{"id":"6f42d2ee-3a38-46c3-a915-8a0b16025644","arxiv_id":"2610.05734","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A training-free post-processor that resamples robot action chunks and scales step sizes by directional consistency lets frozen policies run at user-requested speeds with little or no loss in task success.","lead":"Robots trained by imitation usually play back the speed of their demonstrations, so slowing down or speeding up during a task can break their performance. This paper presents a small add-on that resamples and rescales a robot's predicted actions at test time so it can change speed on command while keeping task success.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The load-bearing mechanism—that cosine inconsistency gates rescaling precisely because it tracks task-phase criticality—is validated on one OOD task and five demonstration traces; before the cross-task transferability claim is trusted, the phase-alignment and observation-distance evidence must be…","rationale":"The reader's weakest_assumption is that the CI proxy and gating schedule generalize across tasks and policy families, with the OOD mechanism validated on a single task. I agree this is the load-bearing concern, and I find it is genuinely the fragile link: the empirical success-preservation claim is broad (50 sim + 4 real tasks, three policy families), with baseline gaps much larger than 50-rollout noise, but the mechanism—the actual novelty—is supported by five demonstration traces and one OOD task. I added two sharpenings: (a) the criticality claim is made on demonstration trajectories while the gating acts on predicted chunks, a gap only bridged on Task 60; and (b) the 'in-distribution' advantage in Figure 6 is confounded by FreeSpeed's systematically gentler rescaling (fs=0.9rho, fc=1), which is consistent with a weaker 'deviate less everywhere' explanation. The authors' own evidence partially mitigates this: the constant-factor sweep in Appendix E.2 suggests no single constant factor matches FreeSpeed, and the lambda/fs sensitivity plateau in Figures 7b-7c shows the schedule is robust. No ad hominem, no internal-inconsistency claim: the paper is honest about its acceleration ceiling (Appendix F) and its single-task analysis. Because the concern is about the width of the mechanism claim rather than the existence of the effect, CONDITIONAL remains the right verdict; my read does not move it in either direction. The concrete test is a direct replication of Section 5.5's measurement on three tasks spanning the paper's own task-type taxonomy, which would settle transferability in one pass.","tokens_in":30311,"tokens_out":6362,"duration_ms":40979,"concrete_test":"Using the same frozen Flow Matching checkpoints, reproduce Section 5.5's kNN observation-distance analysis (Figures 6b/6c and A5) plus Figure 4a-style phase alignment on three additional LIBERO-90 tasks spanning the paper's own placement groups: task 2 (structure-constrained), task 48 (bounded receptacle), and task 69 (open relational), at the 2x and 0.3x commands with 50 rollouts each. A meaningful pass requires (a) predicted-chunk Icos (not demonstration Icos) to peak at grasp/release for all three tasks, and (b) FreeSpeed's median observation-distance ratio to the 1x reference to stay near the Task 60 value (~1.09-1.13) while w/o CR stays above ~1.6. If either condition fails on any task, the CI-proxy mechanism does not transfer and the claim should be narrowed to 'conservative rescaling preserves success' rather than 'CI-gated rescaling tracks criticality'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two layers: (1) empirically, FreeSpeed preserves or exceeds the 1x reference success across 50 simulated and 4 real tasks; (2) mechanistically, it does so because Icos = max_i(1 - cos angle between adjacent delta-p actions) is a policy-independent proxy for phase criticality, so that Equation (5)'s CI-gated rescaling keeps observations in-distribution. Layer (1) is well supported; layer (2), which is the stated novel contribution (§1, §4.1), rests on narrow evidence. Figure 4a shows CI peaks during grasp/placement for only five LIBERO-90 demonstration trajectories, and the observation-distribution mechanism in Section 5.5 (Figures 6, A5, A6) is measured on exactly one task, Task 60. Moreover, Figure 4a measures CI on human demonstrations, whereas the gating operates on predicted chunks; the claim that 'the policy learns these features and reproduces them' (§5.2) is only checked for Task 60. A further confound: at 2x, w/o CR scales every retained action by 2.0 while FreeSpeed caps at fs=0.9rho=1.8 and drops most chunks to fc=1, so FreeSpeed systematically deviates less per