{"id":"a6992a0c-e643-4965-baa0-9a8da3fecdbe","arxiv_id":"2608.06374","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DyPES-VLA trains a single cross-embodiment policy using future-video prediction to share dynamics and a per-embodiment Mixture-of-Experts head for control, reporting state-of-the-art simulation and real-world success rates.","lead":"This paper presents DyPES-VLA, a robot control model that trains one policy for several different robot bodies by learning common dynamics priors from video prediction and then translating them into each robot's native action format. It reports top results on standard simulation benchmarks and on three physical robots.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage-1 pretraining ablation is confounded: it adds 100k extra steps and in-distribution benchmark video, so the claimed shared-dynamics-prior mechanism is not established by Table 5.","rationale":"The reader's weakest assumption identifies human-video transfer, and I agree that this is untested, but the sharper problem is that the only ablation supporting Stage-1 (Table 5) is confounded with additional compute and in-distribution benchmark video. The full model gets 100k Stage-1 steps plus 200k Stage-2 steps, while the w/o Stage-1 row gets only 200k and the VLM/query tokens are not pretrained. Also, 50% of Stage-1 video is generated from the same three benchmarks used for evaluation, so even a compute-matched ablation would not separate transferable priors from domain-specific visual pretraining. The EgoDex-only condition is needed to support the cross-embodiment human-video claim. This is load-bearing because the paper's central contribution is the paradigm of shared dynamics priors plus embodiment-specific control; if the Stage-1 gain is just extra compute or in-domain pretraining, the method reduces to a standard VLA with an auxiliary future loss and the cross-embodiment transfer story collapses. I did not find an internal inconsistency in the architecture, and the empirical point estimates, if accurate, are plausible; hence the verdict remains conditional pending the ablation and release of code and data.","tokens_in":15140,"tokens_out":8455,"duration_ms":83112,"concrete_test":"Retrain the \"w/o Stage-1 pretraining\" condition with the full optimization budget (300k steps, batch 512) and add two Stage-1 data ablations: EgoDex-only and simulation-video-only, with the future objective otherwise identical. If the compute-matched no-Stage-1 control matches full DyPES-VLA, or if EgoDex-only does not outperform the no-Stage-1 control, then Table 5's gains are explained by extra in-distribution training and the shared-dynamics-prior claim fails. Report per-seed means and confidence intervals for these conditions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sections 3.5 and 4.4, the evidence that future-supervised dynamics priors drive cross-embodiment gains is the Stage-1 ablation in Table 5. But Stage 1 is pretrained on a mixture that is 50% EgoDex and 50% video rendered from the exact simulation benchmarks used for evaluation (RoboTwin 2.0, RoboCasa-GR1, LIBERO), and it runs for 100,000 additional optimization steps before Stage 2 begins. The row \"w/o Stage-1 pretraining\" therefore differs from the full model in at least three ways: total training steps (200k vs 300k), initialization of the VLM/query tokens, and prior exposure to the evaluation environments' visual statistics. The reported gain cannot be attributed specifically to transferable shared dynamics priors; it may be entirely extra in-distribution pretraining or additional compute. The EgoDex contribution is never isolated, so the paper's central causal claim, that action-free human and robot video pretraining produces dynamics priors that transfer to robot action generation, is not supported by the provided ablations. The linear probe in Section 4.5 only shows that future supervision improves decodability of contact transitions on LIBERO, which is not evidence that Stage-1 video transfers across embodiments or improves downstream control.