA Foundation Model\" by Wu et al.; please verify the citation key and use the correct model name (LingBot-VLA) throughout.","section":"Table 1"},{"comment":"Figure 3 shows success rates without any uncertainty indication; adding error bars or per-trial counts would make the real-robot comparison more interpretable.","section":"Figure 3"},{"comment":"The notation \"ConcatC\" is used but not defined; please define it as concatenation along the channel dimension, and state the channel counts of x_rgb and x_mot explicitly for reproducibility.","section":"Method, Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the central idea is interesting, but the missing validation of the motion target is the main blocker: the ablation that establishes the paper's headline mechanism is not backed by a check that the supervision actually preserves motion information. I would be willing to reconsider after the authors add the requested fidelity checks and statistical uncertainty quantification. The parameter-matching issue is secondary but should also be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a well-executed, controlled study and the central empirical claim is probably right. The new idea is to supervise a VideoDiT–ActionDiT WAM with motion, depth, and semantic targets during training, then drop all extra branches at inference. The matched Fast-WAM baseline is the right control, the gains on LIBERO-Plus are consistent across all seven shift dimensions, and the real-robot numbers reinforce the trend. The code release helps. This deserves a serious referee.\n\nThe soft spots are real but not disqualifying. First, there are no confidence intervals or significance tests, and only two seeds in simulation. Given the size of the effect, that is a minor-to-moderate issue. Second, the D/R/R routing was selected on LIBERO after the ablation, so LIBERO-Plus is an out-of-distribution check of a configuration chosen in-distribution. That weakens the generalization story, but the effect is large enough that it is unlikely to be pure selection. Third, and most important: the motion target is color-coded RAFT flow encoded by the Wan2.2 VAE. The appendix honestly says it is a visualized motion sequence, not raw flow, but the paper never checks that the VAE latent preserves flow direction or magnitude. Since the ablation says motion is the most consequential view, the mechanism credited to 'motion' might partly be an arbitrary auxiliary appearance target. That is a genuine hole in the interpretation, and I would want the authors to validate the target or at least run a sensitivity test on the compression rank. The stress-test note is right about this being the weakest link.\n\nAll that said, I take the central result seriously. The overall pattern—gains amplified under shift, present in both no-rollout and joint modes, and on hardware—is stronger than most work in this area. The authors are not hiding the proxy; they just do not test it. I would send this to review, expecting revision to add error bars, hyperparameter specifics, and some check of the flow target. For my own work, I would cite it cautiously.","headline":"Well-matched empirical study; central claim likely holds, but the motion target's fidelity is an unvalidated proxy that should be checked before the mechanism is taken at face value.","tokens_in":16440,"tokens_out":2210,"would_cite":true,"duration_ms":23298,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"World action models should predict the future in motion, geometry, and semantics during training—not just RGB—and this structured dreaming improves robustness to visual change without altering deployment.","keywords":["world action models","future prediction","optical flow","gated residual branches","robotic manipulation robustness","visual distribution shift","video diffusion transformer","multi-view supervision"],"falsifier":"Train DreamWAM with the motion target replaced by color-coded static frames or noise that goes through the identical VAE pipeline; if LIBERO-Plus success stays near 75.47%, the motion view is not what carries the gain. A second check: compare against a variant that supervises raw flow vectors in a learned latent space instead of color-coded flow, and see which version keeps the perturbation gains.","tokens_in":15323,"feed_emoji":"🤖","tokens_out":4678,"duration_ms":46273,"temperature":0.7,"pith_summary":"This paper argues that a world action model—a policy that learns to act by predicting how the observed world will evolve—should be trained to imagine that future in several complementary forms at once: appearance, motion, geometry, and semantics, rather than pixels alone. The proposed method, DreamWAM, adds these extra views only during training and removes them at deployment, so at run time the robot still takes in RGB and outputs actions exactly like the baseline it improves on. The claim is that this change makes the learned representation preserve the state transitions that actually determine task success, which matters most when the visual scene shifts. Supporting evidence: average success on LIBERO-Plus perturbations rises from 69.16% to 75.47% with joint inference, and real-world success under unseen lighting, background, and layout changes rises from 55.6% to 74.4%.","feed_headline":"Dreaming in motion, depth, and meaning beats RGB-only robot foresight","feed_subtitle":"A training-only multi-view dream makes robot policies hold up under unseen lighting, background, and layout changes.","key_machinery":"Two mechanisms carry the argument. First, RGB video latents and RAFT optical-flow latents are jointly denoised in a shared VideoDiT stream: they are perturbed at the same flow-matching timestep, concatenated along the channel dimension, and predicted together, which makes temporal change an explicit part of the learned future. Second, geometry and semantic features from Depth Anything V3 and DINOv2 are injected through lightweight gated residual branches that add a gated correction to selected