{"id":"21c92a70-2162-4efd-b747-11c4ed8f58ea","arxiv_id":"2506.11948","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A full-stack speed-adaptation system lets imitation-learned robot policies execute up to 3-4x faster than human demonstrations while preserving task success rates.","lead":"This paper presents a system that runs robot manipulation policies learned from human demonstrations faster than the demonstrations themselves, up to 4x in simulation and 3.2x on real robots. The system combines faster control, smoother action generation, and speed adaptation to keep success rates high.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reached-pose controller-invariance (Sec. 4.2) is supported only by open-loop replay on one non-contact simulated task; the paper's own Limitations section concedes dynamic and contact shifts, so the central speedup claim may not transfer to contact-rich tasks.","rationale":"The reader identified the reached-pose controller-invariance assumption as the weakest link, and I agree. The paper's own Limitations section contains the clearest admission that the invariant-target assumption does not cover robot-object interaction dynamics, which are exactly the regimes where the headline speedup must still maintain success. The simulated Can 'throwing' failure and the real Wiping Board contact failure are concrete, self-reported instances where the assumption breaks. This does not invalidate the paper: the ablation study (Table K.6) shows that replacing reached poses with commanded poses degrades performance, the ACT experiments (Table K.7) provide some generality across policy architectures, and the real-world results show consistent TPR gains on 6/7 tasks. These are real supporting evidences. However, the decisive test for the central mechanism is not open-loop replay in a frictionless simulated task; it is closed-loop execution in tasks with dynamics and contact. Until that is measured, the claim that SAIL 'maintains high task success rates' across diverse tasks remains conditional. The reader's CONDITIONAL verdict captures this accurately, so I would not change the verdict.","tokens_in":26785,"tokens_out":7743,"duration_ms":222448,"concrete_test":"Re-run the Fig. G.4 demo-replay study closed-loop: for the simulated Can task with the object free to move and for a contact-rich task such as the real Wiping Board, replay a fixed reached-pose demonstration at c=1 with K_teleop and at c=0.5 with the high-gain Ke, and record executed end-effector trajectories and task success under the actual SAIL policy. If the closed-loop trajectories deviate from the reference by more than the EAG error threshold rho on more than 10% of timesteps, or if success drops by more than 10%, the controller-invariance assumption fails for dynamic or contact-rich tasks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that a trajectory of reached poses from teleoperation remains a valid, trackable reference under a different high-gain controller at compressed time scales (Sec. 4.2). The only direct evidence is the open-loop demo-replay study in Fig. G.4, which uses one simulated task (Can) with no sustained contact. The paper's own Limitations section states that SAIL 'does not explicitly tackle the dynamics shift of robot-object interaction,' and reports that in simulated Can the robot can throw the can at higher speed, and in real Wiping Board the high-gain controller cannot adjust to the new robot-object dynamics. These are precisely cases where the reached-pose trajectory is not controller-invariant: contact forces and object momentum make the achievable pose at speed c_t depend on the controller and dynamic state, not just on the reference. Because the invariance assumption is the foundation of component (2), and because the full SAIL system also includes EAG and adaptive speed modulation that could partially compensate, the empirical speedup may be contingent on task classes where the invariance assumption holds. A controller-matched baseline and a closed-loop invariance test are needed before the generality of the claim is established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SAIL, a full-stack system for executing offline visuomotor imitation policies faster than the demonstration speed. SAIL combines error-adaptive guidance (EAG) to maintain temporal consistency of action chunks, training on reached poses with a high-fidelity tracking controller to reduce controller-induced distribution shift, adaptive speed modulation that slows critical phases of a task, and latency-aware action scheduling with a derived lower bound on the action interval. The authors evaluate SAIL on five simulated RoboMimic/MimicGen tasks and seven real-world tasks on two robot platforms, reporting up to a 