{"id":"2c5e9ce4-eb2f-46c1-b493-b20c7bb8130f","arxiv_id":"2607.09648","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":6,"one_line_summary":"Predicting continuous B-spline action curves instead of discrete chunks lets visuomotor policies finish real and simulated manipulation tasks substantially faster without large success-rate loss.","lead":"Robot policies usually spit out short lists of discrete moves; this paper instead has them output smooth B-spline curves that can be sped up at run time. That change cuts task time on real arms while keeping success rates competitive with standard chunking policies.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection beyond the reader's already-identified controller-trackability limit; that limit is correctly scoped by the paper.","rationale":"The strongest claim is empirical and carefully hedged: BSP reduces completion time while preserving success in most 1\times–2\times settings across three real tasks and several sims, with honest reporting of 4\times collapse. The single condition that must hold is that the low-level controller can track the temporally rescaled spline samples. That condition is already identified by the reader, is demonstrated to fail by the authors themselves, and is discussed in Limitations. No additional load-bearing flaw (e.g., invalid knot construction, mis-specified alignment, or circular evaluation) is present in the manuscript. Therefore the CONDITIONAL verdict—accept-shaped contribution once claims are scoped to controller-feasible regimes and reporting is strengthened—needs no adjustment.","tokens_in":13246,"tokens_out":515,"duration_ms":6509,"concrete_test":"On the same ARX5 Speed Stacking layouts, re-run Reg.+BSP at 3\times and 4\times while logging joint tracking error (commanded vs. measured) and success; if success collapses exactly when peak tracking error exceeds the demonstration envelope (or a fixed threshold such as 5°), the controller-trackability bound is confirmed and the speedup claim remains correctly scoped rather than overstated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption is the right one and is already load-bearing for the central speedup claim: temporal rescaling a(u)\to a(nt) (Sec. 4) only yields reliable task completion when the low-level joint controller can track the denser, faster samples without leaving the policy's training distribution. The paper itself demonstrates the failure mode (Speed Stacking Reg.+BSP 4\times: 0/20; Sec. 5.2 Finding 2 and Sec. 6) and attributes it to ARX5 stiffness/accuracy limits rather than to the B-spline representation. No deeper internal inconsistency appears: adaptive fitting (Alg. 1), fixed-size [U;C] prediction, and segment alignment (Eq. 2) are coherent, and the 1\times–2\times real-world and Push-T results support the claim inside the feasible regime. The remaining gaps (no error bars, single hardware platform, limited statistical reporting) are presentation/strength issues, not hidden assumptions that would overturn the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper introduces B-spline Policy (BSP), an action representation that replaces discrete-time action chunks with continuous cubic B-spline curves parameterized by knots and control points. Demonstrations are converted via adaptive knot insertion (Alg. 1); a standard visuomotor policy (Diffusion or ACT-style regression) then predicts fixed-size local B-spline parameters [U;C]. At inference, segments are temporally rescaled (a_exec(t)=a(nt)), sampled at high frequency, and stitched with an MSE-based segment-alignment step (Eq. 2). Real-world evaluation on Cube Picking, Table Cleaning, and Speed Stacking (ARX5 arms; 20 fixed layouts per method) and simulation on Push-T, RoboMimic, and RoboCasa show reduced average completion times versus discrete chunk baselines and DemoSpeedup, with success rates matched or improved in most 1×–2× settings; aggressive 4× speedup can collapse success when low-level tracking fails.","tokens_in":13558,"tokens_out":1502,"duration_ms":26912,"significance":"Execution speed is a genuine bottleneck for learned visuomotor policies, and the paper cleanly identifies uniform temporal resolution and chunk-boundary discontinuities as algorithmic contributors. Representing actions as B-splines is a natural, compact continuous alternative that integrates with existing Diffusion/ACT pipelines with minimal architectural change, and the inference-time segment alignment mechanism is a practical, falsifiable contribution. The multi-backbone, multi-task real-world results (Table 1), the alignment ablation (Fig. 6), and the Push-T speedup study (Table 2b) provide concrete evidence that continuous spline actions can cut completion time while preserving success inside a controller-feasible regime. Limitations of hardware tracking are stated explicitly (Sec. 6). This is a solid systems/engineering contribution for robot learning venues.","major_comments":[{"comment":"Table 1 and Sec. 5.2: Success is reported as raw fractions over 20 fixed layouts with no binomial confidence intervals, standard errors, or hypothesis tests. Several central “preserves success” claims rest on small absolute differences (e.g., Table Cleaning Reg. 4×: 14/20 vs 13/20; Cube Picking many 19–20/20 ties). With n=20 these differences are underpowered; the paper should report CIs (or bootstrap intervals) and, where it claims improvement or parity, a simple test so readers can judge whether the success–time trade-off is statistically supported.","section":"Table 1 / Sec. 5.2"},{"comment":"Sec. 4 (temporal rescaling a_exec(t)=a(nt)), Sec. 5.2 Finding 2, and Sec. 6: The central speedup claim is conditioned on low-level joint controllers tracking denser, faster spline samples without leaving the policy’s training distribution. The paper correctly shows the failure mode (Speed Stacking Reg.