{"id":"275efc3c-53e8-426a-afc7-e6cb515cdeb9","arxiv_id":"2607.24159","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Decoupling video dynamics from action prediction with multi-level feature transfer and affordance/depth guidance yields more data-efficient robot policies than coupled video-action models.","lead":"DeVA splits robot policy learning into a video expert that imagines future scenes and an action expert that outputs controls, linked by multi-layer features plus affordance and depth cues. It reports stronger data-efficient manipulation results than unified video-action models and several VLAs in sim and on a real bimanual robot.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The real-world \"success rates\" (74% vs 48%/34%) are partial-credit averages over only 10 trials per task, and the strongest baseline (π0.5, 0.67) appears only in the appendix — the headline real-world margin is statistically fragile and scored on a lenient metric.","rationale":"The reader flagged pseudo-label guidance quality and capacity/tuning confounds as the weakest assumption, and judged real-world generality under-determined. I agree those are concerns, but the sharper, checkable load-bearing issue is one step downstream: the real-world evaluation itself. The pseudo-label concern only bites in real-world settings (sim uses oracle contacts), and in those settings (a) the metric is partial credit mislabeled as success rate, (b) n=10/task makes the headline margins statistically thin, (c) the strongest baseline (π0.5, 0.67) is omitted from the headline comparison, and (d) guidance is never ablated on hardware. These compound: the claim \"strong real-world performance with gains from physical guidance\" currently rests on a leniently scored, small-N comparison against a weakened baseline set. I do not recommend REJECT because the simulation evidence (RoboCasa with 3,600 rollouts, LIBERO 99.0%, LIBERO-plus 80.8% with an 11.2-point margin, plus the decoupling and guidance ablation chains) is broad and internally consistent, and the efficiency comparison (Fig. 7) is conducted under matched budgets. The paper is an accept-shaped contribution whose real-world and guidance-necessity claims need the recomputation above before they should be cited at face value — which is exactly what the reader's CONDITIONAL already encodes. Hence UNCHANGED, with the condition made concrete: binary rescoring with CIs, inclusion of π0.5, and a hardware guidance ablation.","tokens_in":21844,"tokens_out":3489,"duration_ms":94446,"concrete_test":"From the released rollout logs/videos, recompute Table 6 and Fig. 6(b) as binary success (episode credited only at score 1.0 per the C.2 criteria), report bootstrap 95% CIs over the 10 trials per task, and add π0.5 to the main-text comparison. If DeVA's binary success is not significantly above π0.5's (or the ranking vs GR00T-N1.6 flips on any task), narrow the real-world claim. As a second step, run one real-world ablation (10 trials/task, Handover Marker) with the affordance/depth guidance tokens removed, to test whether pseudo-label guidance contributes on hardware at all.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim leans on three pillars: sim benchmarks, efficiency/convergence, and real-world deployment. The sim pillar is solid (3,600 RoboCasa rollouts, 3 seeds; LIBERO/LIBERO-plus standard protocols; component ablations). The real-world pillar is not. Appendix C.2 defines the evaluation as stage-completion partial credit in [0,1] (e.g., Lift Pot awards 0.25/0.5/0.75 for incomplete episodes), yet Fig. 6(b) and Table 6 label these numbers \"Success Rate.\" The values confirm this: with 10 trials per task, binary success rates must be multiples of 0.1, but the table reports 0.88, 0.45, 0.23, 0.68, 0.65 — impossible under binary scoring. So \"real bimanual average 74%\" is a partial-credit score, not a success rate, and it inflates every method differently depending on where partial credit is granted (the 0.5 bin for \"one gripper lifts\" is generous). Second, n=10 trials/task/method gives a standard error of ~8-10 points per task; the DeVA-vs-GR00T gap (0.74 vs 0.48) is only ~2σ, and the appendix's own Table 6 shows π0.5 at 0.67 — within noise of DeVA's 0.74 — yet π0.5 is absent from the main-text Fig. 6(b) comparison and from the abstract-level claim. Third, the physical-guidance gains (66.0→72.0) are ablated only in simulation, where affordance targets are oracle simulator contacts; in the real world the guidance uses off-the-shelf pseudo-labels [50,57] whose contribution is never ablated on hardware, so \"clear performance gains from physical guidance\" is unverified exactly where the labels are noisy. If binary success on hardware shows DeVA ≈ π0.5, the real-world claim reduces to \"competitive with a strong VLA,\" and the guidance claim rests entirely on sim-oracle supervision.