4x speedup in simulation and up to a 3.2x speedup in the real world while claiming high task success rates.","tokens_in":26980,"tokens_out":8267,"duration_ms":104781,"significance":"The problem of faster-than-demonstration execution is well motivated and rarely addressed end-to-end; the decomposition into policy-level consistency, controller-invariant targets, adaptive speed, and latency management is a useful systems contribution. The empirical breadth is substantial, with 12 tasks, two distinct real platforms, ablations, and a variant using ACT. The EAG design insight that CFG conditioning should be disabled when tracking error is large is plausible and is supported by the Appendix H correlation studies and by the simulation ablation. However, the headline speedup numbers are not yet fully trustworthy because of an asymmetric simulation setup and the weak evidential support for the central reached-pose invariance claim.","major_comments":[{"comment":"The simulation evaluation removes joint torque limits for the Franka robot, while at least some baselines are reported with torque limits kept when removing them hurt performance. This asymmetric treatment directly favors SAIL in the headline simulation speedup comparisons reported in Table 1. Please re-run all methods under identical torque-limit settings and add a DP-Fast baseline that uses the same high-gain controller and reached-pose targets as SAIL, so the effect of the policy-level components is isolated from the controller change.","section":"Sec. C.1, Table 1"},{"comment":"The controller-invariance of reached poses is the foundation of component (2), but the only direct evidence is the open-loop demo replay study in Fig. G.4 on the Can task, which does not involve sustained contact. The Limitations section explicitly concedes that SAIL does not tackle robot-object dynamics shift and reports that the simulated Can can be thrown at higher speed and that the real Wiping Board task degrades under the high-gain controller. These are precisely regimes where the achievable pose at a given speed depends on the controller and dynamic state rather than only on the reference. Please provide a closed-loop test of reached-pose invariance on contact-rich tasks, or explicitly narrow the claim to task classes where the invariance holds.","section":"Sec. 4.2, Fig. G.4, Sec. 7"},{"comment":"The real-world headline of 3.2x speedup is taken from the Wiping Board row, where SAIL has lower success rate (0.70 vs 0.90) and lower throughput-with-regret (3.18 vs 3.48) than DP-Fast. Reporting the maximum SOD from a task where the method is worse than the baseline on the primary throughput metric overstates the result. Please report confidence intervals or per-task error bars for the 10-rollout evaluations, and either match or exceed baseline success rate before claiming a speedup, or explicitly present the speed-success trade-off.","section":"Sec. 5.2, Table 2"},{"comment":"The derivation of the lower bound delta_lb is under-specified: it uses the full prediction horizon H_p, but the interval that must cover the sensing-inference delay in a receding-horizon schedule is the remaining plan length after the already-executed steps H_e (defined as 8 in Sec. 5.1), not the full H_p. The conditioning length H_c also appears in the inequality without a precise timing diagram. Please state the schedule algebraically with all variables (H_p, H_e, H_c, delta_delay) and re-derive the bound; as written, the inequality may overestimate the achievable speedup.","section":"Sec. E, Eq. (8)"}],"minor_comments":[{"comment":"The real-world results are based on only 10 rollouts per task and no error bars or statistical tests; given the large variance implied by success rates such as 0.30 and 0.40, confidence intervals are needed before quantitative comparisons can be assessed.","section":"Table 2"},{"comment":"The EAG error thresholds and CFG weights are tuned per task and are listed only in the appendix; this should be acknowledged as a limitation in the main text, since the adaptive guidance therefore depends on task-specific hyperparameters.","section":"Table J.3"},{"comment":"The figure contains draft annotations such as \"(1/28) DX: Made a few changes\" and \"(2/1 5AM) SK: fixed notation\" that should be removed before publication.","section":"Fig. E.2"},{"comment":"There is a typo: \"oreintation\" should be \"orientation\" in the definition of tracking error.","section":"Sec. H.3"},{"comment":"The notation for c_slow and c_fast is confusing: the text says c_slow > c_fast, which is correct only if the reciprocal of c is the speedup, and Table J.4 labels a single \"slowdown c\" per task. Please define the speedup factor consistently and reconcile the table with the equations.","section":"Sec. 