{"id":"86d6106d-2af2-40ae-8db4-056fd840e1cf","arxiv_id":"2607.24079","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Finite simulation frequency leaks proportional gain into effective damping and derivative gain into effective inertia, so simulators should use deliberately wrong parameters to match real behavior.","lead":"This paper argues that robot simulators can become more accurate by deliberately using 'effective' parameters that differ from physically measured values, absorbing the physics a simulator leaves out. It derives explicit corrections for PD control at finite simulation frequency and re-frames rope manipulation and underwater swimming as examples of the same renormalization idea.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The missing link between finite f_sim and a pure state delay (Eqs. 3–4) is load-bearing: standard discrete-time integrators such as semi-implicit Euler do not yield the first-order K_P→K_D and K_D→J mixing in Eqs. (9)–(17), so the central quantitative claim is unverified.","rationale":"The paper is a clear perspective; the analytic derivation of Eqs. (9)–(17) is internally consistent under the delayed-feedback model, and the dimensional analysis is a useful consistency check. The problem is that the bridge from 'finite simulation frequency' to a pure state delay is asserted (Eq. 3), not derived. The reader flagged this, and I can sharpen it: common symplectic integrators do not have the same first-order effective dynamics. For semi-implicit Euler, the first-order effective damping is K_D and the effective inertia is J; no K_P mixing appears. Thus this is not merely an unquantified scale factor δt; the structure of the correction depends on the discretization. Because the paper's main quantitative contribution is the sim-parameter correction, this gap is load-bearing. The rope and fish examples are interpretive and do not validate the PD formulas. I keep the reader's CONDITIONAL verdict: if the concrete simulation test reproduces the delay-model formulas in a representative simulator, the central claim is supported; if not, the quantitative claim should be rejected or restricted to deliberately delayed feedback loops. The concern is about the argument's assumption, not about the authors.","tokens_in":6561,"tokens_out":14470,"duration_ms":128813,"concrete_test":"Run a 1-DOF PD step response in a representative semi-implicit-Euler simulator (or MuJoCo with default integrator) at f_sim=30, 60, and 120 Hz, with J=1, K_P=100, K_D=10, qd=0, q̇d=0. Fit the effective damping and natural frequency from the decaying oscillation and compare with Eqs. (16)–(17) using δt=1/f_sim and δt=1/(2f_sim). If the fitted K_D^eff and J^eff do not shift by approximately δtK_P and δtK_D, the pure-delay equivalence is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 replaces finite simulation frequency by q(t−δt), q̇(t−δt) in the continuous PD ODE (Eq. 4), then expands to obtain effective parameters (Eqs. 9–17). This pure-delay model is not the discrete-time law implemented by common physics simulators. A simulator with control at f_sim computes torque from the current q_n,v_n and either holds it over the step (zero-order hold) or uses a symplectic/semi-implicit update. For a 1-DOF PD at constant setpoint, semi-implicit Euler has characteristic multiplier magnitude sqrt(1−hγ)=1−hγ/2+O(h²) and phase hω+O(h³), where γ=K_D/J, ω²=K_P/J, h=1/f_sim. To first order, the damping is K_D and the inertia is J, with no K_P→K_D or K_D→J mixing. The paper's Eq. (17) would require J^sim=J_measured+δtK_D; for this integrator the first-order shift is absent. Since δt is never derived from a concrete integration scheme (the paper only says its magnitude depends on the simulator), the central quantitative advice may be wrong for the simulators it addresses. The analytic derivation is correct under its stated model, but the model is assumed, not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that sim-to-real transfer can be improved by deliberately using 'renormalized' simulator parameters that differ from measured physical values. The central analytical result is in Section 2: modeling the finite simulation frequency as a small pure delay δt in the PD-control law leads to K_D^eff = K_D − δt K_P and J^eff = J − δt K_D, and hence the recommended simulator parameters K_P^sim = K_P^real, K_D^sim = K_D^real + δt K_P^real, J^sim = J_measured + δt K_D^real (Eqs. 9–17). Appendix A extends this to multi-joint robots with an effective-action derivation. Sections 3 and 4 reinterpret existing rope-manipulation and