{"id":"f40299c9-1cec-4162-bde1-cf89c44ae368","arxiv_id":"2607.29482","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Starting generative action generation from a robot's recent state history instead of Gaussian noise reduces transport cost and enables fast, low-latency control without losing success rate.","lead":"This paper introduces a robot-control method that starts its AI-generated motion from the robot's recent past positions instead of random noise, making generation faster while keeping task success rates similar. It could help robots act quickly and smoothly from human demonstrations in real time.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline latency and low-NFE claims are confounded by a 15x model-size difference; the paper's own scaled-down CFM note is not backed by reported latency or success curves.","rationale":"The reader's verdict is already CONDITIONAL, and I agree with that verdict, but my primary load-bearing concern differs from the reader's stated weakest_assumption. The reader focused on the static transport-cost correlation assumption and the state/action representation mismatch. Those are real issues, but the most decisive threat to the central claim is the uncontrolled comparison behind the headline latency and low-NFE numbers. If a 17M noise-initialized CFM achieves the same latency and success, then history initialization is not what enables high-frequency control; the contribution reduces to a minor geometric observation. The paper itself introduces the same-size CFM check but only on one task and without latency data, which is exactly the missing control needed. This is not a fatal flaw in the method's correctness, but it is a load-bearing gap in the evidence: the abstract's quantitative promises (19.1 ms, matching success rates) would not be uniquely supported. I therefore recommend keeping the verdict at CONDITIONAL, with the condition being a same-size baseline evaluation. I mark agreement as 'partial' because the reader's rationale contains this concern but the reader's weakest_assumption field points elsewhere.","tokens_in":11784,"tokens_out":11542,"duration_ms":143614,"concrete_test":"Train a 17M-parameter CFM baseline with an independent Gaussian source, using the same U-Net/ResNet-18/FiLM backbone, conditioning, and hyperparameters as Temporal Policy, on at least Square ph, Tool Hang ph, and Transport mh. Measure success rate versus NFE (1, 2, 4, 8, 10) and wall-clock inference latency on the same RTX 4080. If the 17M CFM matches Temporal Policy's success-latency curve (e.g., >=0.9 success at NFE<=4 and <25 ms), the efficiency gain is not attributable to history initialization. If Temporal Policy dominates at low NFE with comparable parameter count, the central efficiency claim is supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The causal claim 'history initialization -> straight field -> low NFE -> 19.1 ms latency' is not established because the efficiency comparison is not architecture-controlled. Table I reports 19.1 ms for Temporal Policy (17M params) versus 63.5 ms for CFM and 615.6 ms for Diffusion Policy, both with 255M params. The 15x parameter reduction alone could explain most of the latency difference. The paper states (Sec. IV-A): 'We trained a scaled-down CFM baseline restricted to the same 17M parameter architecture on the transport task and did not observe degraded performance' — but it does not report that baseline's success across tasks, its NFE–success curve, or its inference latency. Thus the 'low NFE enabled by geometry' claim rests on comparing different architectures at different scales. Table II further weakens the straight-field mechanism: straightness ratio is 1.08 for CFM vs 1.02 for Temporal Policy, so the vector field is not substantially straighter than an already-straight flow-matching baseline. The remaining differentiator is the shorter source-to-target distance, which is close to tautological when the source is chosen as recent history and the target as future actions. Without a same-size, same-backbone noise-initialized baseline evaluated across tasks at NFE 1–10, the headline efficiency advantage is a confounded assertion rather than a demonstrated property of history-initialized coupling.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Temporal Policy proposes a generative policy for robot learning from demonstration based on stochastic interpolants. Instead of initializing the generative flow with independent Gaussian noise, the method initializes it with the robot's recent state history, coupling past states to future action chunks. The paper derives the interpolant, drift objective, and score-recovery equations, presents training and inference algorithms, and evaluates the method on five Robomimic tasks and a real Barrett WAM teleoperation platform. It reports competitive success rates, reduced transport cost, low NFE, and a 19.1 ms inference latency on an RTX 4080, alongside data-efficiency and real-world results.","tokens_in":12088,"tokens_out":7019,"duration_ms":85169,"significance":"If the central claims hold, the paper addresses a real bottleneck in deploying generative policies for closed-loop control: inference latency. The data-dependent coupling