{"id":"125aecc3-2020-43fc-8c55-4edfa465181c","arxiv_id":"2607.25798","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single diffusion transformer trains on tokenized robot bodies and motions to generate and optimize robot designs for unseen rewards and trajectories, outpacing evolutionary search in speed and often in reward.","lead":"One AI model takes a robot's body, movements, and controls as a single stream of tokens, then designs new robot bodies for movements it has never seen. This could replace slow trial-and-error robot design with fast, zero-shot optimization, and the authors built a real robot arm that tracked motions with about 73% lower error.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim hinges on reward-prediction fidelity; the paper's own quadruped results show model-based optimizers underperform CMA-ES on tracking error, so zero-shot optimization across all design spaces is not established.","rationale":"The reader's weakest assumption—that predicted RoboToken dynamics are faithful enough for reward ranking and gradient guidance—is exactly the load-bearing point. The paper's own Fig. 7 acknowledges lower correlation for Quadruped, and Tables 4/8/11 show that CMA-ES beats both Zeroth and DGS on position error in that space. This is not a minor metric discrepancy: it undercuts the abstract's blanket 'improving performance' claim and the generality of the central contribution. The hand-tuned guidance scales and stochastic DDIM also weaken the 'zero-shot' claim, since those hyperparameters are not derived from the unseen reward itself. I do not think this rises to rejection: the fixed-base and bimanual results are substantially stronger, and the hardware validation is a useful sanity check despite lacking error bars. Conditional acceptance, with the recommendation to either narrow the claims to the empirically supported design spaces or provide the missing fidelity analysis, is appropriate. The reader's verdict already reflects this, so no adjustment is needed.","tokens_in":21020,"tokens_out":3304,"duration_ms":36045,"concrete_test":"On the Quadruped space, use the released checkpoints (or retrain per §2.2) and a held-out set of unseen reward functions and trajectories. Generate 128 designs with DGS (n=128, guidance scales from §8.4/§9.4) and 128 with Zeroth. For each, compute model-predicted reward and MuJoCo reward using the 128 RL experts. Check (1) Spearman rank correlation between predicted and actual rewards; (2) actual reward of top-1 DGS vs top-1 Zeroth vs CMA-ES at matched wall-clock or sample budget. If rank correlation < ~0.5 or DGS top-1 is not significantly better than Zeroth top-1, then the dynamics model is not faithful enough to support the claimed zero-shot optimization, and the conclusion should be restricted to fixed-base and bimanual spaces.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Transformer Transformer's central assertion is that a single learned dynamics model can replace the simulator for design evaluation and that its reward gradients improve designs (§2.3). This requires the model's predicted rewards to rank designs the way MuJoCo does, especially for dynamics-based terms. The paper's own Fig. 7 and Tables 4, 8, and 11 show this is weakest in the Quadruped space: CMA-ES achieves 1.9 cm position error vs 2.6 cm for both Zeroth and DGS (Table 4), and 2.1 cm vs 2.8/3.5 cm for the velocity reward (Table 11). Thus the abstract's blanket claim of 'improving performance ... over the evolutionary baseline' is contradicted in a full design space. The DGS gradient is also not shown to add value: in several quadruped rows DGS equals or underperforms Zeroth (Tables 4, 8, 11), and the method relies on per-space/per-reward guidance scales (50/100/0.2/500) and stochastic DDIM η=1.0 tuned for DGS. These hyperparameters are not specified for unseen rewards; without them zero-shot behavior is underspecified. Since no code/data are released, the possibility remains that reported gains rely on favorable guidance tuning rather than on faithful dynamics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Transformer Transformer, a diffusion-transformer-based generative model over RoboTokens, a tokenization of robot embodiments, states, and actions. The same architecture is trained with masked modeling to perform motion-to-robot generation (diffusing embodiment and dynamics tokens conditioned on target end-effector trajectories) and cross-embodiment control. At inference, reward-agnostic dynamics predictions are converted into reward predictions, and their gradients steer the embodiment diffusion process ('Dynamics Self-Guidance'). The authors evaluate on three procedurally generated design spaces (fixed-base ViperX, quadruped manipulator, mobile bimanual), compare against random, CMA-ES, and zeroth-order diffusion sampling, and validate a fabricated ALOHA design in cloth-flinging experiments. The main claims are zero-shot optimization of unseen rewards/trajectories, improved performance and runtime over the evolutionary baseline, and unification of generator/critic/controller in one model.","tokens_in":21341,"tokens_out":8174,"duration_ms":79333,"significance":"If the claims held, the contribution would be significant: a single learned dynamics model that can generate, evaluate, and control robot designs would remove the need for separate learned generator, critic, controller, and evolutionary search, and