robotic-fish studies, and Section 5 sketches a practical procedure for applying effective parameters to new problems.","tokens_in":6943,"tokens_out":6653,"duration_ms":57793,"significance":"The paper's core idea is attractive and the Section 2 algebra is internally consistent and parameter-free under the pure-delay assumption. The effective-action appendix is a nice formal complement. The paper is honest about the need to determine δt from the simulator, but this honesty exposes the central gap: no derivation or experiment shows that a finite-f_sim simulator is equivalent to a pure state delay. If that equivalence holds, the quantitative correction is a valuable, simple recipe. As it stands, the paper is a conceptual proposal with an unverified central quantitative claim.","major_comments":[{"comment":"The central assumption is that finite simulation frequency can be represented by q(t−δt), q̇(t−δt) with δt ∝ 1/f_sim. This is not derived. For common discrete-time simulators, it is false at first order. For example, a 1-DOF PD controller at constant setpoint integrated with semi-implicit Euler at step h has the update q_{n+1}=q_n+h v_{n+1}, v_{n+1}=v_n+h(−ω² q_n − γ v_n), where ω²=K_P/J and γ=K_D/J. The characteristic multiplier magnitude is √(1−hγ) = 1−hγ/2+O(h²), i.e. the effective damping is K_D and the inertia is J to first order; there is no K_P→K_D or K_D→J mixing. Thus Eqs. (9)–(17) are not the generic first-order corrections for finite-frequency simulation. Since the paper explicitly states that 'the precise magnitude of δt depends on the simulator implementation,' it leaves the structure of the correction unverified. This is load-bearing: the central advice J^sim = J_measured +","section":"Section 2, Eqs. (3)–(4)"},{"comment":"No experiment or simulation validates the renormalized parameters. The paper claims that a deliberately wrong simulator inertia produces a more faithful simulation, but it never demonstrates this. A simple discrete-time simulation with a PD controller at finite f_sim, comparing J_sim = J_measured against J_sim = J_measured + δtK_D, would directly test Eq. (17). Similarly, Sections 3 and 4 cite successful coarse models, but those models were fitted/tuned in the original studies and do not use the derived corrections; they illustrate the philosophy but do not validate the quantitative claim of this paper.","section":"Abstract and Section 2"},{"comment":"The rope and swimming examples are retrospective interpretations. Equation (21) and Eq. (22) merely assert that a simplified model with effective parameters matches a task-relevant observable. This is standard model identification, not a demonstration of the specific renormalization mechanism (resolution-dependent parameter shifts) derived in Section 2. If these examples are intended as supporting evidence, the paper should provide at least one case where the effective parameter is varied with resolution and shown to preserve the chosen observable.","section":"Sections 3–4, Eqs. (20)–(23)"}],"minor_comments":[{"comment":"The proportionality factor between δt and 1/f_sim is not specified. For a fixed-step integrator, is δt = h, h/2, or scheme-dependent? This matters for the numerical values in Eqs. (16)–(17).","section":"Section 2, after Eq. (3)"},{"comment":"The statement 'Dimensional analysis fixes the form of the corrections' is overstated: dimensional analysis alone gives possible combinations, not signs or coefficients. The Taylor expansion provides those, so the sentence should be rephrased.","section":"Section 2, Eq. (19)"},{"comment":"The notation O_sim(a, θ*_sim(a)) ≈ O^real uses O for both the simulator output and the real observable. Please use distinct symbols or define the two clearly.","section":"Eq. (1)"},{"comment":"The symbol q^{(3)} is not defined; specify that it denotes d³q/dt³.","section":"Appendix A, Eq. (50)"},{"comment":"Several references are listed with 2026 dates and appear to be preprints. If they are not yet publicly available, add arXiv identifiers or other access information.