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper presents DeVA, a video-action model for language-conditioned manipulation in which a pretrained video diffusion transformer (Cosmos-Predict2, 2B params) and a separate flow-matching action expert are trained jointly but kept parameter-decoupled, exchanging information through layer-wise cross-attention from eight video layers plus learnable bridge tokens. Intermediate video features are additionally supervised with task-conditioned affordance heatmaps and relative depth via lightweight DPT-style decoders, whose features are injected as guidance tokens into the action expert's cross-attention. The authors report: (i) RoboCasa 72.0% success with only 50 demos/task, exceeding baselines trained on 300–3,000 demos; (ii) 99.0% average on LIBERO and 80.8% on LIBERO-Plus (+11.2 over the best baseline); (iii) faster convergence and much higher success than a matched unified-backbone variant (66–71% vs 34% at ~40–50K steps); (iv) real-world bimanual results on three tasks with an average \"success rate\" of 74% vs 48% (GR00T-N1.6) and 34% (Cosmos-Policy); and (v) ablations attributing ~6 points on RoboCasa to affordance+depth guidance. The simulation evidence is substantial; the real-world evidence is weaker than the headline numbers suggest, for reasons detailed below.","tokens_in":22315,"tokens_out":3688,"duration_ms":31017,"significance":"If the results hold, DeVA is a useful contribution to the video-action-model line of work: it gives a concrete, ablated answer to how to interface a pretrained video diffusion backbone with an action expert (layer-wise cross-attention plus bridge tokens, multi-step denoising features), and the RoboCasa data-efficiency result (72.0% at 50 demos/task, with 3,600 rollouts over 3 seeds) and the LIBERO-Plus robustness margin (+11.2 over the strongest baseline) are practically meaningful. The component ablations (Fig. 8), coupled-vs-decoupled convergence comparison under matched budget (Fig. 7(b)), and detailed appendix (parameter accounting in Table 2, per-dataset normalization in Table 4, denoising-step and interaction-mechanism ablations in Tables 7–9) are the kind of evidence that lets others build on the work. The promised release of code, configs, and checkpoints, and the documented real-world evaluation criteria (App. C.2), are genuine reproducibility strengths. The contribution is incremental relative to concurrent dual-DiT VAMs [9,10] in architecture, but the physical-guidance mechanism and the systematic efficiency analysis differentiate it.","major_comments":[{"comment":"Fig. 6(b) and Table 6 vs. Appendix C.2: the quantity reported as 'Success Rate' for real-world deployment is not a success rate. Appendix C.2 defines a stage-completion partial-credit score in [0,1] (e.g., Lift Pot awards 0.25/0.5/0.75 for incomplete episodes; Handover Marker awards 0.5 when one gripper lifts but the transfer fails). With 10 trials per task, a binary success rate must be a multiple of 0.1, yet Table 6 reports 0.88, 0.45, 0.23, 0.68 — values only possible under partial credit. This is load-bearing for the abstract and §4.4 claim 'DeVA achieves an average success rate of 74%, compared with 48%/34%': partial credit inflates all methods, and it does so differentially depending on where credit is granted (e.g., the 0.5 bin for 'one gripper lifts' is generous and may reward failure modes unevenly across methods). The fix is straightforward but mandatory: relabel the metric as","section":"§4.4, Fig. 6(b), Table 6, Appendix C.2"},{"comment":"Table 6 and Fig. 6(b): the real-world comparison in the main text omits the strongest baseline. Table 6 (appendix) reports π0.5 at 0.67 average vs. DeVA's 0.74, yet Fig. 6(b) and §4.4 compare only against GR00T-N1.6 (0.48) and Cosmos-Policy (0.34), and the abstract-level real-world claim rests on that comparison. With n=10 trials per task per method, the per-task standard error under the partial-credit metric is roughly 8–12 points, so the 0.74 vs. 0.67 gap over π0.5 is well within noise, and even the 0.74 vs. 0.48 gap over GR00T-N1.6 is only about 2σ on