idea is well motivated, and the SDE/ODE derivation in Section III-C is correct. The paper also provides public code and reports success-rate parity with strong baselines across multiple tasks, which is valuable. However, the headline efficiency claim is not architecture-controlled, and the default deterministic inference mode has a fundamental limitation for a point-mass source that is not discussed in the proper framework. The transport-cost claim is supported by only one task and relies on an unmeasured correlation assumption. These issues are load-bearing for the paper's main contributions, so the manuscript requires substantive revision rather than minor cleanup.","major_comments":[{"comment":"The efficiency comparison is not architecture-controlled. Temporal Policy uses 17M parameters while Diffusion Policy and CFM use 255M, so the 19.1 ms latency improvement could be largely due to model size. The paper states that a scaled-down 17M CFM baseline was trained on Transport and did not degrade, but it does not report that baseline's success rates, NFE-success curve, or inference latency. To support the claim that history initialization enables low NFE and fast inference, the authors should provide a same-backbone, same-size noise-initialized CFM baseline evaluated across tasks at NFE 1-10, with latency numbers.","section":"IV-A, Table I"},{"comment":"The recommended deterministic sampling mode (g=0) cannot generate a distribution over actions from a fixed history. A deterministic ODE maps a point to a point; starting from x0 yields a single output trajectory, not a sample from a multimodal action distribution. This is especially important because the paper motivates generative models by their ability to capture multimodal LfD behaviors. The authors should either evaluate and report the stochastic SDE mode at matched NFE, or explicitly reframe the default policy as a deterministic conditional-mean-style policy and provide evidence that this preserves the useful multimodality (e.g., via success metrics or diversity measures). The current text does not resolve this issue.","section":"III-E, Algorithm 2, Eq. (6)"},{"comment":"The transport-cost reduction is claimed as a general property, but the static argument assumes E[x1^T x0] > 0 and the dynamic transport cost is measured only on Square ph (Table II). Expert demonstrations with pauses, reversals, or jitter can make this correlation small or negative, so the geometric benefit is task-dependent. The authors should report the empirical cross-correlation between source and target for each task, and ideally dynamic transport costs across all five tasks. Without this, the 'nearly an order of magnitude' reduction is not established beyond a single task.","section":"III-B, IV-B"},{"comment":"The real-world experiment appears to violate the structural equivalence S ~= A required in Section III-A. The state-action coupling is described as using follower joint positions and gripper velocities, but the formulation assumes both source and target live in the same configuration space (joint positions or Cartesian poses). Please clarify the exact representation of x0 and x1 in the real-world experiment, and explain how gripper velocities fit the S ~= A requirement, or relax the formalism to accommodate them. This matters because the real-world deployment is used as evidence that the method works outside simulation.","section":"IV-E"}],"minor_comments":[{"comment":"The symbol ε is used both for the training noise scale in Eq. (1) and for the inference noise schedule in Eq. (6). In the ablation, ε is apparently used as gλ without a clear definition. Please distinguish these two quantities.","section":"IV-C, Fig. 3"},{"comment":"The table reports Max/Average success rates but no variance or confidence intervals, despite stating that values are averaged across 3 seeds. Reporting standard errors would help assess the significance of the small differences between methods.","section":"Table I"},{"comment":"The data-efficiency plot appears to be for Square ph only and no error bars are shown. Clarify this in the caption and text, and consider adding error bars or additional tasks.","section":"Fig. 4"},{"comment":"The score formula divides by ε^2 λ(1-λ), so the expression is singular at λ=0 and λ=1 when ε=0. The text should state how the deterministic ε=0 case is handled numerically, especially for the first Euler step.","section":"Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The paper addresses a timely problem and the central SDE/ODE framework is sound, but the current evidence does not cleanly separate the benefit of history initialization from the benefit of a smaller network. The deterministic-inference issue is also conceptually important: if the default mode is deterministic, the paper should not claim to sample multimodal action distributions without reporting stochastic sampling results. These are fixable with additional controlled experiments and clearer framing, so I do not recommend rejection, but the revision needs to be substantive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: the idea is genuinely new — using a point-to-distribution stochastic interpolant initialized at the robot's recent state history as an action chunk generator — and the math is done carefully. The SDE/ODE derivation and the score recovery formula check out. The code is public, and the Robomimic results are competitive. Credit where due: the paper is explicitly honest that the reduced transport cost comes mostly from starting closer to the target, not from straightening the flow (their straightness ratio 1.02 vs CFM's 1.08 says the same). That kind of candor is rare.