replan (realized 1.34x vs 1.55x for vanilla on LIBERO). Figure 6's in-distribution advantage is therefore consistent with the weaker statement that FreeSpeed simply executes less aggressively, not that CI specifically identifies which phases tolerate rescaling. The constant-factor sweep (Appendix E.2) partially answers this, but on three tasks only. None of this refutes the empirical success claim, and the authors honestly state the acceleration ceiling limitation (Appendix F), but the mechanism that distinguishes FreeSpeed from 'rescale conservatively everywhere' is not yet shown to transfer to settings where CI does not peak at critical phases (e.g., tasks with turning free-space motion or oscillatory approaches). If on some task the gating misaligns with criticality, FreeSpeed's advantage over w/o CR could vanish while the frozen-policy claim would still hold it harmless.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper proposes FreeSpeed, a training-free module that controls the execution speed of frozen action-chunking robot policies by post-processing predicted action chunks. At each replan, the chunk is temporally resampled at a commanded stride rho_t, and the resampled translation/rotation increments are rescaled by a factor f_t = f_c + (f_s - f_c) exp(-lambda * I_cos) (Eq. 5), where I_cos is the maximum cosine inconsistency between adjacent translation increments (Eqs. 2-3) and f_s, f_c are per-command scaling endpoints. The paper claims (i) that I_cos is a training-free proxy for task-phase criticality, high during grasp and release and low during transport, so that gating rescaling on it keeps the frozen policy's observations close to the training distribution, and (ii) that, across three policy families (Flow Matching, pi0.5, Fast-WAM), 50 LIBERO tasks and four real-robot tasks at commands from 0.2x to 4x, FreeSpeed preserves or exceeds the 1x reference success while realizing rates from 0.22x to 2.53x, outperforming vanilla temporal resampling and constant-factor rescaling (the w/o CR ablation).","tokens_in":30898,"tokens_out":9510,"duration_ms":53166,"significance":"The empirical layer of the claim is well supported and the evaluation is unusually well controlled: all methods run on identical frozen checkpoints and initial states, the w/o CR ablation isolates the CI-based rescaling, the Appendix E.2 constant-factor sweep is a genuine control for the 'simply less aggressive' confound, and the sensitivity analysis in Figure 7 explores lambda and f_s. The reproducibility appendix (architectures, controller bounds, hardware, full per-task tables) is complete to an uncommon degree. If the result holds, FreeSpeed is a practical plug-and-play speed-control mechanism with no retraining, user-set and in-episode rates, and both acceleration and deceleration, a capability combination not jointly demonstrated by the compared methods. The risk sits in the mechanistic layer: the CI criticality proxy is validated on five demonstration traces and one out-of-distribution task, and the transferability of the hand-set schedule constants across tasks and policy families is asserted more strongly than the current evidence warrants.","major_comments":[{"comment":"The central mechanistic claim, that I_cos tracks task-phase criticality in predicted chunks, is supported by Figure 4a computed on only five LIBERO-90 human demonstration traces, and the observation-distribution mechanism in Section 5.5 is measured on a single task (Task 60). The bridging statement in Section 5.2, that the policy 'learns these features and reproduces them in its predicted action chunks', is asserted rather than demonstrated across tasks and policy families. Please plot I_cos over predicted chunks (rather than demonstrations) for all three policy architectures across more tasks, showing peak alignment with grasp and release phases, and provide evidence that the max aggregation in Eq. 3 is the appropriate statistic rather than, say, the mean directional inconsistency.","section":"5.2, Eqs. 2-3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The in-distribution advantage of FreeSpeed is partly confounded by systematically less aggressive execution: at the 2x command, w/o CR scales every retained action by 2.0 while FreeSpeed caps at f_s = 0.9*rho = 1.8 and reverts to f_c = 1 on high-CI chunks, giving a realized rate of 1.34x versus 1.55x for vanilla on pi0.5 LIBERO (Section D.6). The Appendix E.2 constant-factor sweep is the right control and supports the claim on tasks 34, 48, and 86, but the coverage is narrow. Please add matched-realized-rate comparisons, for example FreeSpeed at 2x against w/o CR at a rate-matched command, across a larger portion of the task set, or extend the E.2 sweep, so that success is attributed to CI gating rather than to overall reduced aggression.","section":"5.5, Appendix E.