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DyPES-VLA, a two-stage cross-embodiment VLA that learns a shared query representation through future-frame prediction and then decodes it into native action spaces through an embodiment-specific Mixture-of-Experts action head. Stage 1 pretrains the VLM, query tokens, and a SANA future-generation head on action-free human and simulation videos; Stage 2 jointly optimizes the future-prediction and action objectives on action-labeled demonstrations from RoboTwin 2.0, RoboCasa-GR1, and LIBERO. The resulting single checkpoint is evaluated on the three simulation benchmarks and, after joint finetuning, on three physical robots (FR3, COBOT Magic, G1), with reported success rates of 89.02%, 59.25%, and 98.0% in simulation and a 75.6% average in the real-world study. Ablations target the future-prediction objective, Stage-1 pretraining, the MoE head, and embodiment metadata, and a linear probe tests contact decodability.","tokens_in":15540,"tokens_out":8107,"duration_ms":76106,"significance":"If the results hold, the paper makes a useful architectural contribution: separating a future-supervised shared representation from embodiment-specific action decoding is a clean way to avoid manual action-space alignment, and the two-stage training protocol is clearly specified. The paper also has genuine strengths: the future-prediction objective is an independent training signal rather than a circular byproduct of action prediction, the linear probe is carefully designed to avoid trivial prediction by conditioning on the current contact state, and the evaluation spans three morphologies in both simulation and real hardware, with per-task breakdowns in the supplementary material. However, the central claim that action-free human/robot video pretraining produces transferable shared dynamics priors is not yet established, because the Stage-1 ablation is confounded, and the headline state-of-the-art claims rest on margins that are often below what can be resolved without reported variance. With cleaner ablations and error bars, this would be a solid contribution.","major_comments":[{"comment":"The 'w/o Stage-1 pretraining' ablation cannot support the paper's central causal claim about shared dynamics priors. Stage 1 adds 100,000 optimization steps, changes the initialization of the VLM/query tokens, and its video mixture contains 50% simulation clips rendered from exactly the three evaluation benchmarks (RoboTwin 2.0, RoboCasa-GR1, and LIBERO, per Section 4.1). The reported drop under this ablation is therefore consistent with additional compute or in-distribution visual pretraining; the contribution of EgoDex human video is never isolated. I request an ablation that matches total compute and initialization while varying only the presence of EgoDex, for example by comparing Stage-1 pretraining on simulation videos alone versus simulation videos plus EgoDex, or by training a 300k-step no-Stage-1 baseline.","section":"§4.4, Table 5"},{"comment":"No variance or number of seeds is reported for any simulation result, despite several headline margins being very small: 0.95 points over ABot-M0 in RoboCasa-GR1 (Table 2), 0.4 points over Fast-WAM in LIBERO (Table 3), and 0.1 points below the best finetuned X-VLA in LIBERO. Without seed-to-seed or rollout variance, the 'state-of-the-art' claim is not statistically supported. Please report means with standard deviations or confidence intervals over at least three training seeds, and state the evaluation rollout counts in the table captions.","section":"§4.2, Tables 1-3"},{"comment":"The real-world comparison reports point estimates from 25 rollouts per cell with no variance, confidence intervals, or protocol details for ensuring identical evaluation conditions across methods. The G1 rows for ACT are 0% on two tasks, and the 16.0-point margin over GR00T-N1.6 would be much more convincing with a per-rollout breakdown and a statement of whether object placements, water amounts, and success criteria were matched across all compared policies.","section":"§4.3, Table 4"},{"comment":"The linear probe demonstrates only that future supervision improves decodability of contact transitions on LIBERO. It does not measure cross-embodiment transfer, nor does it test whether EgoDex human video is the source of the improvement. Since the claimed mechanism is that shared dynamics priors are learned from action-free human and robot video and then transfer across embodiments, the probe should either be run on held-out embodiments or be complemented by an ablation that isolates the EgoDex contribution to downstream control.","section":"§4.5, Table 6"}],"minor_comments":[{"comment":"The row labels 'w/o future prediction' and 'w/o Stage-1 pretraining' are easy to misread. Please state explicitly whether the former removes the future loss only from Stage 2 or from both stages, and whether Stage 1 is still run in that variant.","section":"Table 5 and §4.4"},{"comment":"The text says 'We use the full EgoDex corpus' and also that the video mixture is '50% EgoDex, 20% RoboTwin 2.0, 20% RoboCasa-GR1, and 10% LIBERO.' Please clarify how the 'full corpus' interacts with the 50% mixing ratio, and report the resulting number of EgoDex frames or sequences actually used.","section":"§4.1"},{"comment":"Please state the total scale of the Stage-1 