{"id":"08b8fe95-4ae1-45c4-a057-d1c5055847b1","arxiv_id":"2509.06342","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"PACE fits a compact set of actuator parameters from brief in-air data and trains energy-aware locomotion policies that transfer zero-shot to real quadrupeds without dynamics randomization.","lead":"This paper presents PACE, a method that aligns a robot simulator to reality by fitting four per-joint parameters and a command delay from 20 seconds of in-air motion, then trains walking controllers without dynamics randomization. It reports a 32 percent reduction in ANYmal's cost of transport (to 1.27) and deployment across more than ten additional robot platforms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Transfer claim rests on load-independent actuator parameters fitted in air; stance loads may invalidate the constant Ia, d, tau_f, and the paper provides no mechanical check of this extrapolation.","rationale":"The reader identifies the in-air-to-stance extrapolation as the weakest assumption. I agree with the direction but not the exact mechanism. The concern is not primarily that Equation 6's constant Ia cannot represent the posture-dependent reflected base inertia: in a full rigid-body simulator, base mass and link inertias couple through multibody dynamics, and the fitted Ia, d, and tau_f are actuator-side parameters intended to capture drive dynamics. The paper explicitly collects in-air data to avoid the base-inertia confound (Section 2.1, Appendix B). The real gap is load dependence: the actuator parameters are identified at near-zero joint torque and no ground contact, yet stance locomotion involves large joint torques and contact forces. Section 2.2 admits that damping and torque-current maps are often state-dependent, and Section 2.4's assumptions do not include a load-independence validation. The fourfold increase in full-robot Ia over the single-drive value (Section 4.1.2) shows that the fitted parameters absorb load- or configuration-dependent effects and are not pure motor constants. If friction, damping, or effective inertia under load differ, the simulator's stance dynamics and energy losses are wrong, and the claim that no dynamics randomization is needed loses its mechanical foundation. The empirical cross-platform results are strong, including single-drive verification, in-air validation across gains, and open-loop base-velocity replays (Figure 6), but those deployments are one-shot demonstrations, not controlled load-dependence tests. The ANYmal CoT reduction is a separate weakness: it compares ANYmal D (this paper) to ANYmal C (prior work, reference 71) with different ambient temperature, firmware, and average speed, so the 32% figure is not a same-hardware A/B result. However, this affects the energy-efficiency headline, not the transfer takeaway itself. The proposed load-dependence test would settle whether the central transfer claim rests on a valid extrapolation. Since the conditionality already captures the need for confirmation, including code release and same-hardware energy comparison, I do not change the verdict.","tokens_in":26240,"tokens_out":11809,"duration_ms":118863,"concrete_test":"Re-run the single-drive mechanical-loop identification of Section 3.2.1 with a controlled external load applied to the output shaft (e.g., a hysteresis brake or a loaded spring) over the joint-torque range seen in stance (approximately 0-100 Nm for TYTAN), re-fitting {Ia, d, tau_f} from chirps at each load. If any fitted parameter shifts by more than the reported experimental uncertainty (e.g., the 0.67-14% spread in Figure 5b), the in-air zero-load parameters are load-dependent and the no-randomization transfer claim must be qualified. As a complementary check on a torque-instrumented platform such as ANYmal, compare predicted joint torques from the fitted simulator against measured torques during stance locomotion; an error distribution exceeding the in-air validation error would confirm the extrapolation gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The takeaway in Section 6 claims that a constant per-joint set {Ia, d, tau_f, q_tilde_b, Td} identified from fixed-base, zero-contact in-air chirps (Section 2.2) is sufficient to eliminate dynamics randomization in stance. The load-bearing assumption is that the fitted actuator parameters are load-independent. Section 2.2 itself concedes that 'load-independent damping... is common but imperfect; actual damping and torque-current maps are often state dependent,' and Section 2.4 lists only that nonlinearities are 'modest in the excited range.' The identification is performed with no ground reaction forces and, on ANYmal, without joint torque sensors, so the fit is anchored only to joint position trajectories under near-zero joint loads. The full-robot fit in Section 4.1.2 already shows Ia about four times the single-drive value, demonstrating that Ia absorbs configuration-dependent link and compensation effects rather than being a pure motor constant. In stance, joint torques reach tens of Nm (Figure 6d) and reflected base inertia dominates (Appendix B, Figure 23), yet no experiment re-identifies the parameters under load or checks predicted torques against measured torques on a torque-instrumented platform. If d and tau_f (or the effective Ia) are load-dependent, the simulator's stance dynamics and its energy model are wrong in a way that dynamics randomization would normally absorb; PACE deliberately omits that safety net. The empirical successes on TYTAN, ANYmal, and MINIMAL are real evidence, but they are single-run demonstrations, and the ANYmal energy comparison is cross-paper (ANYmal C vs D, different firmware, temperature, and speed), so they do not by themselves establish that the in-air zero-load parameters are valid in stance.