+BSP 4×: 0/20) and attributes it to ARX5 stiffness/accuracy, but does not quantify tracking error, lag, or a per-task maximum reliable speedup. A short characterization (e.g., measured joint tracking RMSE vs. n, or the largest n that keeps success within a stated band of the 1× baseline) would make the feasible regime precise rather than anecdotal.","section":"Sec. 4 / Sec. 5.2 Finding 2 / Sec. 6"},{"comment":"Sec. 5.1 baselines and Appendix A.1: Naive n× chunk execution and DemoSpeedup are the right comparisons, but DemoSpeedup’s phase downsampling rates are hand-chosen per task (precision phase fixed at 1; non-precision 4 or 2) after observing that ≥2 on precision “prevents reliable completion.” This makes the DemoSpeedup column a best-effort rather than a fixed protocol. Either freeze the hyperparameters a priori or report a small sensitivity sweep so the comparison cannot be read as tuned against BSP.","section":"Sec. 5.1 / Appendix A.1 DemoSpeedup"}],"minor_comments":[{"comment":"Fig. 1 and Fig. 3: The “3.76X faster” callout and the 10 Hz → 100 Hz pipeline are clear, but the figures do not state the corresponding baseline wall-clock time or which task/backbone the 3.76× refers to; add a caption note.","section":"Fig. 1 / Fig. 3"},{"comment":"Eq. (1) and Sec. 3: Degree p is fixed to cubic in experiments; state explicitly whether open/clamped knot vectors are used and how endpoint multiplicity is handled when appending repeated final control points for fixed-size targets.","section":"Eq. (1) / Sec. 3–4"},{"comment":"Appendix Table 3 (fitting ε ablation) is only on Push-T. A one-sentence note on whether real-world tasks used the same ε=0.002 and whether success was sensitive to it would strengthen reproducibility.","section":"Appendix A.2 / Alg. 1"},{"comment":"Related work (Sec. 2): BEAST and DMPs are cited; a short explicit contrast—BSP as a control-oriented continuous action head + execution pipeline versus tokenizer or motion-primitive encoding—would help readers place the contribution.","section":"Sec. 2"},{"comment":"Minor wording: “stablizes” → “stabilizes” (Sec. 3); “segmentment alignment” → “segment alignment” (Alg. 2 caption); arXiv id in the header is 2607.09648 while the abstract footer shows the same—ensure consistency with the public listing.","section":"Sec. 3 / Alg. 2"},{"comment":"Simulation: segment alignment is disabled for delta actions “because it does not have a clear physical interpretation.” A brief alternative (e.g., velocity-consistent stitching) or an explicit statement that sim speedups therefore isolate only the continuous representation would avoid confusion when comparing Table 2b to real-world results.","section":"Appendix A.1 Controller"}],"recommendation":"minor_revision","confidential_remarks":"The controller-trackability limit is already scoped honestly by the authors; I do not treat it as a hidden flaw. The main revision burden is statistical reporting and a clearer feasible-speedup envelope, not a redesign of the method. Novelty relative to BEAST/DMPs is incremental but the end-to-end acceleration pipeline and real-robot evidence are sufficient for a robotics venue. Fit for a serious robotics/ML journal is good if the statistical and regime-characterization points are addressed."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: replace discrete action chunks with fixed-size local B-spline parameters (knots + control points), train the usual Diffusion or ACT-style head on adaptively fitted demo targets, then temporally rescale and align segments at inference. On their ARX5 tasks that shortens wall-clock time a lot while success mostly holds at 1–2×.\n\nWhat is actually new is not B-splines themselves (classical, and BEAST/DMPs already sit nearby). It is the full control-oriented pipeline: FITPACK-style adaptive fitting to fixed-size [U;C] targets, plug-in prediction, high-rate sampling after a_exec(t)=a(nt), and the MSE segment-alignment step that keeps consecutive predictions continuous under latency. That combination is cleanly specified (Alg. 1, Eq. 2, Alg. 2) and works as a drop-in for two standard backbones.\n\nThey do the empirical work properly for this venue. Fixed test layouts, 20 real rollouts, naive n× and DemoSpeedup baselines, both Diffusion and regression, plus a real alignment ablation that shows continuity matters at speed. Sim tables (Push-T, RoboMimic, RoboCasa) show parity or gains; Push-T speedup even improves score because the agent gets more corrective pushes inside the time budget. The 4× Speed Stacking collapse to 0/20 is reported honestly and attributed to ARX5 tracking limits, not hidden.\n\nSoft spots are real but proportional. No error bars or CIs on the 20-trial fractions; single low-cost arm platform; free parameters (ε, 16-knot segments, λ window) are chosen without much sensitivity beyond a small Push-T table. The load-bearing premise is exactly the one the reader flags: rescaling only helps while the low-level controller stays inside the policy’s distribution. The paper already scopes that in Findings 2 and Limitations, so it is not a hidden flaw—just the boundary of the claim.