and Stage-2 training data (number of videos, frames, or action sequences), since the compute-versus-data tradeoff is central to interpreting the Stage-1 ablation.","section":"§4.1"},{"comment":"The Beta schedule is introduced only in prose as τ = s(1-u) with u ~ Beta(1.5, 1.0) and s = 0.999; please define s and u explicitly near Eq. (4) and state whether the same schedule is used for the future-generation head in Eq. (2).","section":"§3.4"},{"comment":"The future offset Δ_e is said to match the action horizon H_e. Please clarify how Δ_e is defined for the action-free Stage-1 videos, where there is no action horizon, and whether the prediction target is exactly H_e frames ahead or a different offset.","section":"§3.3"},{"comment":"Figure 2 is visually dense; labeling the Stage-1 and Stage-2 branches directly in the figure, or splitting it into two panels, would help the reader map the textual description to the architecture.","section":"Figure 2"},{"comment":"The 59.25% versus 58.3% comparison with ABot-M0 is a 0.95-point margin on a point estimate; please qualify this as within expected noise until variance is reported.","section":"Abstract and §4.2"},{"comment":"The manuscript does not state whether code, model weights, or evaluation scripts will be released; please add an availability statement describing what will be made public.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The related work cites several 2025/2026 arXiv preprints, some from the same research group (e.g., Li et al. 2026a, Yan et al. 2026, Zhong et al. 2026, Chen et al. 2026). The editor may wish to verify that these references are publicly available and that the comparison numbers in Tables 1-3 are drawn from accessible sources. I saw no evidence of misconduct, but the heavy self-citation pattern and the absence of a code release warrant a reproducibility check. The topic fits a robotics/ML venue; the main risk is the confounded Stage-1 ablation, which is fixable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The architecture is genuinely new: future-supervised query states feeding an embodiment-routed MoE action head, with no manual action alignment. If the numbers hold, a single checkpoint across single-arm, dual-arm, and humanoid robots is a useful advance. The evidence is less clean than the paper claims, because the Stage-1 ablation is confounded. But this is fixable and the paper is worth refereeing.\n\nWhat it does well: the design is coherent and the ablations are internally consistent. Removing future supervision, Stage-1 pretraining, the MoE head, or metadata all degrade performance, which supports the architectural choices. The real-world experiment across three physical embodiments is a real plus, and the linear probe for contact onset/release is a thoughtful diagnostic. The write-up is clear about the two-stage training and the inference-time removal of the future head.\n\nThe soft spots are real but not fatal. The Stage-1 ablation does not isolate the mechanism: 'w/o Stage-1 pretraining' differs by 100k optimization steps and the Stage-1 mixture includes 50% simulation video from the exact evaluation benchmarks. So the gain over that row could come from extra compute or in-distribution visual exposure, not from transferable dynamics priors. The EgoDex contribution is never separated, so the paper overclaims when it says the experiments demonstrate that dynamics priors transfer across embodiments. What is demonstrated is that more pretraining on a video mixture helps. Also, simulation results are point estimates with no error bars or seeds; two SOTA margins are under one point (0.95 on RoboCasa, 0.4 on LIBERO). No code or data are released, so the headline numbers are not independently checkable. Real-world rollouts are 25 per cell with no variance, which is standard but still worth noting.\n\nNone of this kills the paper. The core empirical pattern — one checkpoint handles three embodiments with native action spaces — is the main claim and it stands relative to the reported baselines. The causal story about dynamics priors is the part that needs tighter evidence.