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Equation (5) contains hand-set constants, lambda = 50, f_s = 0.9*rho, the f_c endpoints in Table A4, and the expanded controller bound L = 4 in Appendix C.3, whose selection provenance is not stated. The sensitivity analysis in Figure 7 shows robustness to lambda and f_s over a range, which mitigates the concern, but the paper should state explicitly whether these values were fixed before or after observing the evaluation tasks, for example whether they were selected on a held-out task subset. If they were tuned on the reported tasks, the claim that the schedule is training-free and transferable is weakened and should be reworded.","section":"4.2, Appendix C.2, Table A4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Success rates are reported as point estimates from 50 simulation and 20 real-robot rollouts per condition without confidence intervals. Several per-task claims that the method matches or exceeds the reference rest on small samples (Table A8's 92 vs 92 at n = 50 has wide binomial uncertainty, and Table 3's per-task real-robot entries use 20 trials each). Please report binomial confidence intervals or tests at least at the task level for the claims that success is preserved.","section":"Tables 2, 3, A8","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"Figure 4a is computed on human demonstrations rather than on predicted chunks; add a sentence in the text or caption making this explicit, alongside the requested predicted-chunk analysis, since the gating in Eq. 5 operates on predicted chunks.","section":"5.2, Figure 4a","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The rate axes are inverted on the slowdown panels; annotate the inversion direction clearly, as the current layout is easy to misread.","section":"Figure 5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"At the 0.3x command the realized FreeSpeed rate (0.402) exceeds the commanded stride; a one-line note explaining that the conservative deceleration endpoints (f_c > rho in Table A4) cause this would prevent confusion.","section":"Table 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The comparison with SuP and TempoVLA uses success rates reported in other papers under different evaluation settings; appendix D.3 acknowledges this, and a one-line caveat near the figure itself would help readers.","section":"5.7, Figure 7a","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"FreeSpeed works, and the paper is honest about its limits. The training-free speed control is a real capability: frozen action-chunking policies can be slowed or sped up at test time with roughly preserved success across three policy families, 50 simulated tasks, and four real-robot tasks. The comparison is well controlled, and the real-robot numbers are credible. The paper reports its confounds honestly — the 4x realized rate averages only over successful trials, the acceleration ceiling is task-dependent, and deceleration can also fail. That is solid engineering. The soft spot is the mechanistic claim, not the empirical result. CI-as-criticality-proxy rests on five demonstration traces and one OOD measurement task. And there is a genuine confound: at 2x, w/o CR scales every step by 2.0 while FreeSpeed caps at 1.8 and pulls most chunks to fc=1, so FreeSpeed is systematically less aggressive. Part of its in-distribution advantage is conservatism, not CI finding the critical phases. The constant-factor sweep in the appendix partially addresses this, but only on three pick-and-place tasks. Success rates lack confidence intervals, and the constants (lambda=50, fs=0.9rho, fc endpoints) are hand-set; a sensitivity sweep mitigates the circularity, but they are still fitted. None of this breaks the central empirical claim. The open question is transfer: on a task where CI does not peak at the critical phase, the advantage over w/o CR could vanish. The paper deserves a serious referee. I would ask for code, confidence intervals, and at least one task class where CI and criticality do not align. For your own work on test-time control, it is worth citing and worth bringing to reading group.","headline":"FreeSpeed is a credible training-free speed controller, with a mechanistic claim thinner than the empirical results.","tokens_in":32108,"tokens_out":1165,"duration_ms":24406},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"The directional inconsistency between consecutive action steps in a predicted chunk marks task-phase criticality, so a frozen robot policy can change execution speed from 