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is essentially a perspective/position piece with a small analytical core. The central derivation is correct under an assumption that is not established. I recommend major revision rather than rejection because the gap can be addressed by deriving δt for a concrete integrator (or class of integrators) and by presenting a simple numerical test of Eq. (17). The current manuscript, without such validation, does not support the quantitative claim in its abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nI read the Sun et al. paper on renormalization for robotics. The general idea—use resolution-dependent effective parameters to absorb omitted physics rather than building ever more accurate simulators—is genuinely fresh and worth taking seriously. The specific analytic results are also new: explicit first-order formulas for how finite-frequency PD control changes effective derivative gain and inertia (Eqs. 9–17), plus a multi-joint effective-action version in Appendix A. Under the paper's stated model (a continuous-time PD loop with a small state delay), the Taylor expansion is correct, and the dimensional analysis is a nice consistency check.\n\nBut there is a load-bearing problem. The paper models finite simulation frequency as a pure state delay q(t−δt). That is an assumption, not a derivation—the text says 'the precise magnitude of δt depends on the simulator implementation' but never connects it to an actual integrator. For a typical simulator using semi-implicit Euler with zero-order hold, a first-order discrete-time analysis gives the usual damping KD and inertia J, with no mixing of KP into KD or KD into J. So Eqs. 16–17, the paper's central quantitative advice, may simply be wrong for the simulators it claims to fix. The delay model is not validated, and the numerical experiments that could confirm it are absent.\n\nThe rope and underwater examples are interesting as illustrations but are reinterpretations of other people's work, not demonstrations. And the title/abstract says 'real behavior,' which overstates what the paper shows.\n\nNone of this makes the paper worthless. The renormalization perspective is valuable, and the formal derivation in Appendix A is a nice piece of work. But the main new result is unverified and likely incorrect under common integration schemes. I'd send this to peer review—the idea is important enough that a careful referee should engage with it—but I would not accept it without a real validation of the delay model or a corrected analysis.\n\nFor your reading group, it could spark a good discussion about modeling assumptions in sim-to-real.\n\nBest.","headline":"Fresh renormalization framing, clean delay-based PD derivation, but the delay model likely doesn't match standard simulators, leaving the central quantitative claim unverified.","tokens_in":7421,"tokens_out":5069,"would_cite":false,"duration_ms":44401,"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":"Part of the sim-to-real gap is the simulator's finite time step, and the fix is to use deliberately renormalized gains and inertia instead of measured values.","keywords":["renormalization","sim-to-real gap","PD control","effective parameters","finite simulation frequency","simulator fidelity","robot simulation","inertia correction"],"falsifier":"Measure a real PD-controlled joint and its simulator at several decreasing frequencies fsim. If the central claim is right, the simulator parameters that best match real motion satisfy K_D^sim = K_D^real + δt K_P^real and J^sim = J_measured + δt K_D^real with δt ∝ 1/fsim, so corrections grow linearly in 1/fsim. If fitting instead reveals a different dependence—say, corrections scaling as 1/fsim² or a shift in K_P itself—the fixed-delay model is not the right representation of the simulator's discretization.","tokens_in":1190,"feed_emoji":"🤖","tokens_out":1294,"duration_ms":34724,"temperature":0.7,"pith_summary":"This paper argues that robot simulators can match real behavior better by using effective parameters that differ from measured physical values, because those parameters absorb the details the simulator leaves out. The central analytical example is proportional–derivative (PD) control at finite simulation frequency: a small state delay mixes the proportional gain into the effective derivative gain and the derivative gain into the effective inertia. As a result, entering a slightly larger derivative gain and a deliberately larger inertia than measured reproduces real motion at low simulator frequency. The paper extends the idea qualitatively to rope manipulation and underwater swimming, then proposes a practical four-step procedure for choosing such renormalized parameters.","feed_headline":"Slightly wrong simulator inertia matches real robots","feed_subtitle":"Finite time steps change effective gains; renormalized parameters, not measured ones, give faithful simulation.","key_machinery":"The central object is the effective delay δt ∝ 1/fsim inserted into the finite-frequency PD law, followed by a first-order Taylor expansion that moves the acceleration