VideoDiT layers instead of replacing the native video state, with a prediction head that regresses those aligned features during training. Shared attention between VideoDiT and ActionDiT lets the action branch be shaped by these structured future representations. At inference, motion channels are zero-filled and all non-RGB branches are removed, so nothing changes at deployment except the weights.","core_discovery":"The paper's central claim is that robust world-action learning depends not only on predicting the future, but on representing it in a form that matters for action. Concretely, it claims that training a VideoDiT-ActionDiT world action model with four future views—RGB, optical flow, depth, and DINOv2 semantic features—reshapes the shared video-action representation so that the policy is more robust to visual distribution shifts, while the extra supervision branches can be switched off at test time with the deployment interface unchanged. The finding is stated in matched comparisons: under unseen LIBERO-Plus perturbations, the no-rollout policy improves from 51.36% to 63.44% and the joint video-action variant from 69.16% to 75.47%, with gains on all seven perturbation dimensions; on a real robot, average success under unseen visual perturbations improves from 55.6% to 74.4%.","pith_inferences":["If the mechanism is right, the same training-time structured supervision could be applied with richer targets—metric 3D scene flow, contact maps, or object-centric trajectories—and the residual-branch design suggests a recipe for adding heterogeneous targets to any pretrained video diffusion policy.","The results leave open whether the gains stem from the specific semantic content of each view or from the extra temporal coherence the combined objectives impose; a control with random-but-structured targets, such as color-coded noise with the same statistics, would separate these explanations.","A practical extension the paper does not test is keeping the motion branch active at test time for partially observed futures, or using the geometry and semantics branches as test-time conditioning, which might push robustness further than the RGB-only deployment they chose."],"forward_implications":["Improvements across no-rollout and joint inference modes, with a smaller no-rollout-to-joint gap (12.02 versus 17.80 percent), suggest structured supervision internalizes part of the action-relevant future that otherwise would need online generation.","The all-denoise control (D/D/D) falling below the RGB-only baseline indicates that supervision quantity alone does not help; the route by which each view enters the pretrained video pathway matters.","The ablation identifying motion as the most consequential view, with geometry and semantics underperforming when motion is removed, positions motion as a temporal scaffold for the feature-level signals.","Because the deployment interface is unchanged, the gains are additive to any future improvement in the underlying video backbone or action decoder."],"supporting_citations":[{"why":"Supplies the RAFT optical-flow estimator used to build the motion view.","marker":"(Teed and Deng 2020)"},{"why":"Supplies the Wan2.2 video VAE that encodes both RGB and color-coded flow latents onto the same grid.","marker":"(Wan et al. 2025)"},{"why":"Supplies Depth Anything V3 features that form the geometry view.","marker":"(Lin et al. 2025)"},{"why":"Supplies DINOv2 patch features that form the semantic view.","marker":"(Oquab et al. 2024)"},{"why":"Provides the Fast-WAM two-expert VideoDiT-ActionDiT architecture that DreamWAM builds on and matches in evaluation.","marker":"(Yuan et al. 2026)"},{"why":"Provides the LIBERO benchmark and task suites used for training and in-distribution evaluation.","marker":"(Liu et al. 2023)"},{"why":"Provides the LIBERO-Plus perturbation suite that defines the out-of-distribution evaluation.","marker":"(Fei et al. 2025)"},{"why":"Provides the flow-matching objective used for joint RGB-motion denoising.","marker":"(Lipman et al. 2022)"}],"fun_headline_variants":["Beyond RGB: DreamWAM's multi-view future yields 19-point real-world gain","Real-world success jumps from 55.6% to 74.4% with multi-view dream training","DreamWAM: Robots that dream in depth, motion, and semantics act sturdier","Predicting future states beyond RGB makes robot policies robust to visual shifts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The measured gains depend on the motion view, which is built by converting RAFT optical flow into color-coded RGB pictures and encoding those pictures with the Wan2.2 video VAE; if that lossy pipeline does not preserve direction or magnitude of motion, then 'motion' supervision is not the mechanism producing the improvement.","fun_headline_variants_meta":{"raw":{"variants":["Beyond RGB: DreamWAM's multi-view future yields 19-point real-world gain","Real-world success jumps from 55.6% to 74.4% with multi-view dream training","DreamWAM: Robots that dream in depth, motion, and semantics act sturdier","Predicting future states beyond RGB makes robot policies robust to visual shifts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000416,"raw_usage":{"total_tokens":2210,"prompt_tokens":1073,"completion_tokens":1137,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":689,"completion_tokens_details":{"reasoning_tokens":1045}},"tokens_in":689,"tokens_out":1137,"duration_ms":12523,"temperature":1.0,"reasoning_tokens":1045,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:54:18.009255+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DreamWAM with the motion target replaced by color-coded static frames or noise that goes through the identical VAE pipeline; if LIBERO-Plus success stays near 75.47%, the motion view is not what carries the gain. A second check: compare against a variant that supervises raw flow vectors in a learned latent space instead of color-coded flow, and see which version keeps the perturbation gains.","supporting_citations":[],"review_version":1}