\n\nThe soft spots are real, though. The headline 19.1 ms number is for a 17M-parameter model versus 255M baselines. The note about a scaled-down CFM baseline is one task and reports no success curve, no NFE sweep, no latency. So the central causal story — history initialization -> low NFE -> low latency — is not actually demonstrated. Without a same-size, same-backbone noise-initialized CFM at NFE 1–10, the efficiency advantage could just be the network size.\n\nSecond, the recommended fast mode is deterministic ODE, which maps a point-mass history to a single action chunk. That is not sampling from a distribution. The stochastic mode does add diversity, but the paper shows it needs more steps and doesn't improve success, so in practice the deployed policy is deterministic. That's fine for some tasks, but it undercuts the 'generative' selling point, and the paper's own limitation paragraph floats the question without answering it.\n\nThird, the real-world experiment uses gripper velocities in the action while the method requires state and action to share a representation. That looks like a direct contradiction of the stated structural assumption. If it's a typo or if they concatenate positions and velocities, they need to say so. Also, the static transport cost argument leans on a strong positive correlation E[x1^T x0] that is never measured; plausible, but unsupported.\n\nNone of this is fatal. The idea is sensible, the derivations are sound, and the simulation results are competitive. The paper just needs a properly controlled efficiency comparison and a cleaner statement of the representation and the generative claim.\n\nWho this is for: robotics researchers working on flow-based or diffusion policies and anyone who cares about inference latency on modest hardware. It deserves a real referee. I'd send it to review with a request for a same-size baseline and the representation clarification.","headline":"A promising but not yet proven application of history-initialized stochastic interpolants to robotic LfD: correct math, competitive results, but the headline latency advantage is confounded by a 15x smaller network.","tokens_in":12584,"tokens_out":5427,"would_cite":true,"duration_ms":58546,"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":"This paper claims that initializing a generative flow at the robot's recent state history, instead of Gaussian noise, makes action generation faster and cheaper while preserving success rates.","keywords":["robotic learning from demonstration","stochastic interpolants","flow matching","data-dependent coupling","action chunking","transport cost","inference latency","generative policy"],"falsifier":"Take any demonstration dataset, randomly shuffle the pairing of history windows and future action chunks to break temporal correlation while keeping marginals, retrain the same Temporal Policy, and measure transport cost and the number of function evaluations needed for a given success rate. If the low-NFE advantage persists despite zero or negative cross-correlation, then the temporal coupling is not the cause. Alternatively, compute E[x1^T x0] directly on a jittery dataset and show the predicted cost reduction J_temporal < J_indep no longer holds.","tokens_in":11631,"feed_emoji":"🤖","tokens_out":6211,"duration_ms":62561,"temperature":0.7,"pith_summary":"Temporal Policy reframes robot action generation as a transport problem: instead of mapping Gaussian noise to future actions, the generative flow starts from the robot's recent state history and carries it to the future action chunk. Because past and future states in smooth demonstrations are correlated, the expected transport distance shrinks and the learned vector field becomes nearly straight, so the policy needs very few sampling steps. The paper reports 19.1 ms inference latency on a consumer GPU with ten function evaluations, overall matching the success rates of noise-initialized diffusion and flow-matching baselines across simulation benchmarks and a physical 7-DoF teleoperation task. If this holds, generative policies become practical for high-frequency closed-loop control without specialized hardware. The same trained model also supports both deterministic and stochastic sampling at inference.","feed_headline":"Robot policies start from past states, cut inference to 19 ms","feed_subtitle":"Coupling past states to future actions straightens the generative flow, so fewer steps match baseline success rates.","key_machinery":"The central object is the history-initialized stochastic interpolant: a continuous probability path x_lambda = (1-lambda)x0 + lambda x1 + epsilon(1-lambda)w_lambda that mixes source history, target action chunk, and Wiener noise, plus a drift network trained by square-loss regression to approximate the target