the per-task level. π0.5 must be included in the main-text figure and discussion (including, per §C.3, the caveat that π0.5 brings large-scale robot-centric pretraining and a 50-step action horizon vs. 25 for DeVA), and the claims should be tempered to what n=10×3 tasks can support — or the trial count increased. As written, the headline","section":"§4.4, Fig. 6(b), Table 6"},{"comment":"Fig. 8(b) and Appendix B.3: the claim of 'clear performance gains from physical guidance' is established only in simulation, where affordance targets are oracle MuJoCo contacts/projected end-effector positions (§B.3). In the real-world pipeline the affordance targets are pseudo-labels from an off-the-shelf model [50] used 'out-of-the-box' and depth targets come from [57]; the authors themselves note these contain 'moderate prediction noise.' No real-world ablation isolates the contribution of physical guidance, so it is possible that on hardware the guidance tokens add little or that noisier pseudo-labels erode the sim-measured 6-point gain (66.0→72.0). The paper should either add a real-world ablation (base vs. +guidance on at least one task) or explicitly scope the physical-guidance claim to simulation in the abstract and §7. Given that the guidance module is a named contribution of th","section":"§5, Fig. 8(b), Appendix B.3"},{"comment":"Fig. 7(b): the 'faster convergence than a unified architecture' claim is supported only over the plotted horizon (~60K steps). The coupled variant is at ~34% at 40–50K steps when the decoupled variants plateau at 66–71%, but the figure does not show whether the coupled model eventually catches up with more steps, i.e., whether decoupling buys asymptotic performance or only optimization speed. This matters because the coupled variant is the paper's own construction (not a published baseline), so the reader cannot rule out that it is simply under-tuned. Please either extend the coupled curve to convergence, or state explicitly that the claim is about matched-budget efficiency, and report the coupled variant's final performance under its own best budget.","section":"§4.5, Fig. 7(b)"}],"minor_comments":[{"comment":"§3.3 defines Stage-1 loss weights λ_aff and λ_depth, but their values appear only in Appendix Table 5 (both 1.0). State the values in the main text where the objective is introduced.","section":"§3.3 / Table 5"},{"comment":"Simulation results use 3 seeds × 50 rollouts, but no variance or confidence intervals are reported in Table 1 or Fig. 5. On LIBERO, DeVA's 99.0 vs. DiT4DiT's 98.6 and Cosmos-Policy's 98.5 is almost certainly within seed noise; please report standard errors so that 'highest overall average' claims are calibrated.","section":"Table 1, Fig. 5"},{"comment":"Fig. 5(c) ('Perturbation') is never referenced or described in the text; the caption mentions 'Robot Init' and 'Camera View' axes but no numbers or discussion are given. Either integrate it into §4.3 or remove it.","section":"Fig. 5(c)"},{"comment":"Fig. 3 contains a label 'Bit Addition' (presumably an addition/fusion op); clarify or correct the terminology.","section":"Fig. 3"},{"comment":"Appendix D.3 reports 30K-step checkpoint results (49.6/53.2/22.4) that are far below the ~55K-step numbers elsewhere (e.g., 66.0 base). Please note the different budget explicitly in the table caption to avoid confusion with main-text numbers.","section":"Appendix D.3, Table 9"},{"comment":"Formatting/typos: 'DeV A' appears with a spurious space throughout the extracted text (e.g., title, §3, §4.1); Table 1's last row reads 'Ours5072.0' with missing column separation; §4.1 says 'three complementary simulation benchmarks' and the introduction says 'three simulation benchmarks, including RoboCasa, LIBERO, LIBERO-plus' — 'including' suggests the list is non-exhaustive; make it exact.","section":"General"},{"comment":"§B.1: training uses Cosmos-Policy's replay-filtered (success-only) demonstrations for LIBERO and RoboCasa. It would help to state explicitly which baselines in Table 1 and Fig. 5(a) were trained on the same filtered data versus their original data, since filtering changes the effective data budget (1,199 episodes for RoboCasa across 24 tasks).","section":"Appendix B.1"}],"recommendation":"major_revision","confidential_remarks":"The simulation evidence is strong and I do not doubt the architecture results. My concern is concentrated in the real-world section: the headline \"74% success\" is a partial-credit score over 10 trials/task, and the strongest baseline (π0.5, 0.67) appears only in an appendix table while the main-text figure compares against two weaker methods. This pattern—whether intentional or not—has the effect of overstating the real-world margin, and the editor may wish to require the labeling fix and main-text inclusion of π0.5 as a condition of acceptance. The paper otherwise appears to be a solid CoRL/RSS-style contribution; if this venue expects only simulation-level rigor for real-world demos, major comments 1–2 could be downgraded to mandatory presentation fixes."