0.2x to 4x without retraining and still keep its original success…","keywords":["robot policy speed control","training-free speed adaptation","cosine inconsistency","action chunking","frozen policy","imitation learning","out-of-distribution observations","action rescaling"],"falsifier":"A decisive test is a task whose critical phase runs along straight, directionally consistent motion, such as a straight-line insertion or vertical placement where Icos stays low exactly when the robot can least afford large steps; if FreeSpeed then accelerates through that phase and fails while a conservatively scaled baseline succeeds, the proxy misfires. The paper's own evidence for the proxy covers five LIBERO-90 demonstrations and one distribution-shift task, so a counterexample task where Icos peaks outside the grasp-and-release moments would settle the question.","tokens_in":31814,"feed_emoji":"🤖","tokens_out":5830,"duration_ms":33534,"temperature":0.7,"pith_summary":"Imitation-learned robot policies inherit a fixed speed from their demonstrations, and the usual route to speed control is expensive retraining tied to task-specific data augmentation. This paper argues that a policy's own predicted action chunk already carries a cheap, training-free signal for how much its speed can be safely modified: the cosine inconsistency between adjacent translation increments, which spikes during critical phases such as grasping and placing. FreeSpeed uses that signal to rescale resampled actions, keeping the frozen policy's observations close to its training distribution while pushing execution toward a user-requested rate. Across three policy families, 50 simulated tasks, and four real-robot tasks, the paper reports that FreeSpeed matches or exceeds the frozen policy's success rate over commands from 0.2x to 4x while realizing execution rates between 0.22x and 2.53x. The contribution is a plug-and-play speed interface for policies that would otherwise require retraining or demonstration reprocessing.","feed_headline":"Frozen robot policies gain user-set speed from one action signal","feed_subtitle":"Reading its own action chunks, a policy runs at 0.2x to 4x while keeping its original success rate.","key_machinery":"The central object is the inconsistency-gated rescaling rule ft = fc(rho_t) + (fs(rho_t) - fc(rho_t)) exp(-lambda Icos(A_t)), applied after temporal resampling of the predicted action chunk on a grid of stride rho_t. Here Icos is the maximum of di = 1 - the cosine of the angle between adjacent translation increments within the chunk, which the paper shows peaks during decisive phases like grasping and placement in human demonstrations. The rule does the work: it edges executed step lengths toward the commanded rate when motion is straight and pulls them back toward the 1x reference when directions disagree, so the frozen policy never strays far from its training distribution. A single decay rate lambda = 50 and the endpoint pairs (fs, fc) are shared across tasks and policy families, which is what makes the module plug-and-play rather than task-specific.","core_discovery":"The paper's central claim is that the maximum cosine inconsistency between consecutive translation increments in a predicted action chunk, Icos = max_i (1 - cos of the angle between adjacent delta-p actions), acts as a training-free proxy for task-phase criticality. Where Icos is small, the chunk is directionally consistent and its speed can be modified aggressively; where Icos is large, the policy is in a decisive phase such as grasping, releasing, or placing, and step lengths should stay close to the original predictions. FreeSpeed operationalizes this by resampling the chunk at the commanded stride and rescaling each action by ft = fc + (fs - fc) exp(-lambda Icos), with a single decay constant lambda = 50 shared across all tasks and policies, fs = 0.9 rho for acceleration, and fc near 1.0 for critical chunks. The paper provides evidence that this adaptive recipe keeps the frozen policy's conditioning observations close to the training distribution, and that it applies to any frozen action-chunking policy without policy updates, additional data, or a learned scheduler.","pith_inferences":["The mechanism suggests a general diagnostic: any action-chunking policy that must run at an off-nominal speed could be instrumented the same way, using chunk geometry rather than learned critics or world models to decide when to trust aggressive step lengths.","If Icos is a faithful phase-criticality signal, it could be combined with the tactile or force feedback the authors mention to gate speed during contact-rich phases, where kinematics alone may not warn of danger.","Because FreeSpeed only