4.3 and Table J.4"},{"comment":"The text refers to an ablation SAIL(-C) while Table K.6 uses columns -HG, -AS, and -C inconsistently; make the naming in the text and table match.","section":"Sec. 5.1 and Table K.6"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the systems contribution is real, but the primary risk is baseline fairness. The torque-limit asymmetry and the absence of a controller-matched DP-Fast baseline are fixable in revision and should not be papered over. The reached-pose invariance claim either needs stronger closed-loop evidence or a narrowed task-class claim. I would not reject the paper, but the headline speedup numbers should not be accepted as stated without these changes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What should you know: SAIL is a serious attempt to formalize faster-than-demonstration execution of offline imitation policies, and the system actually works on a broad set of tasks. The main thing to watch is not the engineering but the theoretical load-bearing claim about controller-invariant reached poses, which has thin evidence.\n\nThe genuinely new piece is the problem formalization: earlier work got speedup as a byproduct or needed online interaction, and this paper makes it the primary objective. EAG—adaptively switching off CFG guidance when tracking error is high—is a clean idea and the authors support it with a careful OOD analysis (Appendix H) and a beneficial ablation (Table K.6). The latency-bound derivation, while simple, correctly identifies the physical limit on speedup. The simulation and real-world experiments are extensive, with two distinct platforms, and the results are consistent: SAIL has higher throughput than DP-Fast on nearly all tasks. Table K.6 shows each component contributes.\n\nThe soft spots are real but nuanced. First, the torque-limit removal (Sec C.1) could favor SAIL; the paper admits that for some baselines keeping torque limits gave better results, and they report those. That is cherry-picking, and it undermines the headline speedup comparison. A controller-matched baseline—DP-Fast with the same high-gain controller—is also missing, so part of SAIL's advantage may come from the controller rather than the learning-side contributions. Second, the controller-invariance claim (Sec 4.2) rests on a single open-loop replay experiment on the simulated Can task (Fig G.4), which involves no sustained contact. The paper's own Limitations section and the Wiping Board result show that with sustained contact, the reached-pose target is not controller-invariant. That is the kind of honest negative evidence that makes the paper credible, but it also means the central assumption is narrower than stated. Third, many hyperparameters (guidance weight, error thresholds, speed factors, DBSCAN parameters, controller gains) are tuned per task, which limits the transferability claim. Finally, the real-world evaluation uses 10 rollouts per task with no error bars; for some tasks the differences are large enough that this is fine, but for closer ones (e.g., Stacking Cups) it is thin.\n\nNone of this breaks the paper. The problem definition and EAG are contributions, and the system is well-built. The paper would benefit from a controller-matched baseline, a closed-loop test of the invariance assumption, and more real rollouts. I would send it to serious peer review and expect a conditional accept after those additions. For anyone working on fast execution or full-stack IL, this is worth reading and citing.","headline":"SAIL is a well-engineered, genuinely novel formalization of faster-than-demonstration execution with a few real evaluation soft spots, the biggest being the thin evidence for controller-invariant reached poses.","tokens_in":27613,"tokens_out":4627,"would_cite":true,"duration_ms":50049,"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":"SAIL enables learned visuomotor policies to run 4x faster than demonstrations in simulation and 3.2x faster on real robots while keeping success rates high.","keywords":["faster-than-demonstration execution","imitation learning","visuomotor policy","distribution shift","controller-invariant target","adaptive speed modulation","action scheduling","manipulation throughput"],"falsifier":"A decisive experiment: on a robot not used in the paper, replay the same teleoperation demonstrations at increasing speedup factors with both commanded-pose references and reached-pose references under the high-gain controller, and measure tracking error and task success. If reached-pose replay at high gain drops toward the commanded-pose baseline rather than staying near the original-speed success rate, the controller-invariance assumption is violated; the paper's own contact-rich wiping case already suggests