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PACE, a sim-to-real framework for legged robots. The method identifies a compact per-joint parameter set (armature inertia, viscous damping, Coulomb friction, joint bias, and a global delay; 4n+1 parameters total) by fitting simulated fixed-base, in-air joint trajectories to real encoder-only chirp data using CMA-ES in massively parallel simulation. Policies are then trained with PPO in the fitted simulator without dynamics randomization, using a four-term reward that includes a physics-based electrical and mechanical loss model. The authors validate the pipeline bottom-up on a single drive, on full robots in air, and in on-ground locomotion on ANYmal, TYTAN, and MINIMAL, with additional deployments on ten other robots. They report a 32% reduction in ANYmal's full Cost of Transport to 1.27, and a TYTAN CoT of 0.97, measured on a 400 m running track via battery state of charge.","tokens_in":26605,"tokens_out":6654,"duration_ms":60672,"significance":"If the central claims hold, the paper would make a valuable practical contribution: roughly 20 seconds of encoder-only in-air data per robot would suffice to train deployable, blind locomotion policies without dynamics randomization, while also improving energy efficiency through a first-principle energy reward. The bottom-up validation is a genuine strength: the single-drive experiments recover analytic inertia values within 2--15%, the full-robot in-air comparison shows clear improvement over a URDF-only baseline and competitive behavior against a learned actuator network, and the cross-platform deployment across three primary and ten additional robots is impressive. The promise of open-sourced code, models, and data would enhance reproducibility. However, the two headline claims -- that in-air identification transfers to stance without randomization, and that the 32% CoT reduction is attributable to the proposed method -- are not yet established by the experiments as presented.","major_comments":[{"comment":"The central takeaway (Section 6) asserts that constant per-joint parameters {Ia, d, tau_f, q_tilde_b, Td} identified from fixed-base, zero-contact in-air data are sufficient to eliminate dynamics randomization in stance. Section 2.2 itself concedes that load-independent damping is 'common but imperfect' and that torque--current maps can be state dependent, and Appendix B (Fig. 23c) shows that the effective inertia at the HFE joint during horizontal on-ground motion reaches roughly 8--10 kg m^2, orders of magnitude above the in-air values. Section 4.1.2 further reports a full-robot Ia about four times the single-drive value, indicating that the fitted inertia absorbs configuration-dependent link and compensation effects. The manuscript offers no experiment that re-identifies the parameters under load, compares simulated stance torques against measured torques on a torque-instrumented platform, or otherwise tests whether the load-independence assumption holds during contact. Since PACE deliberately omits dynamics randomization as a safety net, this extrapolation is load-bearing and needs direct evidence rather than the single-sentence assertion 'In-air data suffices for contact tasks.'","section":"Section 5, Section 2.2, Appendix B"},{"comment":"The headline 32% Cost of Transport reduction compares ANYmal D (CoT 1.27) with ANYmal C (CoT 1.86) from the previously published reference [71], not with a same-hardware baseline. Table 3 shows that the two runs also differ in ambient temperature (26 deg C vs 10 deg C), and ANYmal C and ANYmal D may differ in actuators, firmware, mass, and controller generation. The entire reduction is attributed to PACE without controlling for these factors. To support the abstract's claim, the authors need a controlled comparison on the same robot unit, for example running the actuator-network baseline or a domain-randomized baseline on the same ANYmal D with the same reward and track protocol, or running PACE on ANYmal C, or otherwise demonstrating that hardware/firmware/temperature differences do not explain the CoT gap.","section":"Section 3.3.3, Table 3, Abstract"},{"comment":"The paper claims the physics-grounded energy reward contributes to the reported efficiency gains, but no ablation isolates the effect of the energy term. The reward scales and scheduling (c_v, c_e, c_ftd, c_c, lambda half-life) are free parameters listed in Eq. (20) and Table 2, and P_pot in Eq. (14) averages to zero over flat cyclic locomotion, so it is unclear how much of the improvement comes from the energetic reward versus the fitted simulator enabling different gaits. Since one of the four stated contributions is the energetic assessment, an ablation on at least one platform -- same fitted simulator with and without the energy term, or with a torque-squared proxy -- measured by battery CoT would quantify this contribution.","section":"Section 2.3.3, Table 4"}],"minor_comments":[{"comment":"The abstract states that all code, models, and datasets are publicly available, while the Data Accessibility Statement says the source code 'will be released soon' and the text elsewhere says 'will be released'; please align these statements.","section":"Abstract and Data Accessibility Statement"},{"comment":"The header for the armature row reads 'Armature I_a [10^-3 