correction to the left-hand side of the equation of motion. This produces renormalized parameters (Eqs. 9–17) that absorb the discretization effect. An appendix gives a complementary multi-joint derivation using a history field y(s,t)=q(t−s) and an effective-action argument, which yields the same first-order corrections.","core_discovery":"At finite simulation frequency, replace the simulator's discrete controller update by continuous-time PD control on delayed states q(t−δt), q̇(t−δt), where δt ∝ 1/fsim. Expanding through first order in δt gives the effective derivative gain K_D^eff = K_D − δt K_P, the effective inertia J^eff = J − δt K_D, and a small velocity-feedforward term δt K_P q̇_d. To make the simulator reproduce the real robot, the paper derives K_P^sim = K_P^real, K_D^sim = K_D^real + δt K_P^real, and J^sim = J_measured + δt K_D^real. In particular, the simulator inertia should not equal the measured inertia; the deliberately incorrect value compensates for dynamics lost at the finite simulation frequency.","pith_inferences":["Editorial inference: the same delay-expansion logic could be applied to other discrete-time effects—zero-order hold, integrator phase lag, or substepping—by fitting an effective δt for each simulator, which would extend the formulas beyond the ideal delayed-state model.","Editorial inference: the central claim predicts a falsifiable quantitative signature—if real and simulated systems are compared at multiple simulation frequencies, the required corrections should grow linearly with 1/fsim, with slopes set by the real gains.","Editorial inference: for deformable objects and fluids, renormalized parameters may be task-dependent; the paper's procedure implies that coefficients calibrated to preserve one observable might fail when the preserved observable changes, so validation under new conditions is essential."],"forward_implications":["Low-frequency simulators can become more faithful by renormalizing gains and inertia, without raising the simulation frequency.","Measured physical parameters are not automatically the best simulator parameters; renormalized values can produce more faithful simulated motion.","The corrections scale explicitly with δt: K_D^sim and J^sim should increase linearly with the effective delay.","A moving target introduces a small additional feedforward correction, absorbable by rescaling desired velocity or by using separate feedforward and feedback derivative gains.","At second order, the effective inertia gains an additional +δt²/2 K_P term and a higher-derivative term appears, bounding the validity of the first-order formulas."],"fun_headline_variants":["Use wrong inertia in sim to match real robot","Effective inertia, not measured, gives faithful simulation","Finite step PD control: renormalize gains and inertia","Simulator's inertia should be off to hit real dynamics","Renormalization for robots: effective params bridge sim-to-real"],"cache_read_input_tokens":8704,"weakest_assumption_plain":"The derivation rests on replacing finite simulation frequency with a small effective delay δt whose precise magnitude is not derived from any actual simulator update rule; if real simulators impart effects that are not equivalent to a pure state delay, the first-order renormalized parameter formulas need not be correct.","fun_headline_variants_meta":{"raw":{"variants":["Use wrong inertia in sim to match real robot","Effective inertia, not measured, gives faithful simulation","Finite step PD control: renormalize gains and inertia","Simulator's inertia should be off to hit real dynamics","Renormalization for robots: effective params bridge sim-to-real"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1197,"prompt_tokens":711,"completion_tokens":486,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":421}},"tokens_in":455,"tokens_out":486,"duration_ms":5278,"temperature":1.0,"reasoning_tokens":421,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T23:04:29.447158+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure a real PD-controlled joint and its simulator at several decreasing frequencies fsim. If the central claim is right, the simulator parameters that best match real motion satisfy K_D^sim = K_D^real + δt K_P^real and J^sim = J_measured + δt K_D^real with δt ∝ 1/fsim, so corrections grow linearly in 1/fsim. If fitting instead reveals a different dependence—say, corrections scaling as 1/fsim² or a shift in K_P itself—the fixed-delay model is not the right representation of the simulator's discretization.","supporting_citations":[],"review_version":1}