rescales and resamples, its ceiling is the frozen policy's own competence: on Robomimic aggressive 4x acceleration still drops success to 55.3%, so the gains are real but conditional on a base policy that succeeds in the first place."],"forward_implications":["A frozen vision-language-action model, world-action model, or flow-matching policy can expose a user-requested speed interface at test time without retraining, finetuning, or demonstration resampling.","Because the rescaling is continuous in Icos, speed can be updated within an episode: a command can slow down for a precision phase and speed up again once motion is straight, with success preserved.","The same single schedule (lambda = 50, fs = 0.9 rho) transfers across three policy families and from simulation to a real humanoid arm, so no per-task tuning is needed.","Realized speedup is capped by how much of the task consists of directionally consistent free motion, a limit the authors acknowledge in Appendix F."],"supporting_citations":[{"why":"Motivates the failure mode FreeSpeed is built to avoid: distribution shift and compounding error when the executed trajectory diverges from the training distribution.","marker":"(Ross et al., 2011)"},{"why":"Establishes the action-chunking execution rule of queued actions at a fixed control interval that FreeSpeed operates on.","marker":"(Zhao et al., 2023)"},{"why":"Provides the diffusion-policy action-chunking formulation and the Robomimic baselines used as a frozen policy in evaluation.","marker":"(Chi et al., 2023)"},{"why":"Supplies the frozen pi-0.5 vision-language-action checkpoint evaluated with FreeSpeed and used for the CALVIN comparison.","marker":"(Physical Intelligence, 2025)"},{"why":"Supplies the frozen Fast-WAM world-action-model checkpoint evaluated with FreeSpeed.","marker":"(Yuan et al., 2026)"},{"why":"Underpins the flow-matching policy class trained for the 10 LIBERO-90 and 4 real-robot tasks.","marker":"(Lipman et al., 2023)"},{"why":"Provides the LIBERO benchmark suites and the 50 simulated tasks used for evaluation.","marker":"(Liu et al., 2023)"},{"why":"Supports the intuition that human demonstrators slow down and hesitate during critical phases, which is why predicted chunks carry the Icos signal.","marker":"(Belkhale et al., 2023)"},{"why":"Provides the vanilla resampling baseline and the SuP comparison on LIBERO.","marker":"(Wu et al., 2026)"}],"fun_headline_variants":["No-training speed dial for robot policies via action consistency","One signal lets frozen robot policies speed up or slow down","Training-free speed control reads action chunks to rescale","Robot policies get user-set pace without retraining","FreeSpeed tweaks action steps to hit any target execution rate"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The load-bearing premise is that the maximum cosine inconsistency of a predicted action chunk is a reliable, policy-independent proxy for how much step-length modification a task phase tolerates, so that one fixed schedule (lambda = 50, fs = 0.9 rho) generalizes across tasks and policy families.","fun_headline_variants_meta":{"raw":{"variants":["No-training speed dial for robot policies via action consistency","One signal lets frozen robot policies speed up or slow down","Training-free speed control reads action chunks to rescale","Robot policies get user-set pace without retraining","FreeSpeed tweaks action steps to hit any target execution rate"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000155,"raw_usage":{"cost":0.000155428,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000155428,"upstream_inference_prompt_cost":0.000112308,"upstream_inference_completions_cost":4.312e-05},"total_tokens":1161,"prompt_tokens":1007,"completion_tokens":154,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":75}},"tokens_in":1007,"tokens_out":154,"cost_source":"openrouter_usage","duration_ms":2335,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":75,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-07T00:10:35.827247+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"A decisive test is a task whose critical phase runs along straight, directionally consistent motion, such as a straight-line insertion or vertical placement where Icos stays low exactly when the robot can least afford large steps; if FreeSpeed then accelerates through that phase and fails while a conservatively scaled baseline succeeds, the proxy misfires. The paper's own evidence for the proxy covers five LIBERO-90 demonstrations and one distribution-shift task, so a counterexample task where Icos peaks outside the grasp-and-release moments would settle the question.","supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}