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is architectural and empirical, not conceptual: fully decoupled video and action DiTs, multi-level cross-attention plus bridge tokens, and affordance/depth tokens injected into the action stream give faster convergence and better demo efficiency than a matched coupled backbone, with strong RoboCasa/LIBERO numbers at 50 demos/task.\n\nWhat is actually new is the package, not the ingredients. Dual-DiT VAMs and physical grounding already exist; DeVA’s contribution is systematic multi-level transfer, the two-stage warmup that freezes the physical decoders while still shaping video features, and the clean coupled-vs-decoupled curves under a fixed budget. Those ablations (Fig. 7–8) are the best part of the paper. Sim evaluation is serious: 3,600 RoboCasa rollouts, standard LIBERO/LIBERO-Plus protocols, and component isolations that actually move the needle (action-only 19.8 → video+decoupled 66 → +aff/depth 72). Citations are fair to Mimic-Video, DiT4DiT, Cosmos-Policy, and the affordance/depth literature. Appendix detail on architecture, losses, and data filtering is above average for the area.\n\nSoft spots, in proportion. The real-world pillar is thinner than the abstract’s “clear gains” language. Appendix C.2 scores stage-completion partial credit in [0,1], yet Fig. 6 and Table 6 call the numbers success rates; non-multiples of 0.1 confirm it. n=10 trials/task, and π0.5 at 0.67 sits in the appendix within noise of DeVA’s 0.74 while missing from the main comparison figure. Physical-guidance gains are ablated only in sim with oracle contacts; real hardware uses off-the-shelf pseudo-labels with no hardware ablation. None of that sinks the paper—the sim and efficiency claims stand—but it means the deployment claim should be read as “competitive under matched fine-tuning,” not a decisive hardware win.\n\nThis is for people building VLA/VAM stacks who care about optimization tractability and demo budgets. It deserves a serious referee. I would engage, cite the architecture and efficiency results, and discount the real-world margin until binary success and fuller baselines appear.","headline":"Solid modular VAM recipe with real efficiency evidence; sim story holds, real-world headline is softer than the abstract implies.","tokens_in":23211,"tokens_out":562,"would_cite":true,"duration_ms":11834,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Decoupling video prediction from action prediction, then guiding both with affordance and depth, yields stronger robot policies from fewer demos.","keywords":["Video-Action Model","robot manipulation","video generative models","decoupled experts","affordance","depth guidance","policy learning","Vision-Language-Action"],"falsifier":"Train a capacity-matched unified backbone and a decoupled model without affordance/depth guidance on the same filtered demos and budget; if success and convergence gaps on RoboCasa, LIBERO, and the three real bimanual tasks disappear, the central architectural claim fails.","tokens_in":22746,"feed_emoji":"🤖","tokens_out":906,"duration_ms":16991,"temperature":0.7,"pith_summary":"Language-conditioned robot policies need to anticipate how a scene will change when the robot moves, but standard vision-language models are pretrained mostly on static images and leave dynamics to scarce robot demos. Video generative models already encode those dynamics by predicting future frames. This paper argues that the useful way to use them is not to force video and actions through one shared backbone, and not to treat the video model as a black-box encoder. Instead it builds two specialized experts: a video expert that imagines future observations, and an action expert that outputs control. Intermediate video features at many layers are passed to the action expert through cross-attention and compact bridge tokens, and the video features are further shaped by affordance and relative-depth