\n\nThis is for people shipping visuomotor imitation who care about wall-clock time, not for theory. Math is standard cubic B-splines; citations are fair; no circular derivation. I would bring it to reading group, cite the representation when we next need faster chunked policies, and send it to peer review. Strengthen stats and code release, keep the speed claims inside controller-feasible regimes, and it is a clear accept-shaped methods paper.","headline":"Solid CoRL-style engineering: continuous B-spline action heads plus segment alignment actually cut real completion time on Diffusion/ACT pipelines, with the controller-trackability limit already scoped by the authors.","tokens_in":14173,"tokens_out":622,"would_cite":true,"duration_ms":6447,"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":"Robot policies that output continuous B-spline curves finish manipulation tasks much faster than discrete action-chunk policies while keeping strong success rates.","keywords":["B-spline policy","visuomotor policy","action representation","fast manipulation","temporal rescaling","segment alignment","imitation learning","robot manipulation"],"falsifier":"On the same real arms and Speed Stacking setup, measure success at 4× temporal scaling with a higher-stiffness controller and perfect segment alignment; if success remains near zero even when tracking error is demonstrably tiny, the continuous-representation claim itself fails rather than the hardware limit.","tokens_in":14107,"feed_emoji":"🤖","tokens_out":829,"duration_ms":21474,"temperature":0.7,"pith_summary":"Most learned robot policies predict fixed-length lists of discrete future actions. That choice stabilizes long-horizon learning but forces uniform timing and creates jerky joins between successive predictions, so even accurate policies stay slow. This paper argues that if the policy instead predicts the parameters of a continuous B-spline—knots and control points—the identical behavior becomes a smooth curve that can be sampled at high frequency and sped up simply by scaling time. An inference-time alignment step keeps successive curve segments continuous so acceleration does not accumulate boundary jumps. On real cube picking, long-horizon table cleaning, and bimanual cup stacking, and on standard simulation suites, the resulting B-spline Policy cuts completion time substantially while matching or improving success in most moderate-speed settings.","feed_headline":"Spline actions finish robot tasks twice as fast","feed_subtitle":"Continuous B-spline curves replace discrete chunks so the same policy can be sped up at runtime.","key_machinery":"B-spline Policy (BSP): the policy emits a fixed-size vector of local knots and control points that define a continuous action curve; adaptive fitting turns discrete demonstrations into these targets, and inference-time segment alignment chooses the point on each new segment that best matches the last executed action so consecutive segments join smoothly under acceleration.","core_discovery":"Parameterizing visuomotor actions as continuous cubic B-splines, predicted directly by ordinary imitation-learning backbones and executed with temporal rescaling plus segment alignment, substantially reduces task completion time relative to discrete action-chunk baselines while preserving or improving success rates on real and simulated manipulation tasks.","pith_inferences":["Stiffer, higher-bandwidth joint controllers would likely recover the 4× regime that currently collapses on contact-rich bimanual work.","Adaptive knot density already marks which phases need precision; that signal could drive automatic recovery or curriculum decisions.","The same knot-and-control-point head could replace discrete action tokens inside vision-language-action models that today emit fixed-length chunks.","Because local spline support already encodes multi-step smoothness, policies may need shorter prediction horizons than discrete chunk methods require."],"forward_implications":["A trained policy can be sped up at runtime by temporal rescaling without collecting new data or retraining for each target speed.","Boundary discontinuities between discrete action chunks cease to be the main obstacle to high-speed visuomotor control.","The same continuous head drops into both diffusion and regression (ACT-style) backbones with only an output-space change.","Long-horizon tasks gain the largest absolute time savings because smooth high-frequency commands can be sustained across many successive segments.","Success holds only while the low-level controller remains able to track the rescaled trajectory."],"fun_headline_variants":["B-spline policies finish robot tasks twice as fast","Continuous B-splines cut robot manipulation times","Predict B-spline curves to speed up robot policies","B-spline actions complete tasks far faster than chunks","Temporal scaling of B-splines halves robot task times"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The robot’s low-level joint controllers must still track the denser, faster samples of the predicted spline without drifting into states the policy never saw.","fun_headline_variants_meta":{"raw":{"variants":["B-spline policies finish robot tasks twice as fast","Continuous B-splines cut robot manipulation times","Predict B-spline curves to speed up robot policies","B-spline actions complete tasks far faster than chunks","Temporal scaling of B-splines halves robot task times"]},"model":"grok-4.5","effort":"low","cost_usd":0.007,"raw_usage":{"total_tokens":1632,"prompt_tokens":650,"num_sources_used":0,"completion_tokens":78,"cost_in_usd_ticks":70000000,"prompt_tokens_details":{"text_tokens":650,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":904,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":650,"tokens_out":78,"duration_ms":7579,"temperature":1.0,"reasoning_tokens":904,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T01:28:09.564141+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the same real arms and Speed Stacking setup, measure success at 4× temporal scaling with a higher-stiffness controller and perfect segment alignment; if success remains near zero even when tracking error is demonstrably tiny, the continuous-representation claim itself fails rather than the hardware limit.","supporting_citations":[],"review_version":1}