velocity. The source choice does the load-bearing work: because the expected cross-correlation E[x1^T x0] is positive for smooth demonstrations, the expected squared transport cost J = 2HD - 2E[x1^T x0] is strictly smaller than the Gaussian baseline's 2HD. Straightness follows from the linear interpolant combined with the short source-target distance, and analytic score recovery lets th","core_discovery":"On its own terms, the paper's discovery is that a data-dependent coupling — using the robot's recent state history as the source and the future action chunk as the target — turns action generation into a short, nearly straight point-to-distribution transport instead of a long curved journey from Gaussian noise. With smooth demonstrations the history and the near-future actions are positively correlated, so the expected squared distance the flow must cover is strictly smaller than in the independent-Gaussian setup. Training a stochastic-interpolant drift with this coupling produces straight vector fields, which is why the policy can generate good actions in one to ten solver steps. On a preci","pith_inferences":["Editorial inference: Because the dynamic transport-cost reduction is demonstrated on only one task, the broad claim that low-NFE success follows from the coupling is plausible but not yet fully tested; measuring transport cost on the multi-human, high-variance datasets would directly probe the correlation assumption.","Editorial inference: The requirement that states and actions share a representation limits the method to tasks where actions are positions or poses; extending to force/torque or velocity commands would need a different source construction, and the paper's own real-world experiment used gripper velocities, a representation mismatch.","Editorial inference: If the benefit is mostly 'starting closer,' then initializing from the previous action chunk or a velocity-projected state, as the authors suggest for future work, may yield similar gains; a head-to-head comparison would isolate the source-coupling effect from the interpolant's straightness.","Editorial inference: The authors themselves question whether history initialization preserves full multimodality rather than collapsing it; a behavioral-diversity metric on multi-human datasets would settle whether the method retains the range of valid demonstration strategies."],"forward_implications":["Generative visuomotor policies can operate at roughly 50 Hz on a single consumer GPU (19.1 ms inference), making them viable for closed-loop control rather than open-loop re-planning.","A single trained model supports both fast deterministic (ODE) sampling for latency-critical control and stochastic (SDE) sampling for diversity, with the noise schedule adjustable at inference time.","The approach is architecture-agnostic, so it can combine with other generative acceleration techniques such as consistency distillation or higher-order ODE solvers.","Fewer demonstrations are needed to reach a given success rate, because the source-target gap is smaller and the learning problem is simpler.","History initialization alone yields usable actions at a single function evaluation, suggesting the learned vector field is nearly straight and well behaved."],"fun_headline_variants":["History initialization straightens robot action flow","Past states seed future actions for 19 ms inference","From history to actions: 10x lower transport cost","Temporal coupling speeds robot policy by 10x","Robot policy uses history to cut inference to 19 ms"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire efficiency gain rests on the assumption that a robot's recent state history and its immediate future action sequence are strongly positively correlated, i.e., that demonstrations are smooth, purposeful motions; with pauses, reversals, or high-frequency jitter the cross-correlation E[x1^T x0] shrinks and the transport-cost and straightness advantages erode.","fun_headline_variants_meta":{"raw":{"variants":["History initialization straightens robot action flow","Past states seed future actions for 19 ms inference","From history to actions: 10x lower transport cost","Temporal coupling speeds robot policy by 10x","Robot policy uses history to cut inference to 19 ms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1192,"prompt_tokens":745,"completion_tokens":447,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":372}},"tokens_in":489,"tokens_out":447,"duration_ms":5955,"temperature":1.0,"reasoning_tokens":372,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T06:10:46.250648+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any demonstration dataset, randomly shuffle the pairing of history windows and future action chunks to break temporal correlation while keeping marginals, retrain the same Temporal Policy, and measure transport cost and the number of function evaluations needed for a given success rate. If the low-NFE advantage persists despite zero or negative cross-correlation, then the temporal coupling is not the cause. Alternatively, compute E[x1^T x0] directly on a jittery dataset and show the predicted cost reduction J_temporal < J_indep no longer holds.","supporting_citations":[],"review_version":1}