the GPU-parallelized non-autoregressive evaluation could offer large speedups. The RoboToken representation is thoughtful, and the authors provide large-scale datasets, three diverse design spaces, and real-hardware validation (73% tracking-error reduction on ALOHA). However, the evaluation as presented does not fully support the strongest claims: the evolutionary baseline outperforms the proposed method in the quadruped single-trajectory settings, the 'zero-shot' optimization relies on per-space/per-reward tuned guidance scales, and no experiment trains a single model jointly on multiple embodiment spaces. These gaps are fixable but currently limit the paper's conclusions.","major_comments":[{"comment":"The abstract states that the method improves 'performance and runtime over the evolutionary baseline.' The runtime advantage is clear, but the performance claim is contradicted by the quadruped single-trajectory results. In Table 4 (Tracking Only), CMA-ES achieves position error 1.9 cm and 100% survival, while Zeroth and DGS achieve 2.6 cm and 99.4%; Table 8 (Tracking Torque) shows 2.0 cm vs 2.5 cm; Table 11 (Tracking Velocity) shows 2.1 cm vs 2.8/3.5 cm and lower velocity penalty. Thus the blanket performance claim is not supported in one of the three design spaces. The conclusion should be qualified or additional evidence supplied.","section":"Abstract; Tables 4, 8, 11"},{"comment":"The paper claims a 'single model to span diverse robot embodiment spaces' and a 'one-stop shop' unifying embodiment generation, control, and evaluation. However, §3 states 'we train a model for each design space, and on only the motion-to-robot task.' No experiment trains one set of weights on multiple embodiment spaces; the cross-embodiment controller experiment (§7.1) is trained and evaluated on the bimanual space only. The 11-robot tokenization demonstrates representation generality, but not a single model spanning multiple spaces. The 'unified' claim is therefore an architectural/tokenization claim, not an empirical one, and should be stated as such or supported by a joint-training experiment.","section":"§1 Contributions; §3 Approaches"},{"comment":"Dynamics Self-Guidance backpropagates gradients of reward predictions through the denoiser. This is only sound if predicted rewards rank designs the way the simulator does. The paper's own Fig. 7 caption acknowledges that predicted-actual reward correlation is lower for Quadruped, and Tables 4, 8, and 11 show DGS fails to improve over the Zeroth-Order baseline in that space (and both underperform CMA-ES). No quantitative correlation values are reported, and no analysis of off-distribution prediction error is provided. Moreover, the guidance scale is tuned per design space and per reward (§9.4: 50/100/0.2/500; §8.4-8.5: alpha_size=0.005, alpha_mass=0.005 for guidance vs 0.1/10 for the actual reward), and eta=1.0 is used only for DGS. This means the 'zero-shot' claim is underspecified for an unseen reward: the user must choose a guidance scale and stochasticity. Please provide quantitative","section":"§2.3; Fig. 7; Tables 4, 8, 11"},{"comment":"The claim that the model can 'faithfully self-validate' designs is weaker than suggested. Fig. 8 reports Pearson r=0.53, which is moderate correlation, with no confidence intervals or per-space breakdown; §7.1 shows the learned controller still trails the oracle Mink controller (Table 1: 5.9 cm vs 4.8 cm position error; reward 957 vs 1064). Since the main co-design results use oracle RL experts/Mink for validation rather than the model's own controller, the 'self-validation' capability is not actually exercised in the reported optimization pipeline. This should be presented as a promising initial result with appropriate caveats, not as an established replacement for the expert controllers.","section":"§3; Fig. 8; Table 1"}],"minor_comments":[{"comment":"The symbols alpha_size and alpha_mass are used for both the actual reward weight and the guidance-scale weight, with different values; this is confusing and should be renamed or clearly distinguished.","section":"§8.4, §8.5"},{"comment":"Axis labels, units, and correlation coefficients are not stated in the main text. Please add quantitative values, not just the qualitative 'higher/lower' description.","section":"Fig. 7"},{"comment":"Data counts are inconsistent across sections: the main text says 76 trajectories with 56/20 train/validation split, but bimanual experiments use 26 UMI dishwashing trajectories. Clarify which trajectories are used for each design space and how the split is applied to the bimanual space.","section":"§2.4, §3"},{"comment":"The statement that composed diffusion generates designs 'favorable over 20/26 trajs' lacks a definition of favorable. Specify the threshold or metric used to count a trajectory as favorable.","section":"§3, 'Diffusion Composition'"},{"comment":"DGS has higher position error than Zeroth in Table 11 (3.5 vs 2.8 cm). If DGS is claimed to improve each individual sample, this result needs an explanation or a caveat.","section":"Table 11"},{"comment":"The runtime comparison excludes the substantial one-time cost of training 128 RL experts (16 hours each on an A100) and collecting 1.3M quadruped episodes. While reporting optimization time is standard, the paper should explicitly acknowledge this exclusion and its implications for adopting the method on a new design space.","section":"§9.3.2, Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is ambitious and has a strong real-world validation component. However, the abstract overclaims relative to the reported tables, and the 'zero-shot' and 'unified' claims need qualification or additional experiments. The core idea is promising, so I do not recommend rejection, but the revision must address the quadruped performance contradiction, the per-space/per-reward tuning of guidance scales, and the lack of a truly joint multi-space model."