supervision so the action stream sees where to interact and how the scene is laid out in depth. On kitchen and tabletop simulation suites and on real bimanual tasks, the design reaches high success with only tens of demonstrations per task, converges faster than a matched unified architecture, and gains measurably from the physical guidance.","feed_headline":"Separate video and action experts beat shared robot backbones","feed_subtitle":"Multi-level video features plus affordance and depth guidance lift success from limited demos","key_machinery":"The decoupled video-action architecture: a pretrained video diffusion expert and a separate action diffusion expert, connected by layer-wise cross-attention plus learnable bridge tokens, with intermediate video features also decoded into task-conditioned affordance and relative-depth maps that supervise the video stream and are injected as guidance tokens into the action expert.","core_discovery":"DeVA shows that robot policy learning from video generative priors works better when video dynamics and action prediction are handled by separate transformer experts linked by multi-level feature transfer and by explicit affordance and depth guidance, rather than by a single shared backbone or by sparse, single-layer video conditioning. Under limited demonstrations this yields higher success, faster optimization, and clear gains from the physical signals on both simulation benchmarks and real bimanual hardware.","pith_inferences":["If the physical decoders stay useful after freezing in stage two, similar auxiliary geometry or contact heads could regularize other generative backbones used only as frozen feature banks.","The real-robot gains with thirty demos per task suggest the same recipe may transfer to other multi-camera bimanual setups where large action pretraining corpora are unavailable.","A natural next stress test is whether multi-level transfer still helps when the video expert is trained on much longer horizons or open-world video rather than short task clips."],"forward_implications":["Video world-model pretraining can be turned into data-efficient robot policies without forcing video and actions into one shared representation.","Multi-layer video features plus compact bridge tokens are a practical interface for transferring predictive dynamics into control heads.","Lightweight affordance and depth supervision can steer generative video features toward interaction regions and geometry that matter for manipulation.","Under matched optimization budgets, specialized experts can converge faster than unified video-action backbones on the same tasks."],"fun_headline_variants":["Decoupled video-action experts beat shared robot backbones","Multi-level video transfer plus affordance depth aids policies","Separate experts speed robot learning from limited demos","Physical guidance lifts success in decoupled video-action model","Video dynamics and actions split for stronger robot policies"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The claim depends on intermediate video features, shaped by imperfect affordance and depth labels, actually carrying the right control information across the multi-level interface rather than the gains coming mostly from extra model capacity or training setup.","fun_headline_variants_meta":{"raw":{"variants":["Decoupled video-action experts beat shared robot backbones","Multi-level video transfer plus affordance depth aids policies","Separate experts speed robot learning from limited demos","Physical guidance lifts success in decoupled video-action model","Video dynamics and actions split for stronger robot policies"]},"model":"grok-4.5","effort":"low","cost_usd":0.002659,"raw_usage":{"total_tokens":1025,"prompt_tokens":764,"num_sources_used":0,"completion_tokens":80,"cost_in_usd_ticks":26588000,"prompt_tokens_details":{"text_tokens":764,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":181,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":764,"tokens_out":80,"duration_ms":4014,"temperature":1.0,"reasoning_tokens":181,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T22:20:04.996419+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train a capacity-matched unified backbone and a decoupled model without affordance/depth guidance on the same filtered demos and budget; if success and convergence gaps on RoboCasa, LIBERO, and the three real bimanual tasks disappear, the central architectural claim fails.","supporting_citations":[],"review_version":1}