sustained contact is the most likely place for this to happen.","tokens_in":1857,"feed_emoji":"🤖","tokens_out":4889,"duration_ms":133996,"temperature":0.7,"pith_summary":"The paper formalizes faster-than-demonstration execution of visuomotor imitation policies: given a policy trained offline on demonstrations with a fixed time step, run it with a smaller, time-varying time step to raise task throughput. It argues that naive speedup fails because faster motion changes the low-level controller's tracking behavior, pushing the policy's observations out of distribution, while high-gain controllers amplify reference noise and system latency caps the fastest feasible loop. SAIL (Speed Adaptation for Imitation Learning) is a full-stack answer: it predicts reached rather than commanded poses, tracks them with a high-fidelity controller, switches on action-conditioned guidance only when tracking error is small, modulates speed down around critical actions, and schedules actions so inference delay never starves the controller. In experiments on 12 tasks across simulation and two physical robot platforms, SAIL achieves up to 4x speedup in simulation and 3.2x speedup on hardware over demonstration speed while retaining high success rates.","feed_headline":"Imitation-learned robots run up to 3.2x faster on real hardware","feed_subtitle":"By predicting reached poses and slowing for critical actions, SAIL keeps success rates high while running 3.2 times faster.","key_machinery":"The carrying object is the receding-horizon action chunk from a generative visuomotor policy, together with three system-level modifications. The first is controller-invariant action targets: the policy regresses reached end-effector poses $x_t$, not teleoperator-commanded poses $x_t^d$, so the reference trajectory is achievable and independent of the teleoperation controller's dynamics; a high-gain operational-space controller $K_e$ tracks these reached poses at a reduced interval $\\delta_t = c_t \\delta^*$. The second is Error-Adaptive Guidance (EAG), which computes a tracking-error proxy $e = \\mathrm{error}(x^d, x_{\\mathrm{current}})$ and applies a classifier-free guidance weight $w>0$ only when $e \\le \\rho$; otherwise it sets $w=0$ and uses the unconditional prediction. The third is adaptive speed modulation: a predicted critical-action flag $k_t \\in \\{0,1\\}$ sets $c_t = k_t c_{\\mathrm{slow}} + (1-k_t)c_{\\mathrm{fast}}$. The fourth is action scheduling, with the lower bound $\\delta_{lb} > \\delta_{\\mathrm{delay}}/(H_p - H_c)$ guaranteeing continuous execution under inference latency.","core_discovery":"The central claim is that the main obstacle to faster-than-demo execution is not simply generating faster actions; it is the coupling between execution speed, controller dynamics, and the policy's implied state-action distribution. The paper argues that this coupling is broken by changing what the policy predicts: train it to predict the actually reached end-effector poses from teleoperation data rather than the commanded poses, because reached poses are largely controller-invariant and remain achievable references when a different, higher-gain controller executes them at a faster rate. On the policy side, temporal consistency between successive action chunks should be enforced by classifier-free guidance only when current tracking error is small; when error is large, the conditioning signal is out of distribution and guidance should be disabled. Execution speed itself is made adaptive: the system identifies critical actions from motion complexity and gripper events and slows down for those phases. Finally, scheduling actions across sensing-inference delay with a lower bound on the action interval prevents action exhaustion and gives a physical ceiling on achievable speedup. Together these four components are claimed to produce up to 4x and 3.2x speedups in simulation and the real world with high success rates.","pith_inferences":["The paper leaves untested whether the reached-pose target could be replaced by any controller-invariant motion descriptor, such as a learned dynamics-aware target in joint space; a controlled experiment that swaps only the action-space representation while holding EAG, scheduling, and adaptive speed fixed would test this modularity claim.","EAG is a guidance-time mechanism, so the same threshold logic should transfer to other action-chunking generative policies beyond the representative model evaluated here; the conditioning-misalignment problem is generic, though the optimal tracking-error threshold likely needs re-estimation per model and task.","The speedup ceiling implied by the paper depends on demonstration data quality: clean, high-bandwidth teleoperation yields reached-pose targets that the