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this carefully. The core idea is genuinely new: a single diffusion transformer over RoboTokens that is simultaneously an embodiment generator, a dynamics model, and a cross-embodiment controller, with reward gradients steered through its own predicted states and actions. I haven't seen that combination. The paper also does the right thing by publishing reward definitions, hyperparameters, and the full token schema in the appendix, and the ALOHA hardware result (73% tracking error reduction) is a real validation, even if it is one un-replicated build.\n\nThe model is genuinely good on the fixed-base and bimanual mobile design spaces: it beats CMA-ES in both reward and wall-clock time, and its learned controller correlates well with the oracle (Table 1). Test-time compute scaling holds in all spaces, and the multi-trajectory composition works.\n\nThe soft spot is the quadruped space, and it is load-bearing. The paper's own tables show CMA-ES achieving lower position/orientation error in single-trajectory quadruped runs (1.9 cm vs 2.6 cm for tracking-only; 2.1 cm vs 3.5 cm for the velocity reward), and higher survival. Dynamics Self-Guidance often matches or underperforms the zeroth-order sampler there. Fig. 7 shows the predicted-vs-actual reward correlation is lowest for quadrupeds—precisely where dynamics-guided search is supposed to help. So the abstract's blanket claim of 'improving performance ... over the evolutionary baseline' is not supported in the hardest design space. This does not sink the paper—you can still argue the speedup and scaling make it useful—but it needs a direct sentence in the abstract or results fixing the claim.\n\nThe guidance scales are hand-tuned per design space and reward (50, 100, 0.2; size/mass scales 0.005 vs 0.1 in the actual reward), so 'zero-shot' overstates things. The paper discloses this, which I appreciate, but it means the unseen-reward claim is weaker than it looks. No code or data is released, a real limitation given the scale.\n\nThis is a serious system paper worth a serious referee. Send it to review, but the referee should push on the quadruped comparison, the guidance-scale sensitivity, and the absence of a model-vs-simulator reward-ranking evaluation on the quadruped space. Who should read it: anyone working on co-design, generative robot design, or cross-embodiment control. I'd bring it to our reading group.","headline":"A single DiT over variable-length RoboTokens is a genuinely new way to unify robot design generation, dynamics, and control—but the abstract's blanket claim of beating CMA-ES fails in the quadruped space, and the 'zero-shot' guidance is hand-tuned per reward.","tokens_in":21822,"tokens_out":4979,"would_cite":true,"duration_ms":45941,"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":"A single diffusion transformer, trained on RoboTokens, can generate complete robot designs, predict their dynamics, and control them, optimizing for unseen trajectories and rewards zero-shot.","keywords":["robot co-design","motion-conditioned generation","diffusion transformer","RoboToken","cross-embodiment control","Dynamics Self-Guidance","embodiment optimization","zero-shot reward optimization"],"falsifier":"On a held-out set of optimized designs, compute both the model's predicted rewards and the rewards from the real simulator/controller (e.g., the RL experts or Mink). The central claim fails if the rank correlation between predicted and actual rewards is not significantly positive across all three design spaces, or if designs selected by Dynamics Self-Guidance score no better on average than random designs when evaluated with the oracle controller.","tokens_in":20873,"feed_emoji":"🤖","tokens_out":6549,"duration_ms":60458,"temperature":0.7,"pith_summary":"The paper asks whether the best robot embodiment for a manipulation task can be found by a single learned model rather than by searching over designs with a simulator. It proposes Transformer Transformer, a diffusion transformer that operates on RoboTokens—a unified tokenization of a robot's links, joints, motors, states, and actions—and learns to model their joint distribution. The central claim is that one reward-agnostic dynamics model can, at inference time, be steered by gradients of any user-specified reward to generate high-value robot designs zero-shot, while also serving as a cross-embodiment controller to validate those designs. If true, this collapses co-design into a single GPU-parallelizable diffusion pass, with rewards comparable to or better than the evolutionary baseline CMA-ES while being orders of magnitude faster, plus a 73% tracking-error reduction on a fabricated ALOHA arm.","feed_headline":"One diffusion model designs, simulates, and controls robot hardware","feed_subtitle":"Trained on robot 'tokens,' it zero-shot optimizes designs for unseen tasks, beating evolutionary search