\n\nFor anyone working on cross-embodiment VLA, this is worth reading. Send it to peer review. Ask for a revised ablation that controls for compute and isolates the human-video contribution, plus error bars on the main tables.","headline":"A solid cross-embodiment VLA with a genuinely new architecture, but the flagship claim about shared dynamics priors rests on a confounded ablation; still deserves a serious referee.","tokens_in":16003,"tokens_out":2093,"would_cite":true,"duration_ms":20692,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single checkpoint controls single-arm, dual-arm, and humanoid robots by learning shared dynamics before embodiment-specific actions.","keywords":["cross-embodiment manipulation","vision-language-action models","future prediction","mixture-of-experts","shared dynamics priors","embodiment-specific control","robot manipulation"],"falsifier":"Ablate only the human-video portion of the first pretraining stage's video mixture: retrain with the same total video budget drawn entirely from the three robot simulation data sources, keeping every other hyperparameter fixed, and compare the three benchmark scores and the real-world average against the reported numbers. If removing human video leaves performance unchanged, the claim that egocentric human dynamics transfer to robot embodiments fails, even though future prediction itself may still help.","tokens_in":14936,"feed_emoji":"🤖","tokens_out":7321,"duration_ms":67440,"temperature":0.7,"pith_summary":"The paper proposes DyPES-VLA, a way to train one vision-language-action policy that controls robots with very different bodies—single arm, dual arm, and humanoid—without manually converting their actions into a common format. It argues that future-frame prediction, applied to action-free human and robot videos, makes the shared visual representation carry the dynamics of contact, object motion, and scene change. A second claim is that action generation should not be shared: an embodiment-specific mixture-of-experts head decodes the shared representation directly into each robot's native action space. If correct, this separates the part of manipulation that transfers across embodiments from the part tied to a specific body, allowing a single checkpoint to match or beat both per-benchmark specialists and generalist baselines. The paper reports 98.0% success on a single-arm simulation benchmark, 59.25% on a humanoid benchmark, and 89.02% on a dual-arm benchmark, plus 75.6% average success across three real-world robots.","feed_headline":"One model controls single-arm, dual-arm, and humanoid robots","feed_subtitle":"Future-video pretraining plus per-embodiment action experts beats separate specialists in simulation and on real robots.","key_machinery":"The load-bearing object is the shared query state $Z$: a set of learnable query tokens processed by a pretrained vision-language model together with the observation, instruction, and verbalized embodiment metadata. In the first training stage these states are supervised only by a rectified-flow future-frame generation loss, with the generator conditioned solely on $Z$, so predictive regularities must be stored in $Z$ itself. In the second stage the same states condition an embodiment-specific mixture-of-experts action head, a flow-matching diffusion transformer whose attention layers are shared and whose feed-forward experts are chosen by static routing; lightweight per-embodiment encoders and decoders map between the shared width and each native action space. The future-generation head is discarded at inference. This mechanism's job is to keep the representation embodiment-neutral at the level of dynamics while keeping control realization embodied.","core_discovery":"The central claim is that cross-embodiment manipulation can be decomposed into shared dynamics priors and embodiment-specific control. DyPES-VLA trains a pretrained vision-language model together with learnable query tokens to predict a future camera frame; because the future-generation head sees only those query states, whatever is needed to synthesize the future must pass through them, and those states become the shared interface between perception and action. A flow-matching action head, with attention layers shared across embodiments and feed-forward experts selected by embodiment metadata, then decodes the same query states directly into action chunks in each robot's native action space. This lets heterogeneous robots be co-trained without pre-aligning their action formats. The paper claims that, with a single co-trained checkpoint, this paradigm outperforms prior specialists and generalists on all three simulation benchmarks and, after joint finetuning on 1,800 real-world demonstrations, averages 75.6% success across three physical robot morphologies.","pith_inferences":["The separation of shared dynamics from embodiment-specific control suggests a scaling recipe: add a new robot family by adding one expert encoder-decoder pair to the action head and adding its data to future-prediction pretraining, without re-engineering a common action representation.","A strong test the paper leaves implicit is whether a genuinely novel embodiment, absent from the second-stage co-training, can be controlled after only expert initialization plus a small amount