high-gain controller can track, whereas noisy human demos would amplify reference noise, so the gains should be largest on clean demonstration sets."],"forward_implications":["Speeding up an offline imitation policy is a full-stack problem, not just a policy-generation problem: action smoothing alone cannot recover failures from controller shift, latency, and critical-phase precision.","If the controller-invariance claim holds, any fixed offline policy can be retargeted to run at a faster time-varying interval by replacing commanded-pose supervision with reached-pose supervision and deploying a high-gain tracker, without additional environment interaction.","Adaptive speed modulation lets task throughput improve without sacrificing precision: the policy automatically slows for grasps and alignments and speeds up during simpler reaching motion.","The latency-derived bound $\\delta_{lb} > \\delta_{\\mathrm{delay}}/(H_p - H_c)$ makes feasible speedup directly tied to sensing-inference delay and prediction horizon, so faster inference or longer horizons translate into higher achievable speedups.","Because EAG disables conditioning guidance when tracking error is high, the same mechanism that smooths motion also keeps the policy from trusting out-of-distribution conditioning signals, which supports high success rates at speed."],"supporting_citations":[{"why":"supplies the generative action-chunking visuomotor policy that SAIL accelerates and serves as the primary baseline.","marker":"[1]"},{"why":"identifies temporal inconsistency between consecutive action chunks and provides the BID-Fast consistency baseline.","marker":"[8]"},{"why":"provides classifier-free guidance, which EAG adaptively enables or disables based on tracking error.","marker":"[32]"},{"why":"supplies the waypoint extraction approach on which SAIL's motion-complexity critical-action detector is built.","marker":"[34]"},{"why":"motivates the system-level treatment of sensing-to-action delay and action scheduling.","marker":"[29]"},{"why":"offers an alternative generative policy architecture used to demonstrate that SAIL's components transfer beyond the representative model.","marker":"[2]"}],"fun_headline_variants":["Imitation-learned robots run up to 3.2x faster on real hardware","Robots execute imitation policies 3.2x faster with SAIL","SAIL attains 4x speedup in simulation, 3.2x on real robots","Faster-than-demo execution via reached-pose prediction and adaptive speed","Robots beat demonstration speed by up to 4x with SAIL"],"cache_read_input_tokens":29696,"weakest_assumption_plain":"The load-bearing premise is that a trajectory of actually reached poses from teleoperation stays a valid, trackable reference when a different, higher-gain controller runs at a faster speed; if that controller-invariance fails on other robot dynamics or contact-rich phases, the distribution-shift mitigation and high-fidelity tracking lose their foundation and the speedup gains do not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Imitation-learned robots run up to 3.2x faster on real hardware","Robots execute imitation policies 3.2x faster with SAIL","SAIL attains 4x speedup in simulation, 3.2x on real robots","Faster-than-demo execution via reached-pose prediction and adaptive speed","Robots beat demonstration speed by up to 4x with SAIL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000337,"raw_usage":{"total_tokens":1901,"prompt_tokens":1015,"completion_tokens":886,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":781}},"tokens_in":631,"tokens_out":886,"duration_ms":8853,"temperature":1.0,"reasoning_tokens":781,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T01:00:08.014522+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive experiment: on a robot not used in the paper, replay the same teleoperation demonstrations at increasing speedup factors with both commanded-pose references and reached-pose references under the high-gain controller, and measure tracking error and task success. If reached-pose replay at high gain drops toward the commanded-pose baseline rather than staying near the original-speed success rate, the controller-invariance assumption is violated; the paper's own contact-rich wiping case already suggests sustained contact is the most likely place for this to happen.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the generative action-chunking visuomotor policy that SAIL accelerates and serves as the primary baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the waypoint extraction approach on which SAIL's motion-complexity critical-action detector is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"motivates the system-level treatment of sensing-to-action delay and action scheduling."}],"review_version":1}