in seconds; a fabricated arm cut err","key_machinery":"RoboToken: a tokenization scheme that represents any rigid articulated robot as a variable-length sequence of continuous vectors for links, joints, motors, states, and actions, with IDs encoding connectivity and timesteps, and canonicalized transforms. Transformer Transformer is a diffusion transformer (DiT) trained to denoise these tokens under different masking schemes, so one model can generate embodiments (motion-to-robot) and predict actions (cross-embodiment control). Dynamics Self-Guidance converts the model's reward-agnostic state/action predictions into reward-specific value predictions by backpropagating user-reward gradients through the denoiser at each diffusion step (classifier-","core_discovery":"The paper's central claim is that a single DiT trained on RoboTokens unifies robot embodiment generation, dynamics prediction, and control. Given target end-effector trajectories, the model diffuses complete robot embodiments together with their state and action trajectories; because the predictions are reward-agnostic, user-defined rewards can be evaluated on them and their gradients backpropagated through the denoiser to steer embodiment diffusion—a procedure the authors call Dynamics Self-Guidance. They show this zero-shot optimizer matches or outperforms CMA-ES across fixed-base, quadruped, and bimanual mobile design spaces, and that the same model, when conditioned on an embodiment, act","pith_inferences":["The method's success hinges on how faithfully the model's predicted dynamics reproduce the simulator; Figure 7 already shows noticeably lower predicted-vs-actual reward correlation for the quadruped space, suggesting that off-distribution designs—especially those involving learned whole-body controllers and contacts—are where the approach is most likely to over-optimize to the model.","The 8-timestep context (§2.2, §9.2) is a surprisingly aggressive compression of long-horizon dynamics; it implies the model can reason about an episode from a handful of representative instants, a claim that could be tested directly by ablating context length to see when predicted rewards diverge from simulated ones.","The authors suggest evolutionary algorithms could use the diffusion model as a high-quality sampler at inference time; a natural extension is to interleave Dynamics Self-Guidance with an outer evolutionary loop, which might fix the local-minima issue the paper observes in single-sample guidance."],"forward_implications":["Robot co-design can be done with one GPU-parallelized diffusion model instead of a separate generator, critic, controller, and evolutionary search; reported optimization times drop from minutes/hours to under a second to tens of seconds.","Unseen reward functions and unseen target trajectories can be optimized zero-shot at inference time, because the model is trained reward-agnostically and rewards are injected only as gradients during sampling.","The same model can act as a cross-embodiment whole-body controller, enabling generated designs to be validated directly without retraining per embodiment; on bimanual robots, the learned controller's tracking reward aligns closely with an oracle controller (Fig. 11), and on quadrupeds the correlation is positive (r = 0.53, Fig. 8).","Generated designs inherit diffusion properties: they lie on the training manifold, respect cross-attribute coherence (mass, inertia, motor sizing), and improve with test-time compute, providing a natural test-time scaling curve."],"fun_headline_variants":["Transformer Transformer: robot co-design in seconds","Diffusion co-design: zero-shot robot optimization","One model designs, predicts, and controls robots","Cut robot tracking error by 70% via diffusion co-design","Beats evolution: zero-shot robot design via diffusion"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the model's predicted states and actions for a given embodiment and trajectory are faithful enough that rewards computed on those predictions rank designs the way the real simulator and controller would, and that reward gradients backpropagated through the denoiser point toward designs that also score well when actually simulated.","fun_headline_variants_meta":{"raw":{"variants":["Transformer Transformer: robot co-design in seconds","Diffusion co-design: zero-shot robot optimization","One model designs, predicts, and controls robots","Cut robot tracking error by 70% via diffusion co-design","Beats evolution: zero-shot robot design via diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001875,"raw_usage":{"total_tokens":7189,"prompt_tokens":737,"completion_tokens":6452,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":6378}},"tokens_in":481,"tokens_out":6452,"duration_ms":42231,"temperature":1.0,"reasoning_tokens":6378,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T01:26:22.850594+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out set of optimized designs, compute both the model's predicted rewards and the rewards from the real simulator/controller (e.g., the RL experts or Mink). The central claim fails if the rank correlation between predicted and actual rewards is not significantly positive across all three design spaces, or if designs selected by Dynamics Self-Guidance score no better on average than random designs when evaluated with the oracle controller.","supporting_citations":[],"review_version":1}