of its own data; the real-world finetuning reuses experts already trained on the matching simulation embodiment, so it does not demonstrate true zero-shot body transfer.","A concrete testable extension is to ablate the source mixture of the first-stage videos, comparing simulation-only, human-video-only, and mixed pretraining; the current ablation removes the whole future-prediction objective, not the human-video component, so it does not isolate where the transfer gain comes from.","If the dynamics-prior story is right, the same future-supervised query states could seed other downstream heads, such as subgoal proposal or grasp-point prediction, without retraining the shared backbone."],"forward_implications":["A single checkpoint can control robots with different morphologies, degrees of freedom, and control frequencies without action-space alignment.","Action-free video sources, including egocentric human manipulation videos, can contribute to robot policy learning through future prediction even though they contain no executable actions.","Embodiment metadata lets one policy disambiguate which robot is acting, and static routing prevents heterogeneous control spaces from interfering during joint training.","Because the future-generation head is removed at inference, the cross-embodiment transfer gains come without additional test-time video generation cost.","Future supervision makes contact onset and release linearly decodable from the query states, evidence that the representation tracks interaction dynamics rather than only visible current contact."],"supporting_citations":[{"why":"Supplies the pretrained vision-language model that produces the shared query states.","marker":"Bai et al. (2025)"},{"why":"Supplies the image-generation architecture and frozen autoencoder used as the future-generation head.","marker":"Xie et al. (2024)"},{"why":"Supplies the rectified-flow objective used for both future generation and action generation.","marker":"Lipman et al. (2022)"},{"why":"Supplies the action-free egocentric human manipulation videos used in the first pretraining stage.","marker":"Hoque et al. (2025)"},{"why":"Provides the humanoid simulation benchmark setting and a per-benchmark specialist and real-world baseline.","marker":"Bjorck et al. (2025)"},{"why":"Provides the dual-arm simulation benchmark with strong domain randomization used for co-training and evaluation.","marker":"Chen et al. (2025a)"},{"why":"Provides the single-arm simulation benchmark used for co-training and evaluation.","marker":"Liu et al. (2023)"}],"fun_headline_variants":["Shared dynamics priors plus embodiment experts unify robot control","Future-video pretraining enables one policy for many robot bodies","No action alignment needed: one VLA for all robot morphologies","Cross-embodiment VLA: co-train different robots without action reformatting","Learn shared dynamics, keep per-robot experts: one policy fits all"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that teaching the model to predict future frames from action-free videos, including egocentric human hand videos, writes shared dynamics into the representation that truly transfer to robot control; if human-hand dynamics do not transfer to robot arms and humanoids, the first pretraining stage is wasted or harmful and the reported gains would need another explanation.","fun_headline_variants_meta":{"raw":{"variants":["Shared dynamics priors plus embodiment experts unify robot control","Future-video pretraining enables one policy for many robot bodies","No action alignment needed: one VLA for all robot morphologies","Cross-embodiment VLA: co-train different robots without action reformatting","Learn shared dynamics, keep per-robot experts: one policy fits all"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00045,"raw_usage":{"total_tokens":2305,"prompt_tokens":1022,"completion_tokens":1283,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":1190}},"tokens_in":638,"tokens_out":1283,"duration_ms":9560,"temperature":1.0,"reasoning_tokens":1190,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:03:17.851658+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Ablate only the human-video portion of the first pretraining stage's video mixture: retrain with the same total video budget drawn entirely from the three robot simulation data sources, keeping every other hyperparameter fixed, and compare the three benchmark scores and the real-world average against the reported numbers. If removing human video leaves performance unchanged, the claim that egocentric human dynamics transfer to robot embodiments fails, even though future prediction itself may still help.","supporting_citations":[],"review_version":1}