kg m^3]'; the unit should be kg m^2, not kg m^3.","section":"Table 5"},{"comment":"The text refers to 'current limits (Section 12)', which appears to be an incorrect internal reference; it should point to Figure 12 or Section 4.2.4 where the 32 A battery limit is described.","section":"Section 6 and Section 4.2.4"},{"comment":"In the discussion of knee extension, the text says 'Table 4.2.3'; this should be 'Table 4'.","section":"Section 5"},{"comment":"Equation (42) appears to have a typo: the inertia term should contain sin^2(q_k/2), not sin(q_k/2), based on the derivation from Eqs. (34)--(40).","section":"Appendix B, Eq. (42)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on prior work from the same group (refs 20, 32, 45, 71), which is scientifically legitimate but makes the need for a same-hardware baseline in the CoT comparison more acute. The central idea is promising and the empirical scope is strong, but the two headline claims require additional experiments or careful re-framing before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"PACE is a genuinely useful piece of engineering: a compact 4n+1 parameter set fitted from about 20 s of in-air encoder-only data, no torque sensors, no dynamics randomization, and it works across three quite different quadrupeds plus a long tail of other platforms. The bottom-up validation is the best part. Single-drive fits recover analytic inertias within a few percent across load cases and PD gains, full-robot in-air replays beat the actuator-network baseline on ANYmal, and the on-ground tracking is convincing. The energy reward built from PMSM losses plus the CoT decomposition into electronics, drives, and locomotion is a real step beyond torque-squared penalties.\n\nThe soft spots are where the abstract overreaches. The headline 32 percent CoT reduction is ANYmal D versus a previously published ANYmal C result, with different firmware, temperature, speed, and ambient conditions. That is not a controlled A/B comparison. The paper reports average speed 0.85 m/s vs 0.74 m/s, so part of the efficiency gain may simply be faster walking. Runs are single-shot per robot, so we do not know the variance. The deeper scientific question is the in-air-to-stance extrapolation: the fitted Ia, d, and tau_f are constant parameters identified under zero load, and the paper itself concedes that actual damping and torque-current maps are often state dependent. The full-robot fit already shows Ia absorbing a fourfold configuration effect, which tells you the parameter is not a pure motor constant. No experiment checks predicted torques against measured torques under stance loads, and the discussion asserts \"in-air data suffices\" without a mechanical argument. That does not sink the paper—the empirical transfer is real—but the central claim is stronger than the evidence. Also, the abstract says code is publicly available while the full text says it will be released; that inconsistency should be fixed.\n\nWho this is for: anyone doing legged sim-to-real will get value from it. It offers a practical, low-data alternative to actuator networks and heavy domain randomization, and the multi-level validation is a model of careful reporting. It deserves a serious referee. With a same-hardware A/B energy comparison, repeated runs, and a loaded-joint torque check, it would be close to a complete story.","headline":"Solid engineering with real hardware validation, but the headline energy claim is a cross-paper comparison and the in-air-to-stance parameter assumption is asserted, not tested.","tokens_in":27191,"tokens_out":2286,"would_cite":true,"duration_ms":23313,"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":"About 20 seconds of encoder-only in-air data are enough to train legged-robot controllers that transfer zero-shot, without dynamics randomization.","keywords":["legged robots","quadrupedal locomotion","sim-to-real transfer","reinforcement learning","energy efficiency","system identification","cost of transport","actuator dynamics"],"falsifier":"Fit the 4n+1 parameters from in-air chirps, then stand the robot on a force plate and inject a known PD torque perturbation during stance; if the simulated joint response diverges from the measured one as posture changes while the in-air fit is held fixed, the claim that in-air data suffices for contact tasks fails. A second decisive check is to train the PACE policy with no dynamics randomization and find any terrain or speed at which it fails while a randomization-trained policy succeeds.","tokens_in":26049,"feed_emoji":"🤖","tokens_out":11746,"duration_ms":96461,"temperature":0.7,"pith_summary":"The paper tries to establish that the simulation-to-reality gap for legged robots is mostly an actuator-dynamics gap, and that a deliberately small set of physically meaningful parameters can close it. Its central claim is that fitting four per-joint quantities plus one global delay, a total of $4n+1$ numbers, from about twenty seconds of in-air and encoder-only motion makes a simulator accurate enough that blind (proprioception-only) locomotion policies transfer to hardware in one shot, with no dynamics randomization. If true, this would make sim-to-real deployment cheaper and more accessible: no torque sensors, minutes of data, or large hand-tuned reward stacks. The paper also reports that the fitted simulator enables energy-efficient gaits, with a 32 percent reduction in ANYmal's full cost of transport to 1.27.","feed_headline":"20 seconds of in-air data send legged robot controllers to hardware","feed_subtitle":"A 4n+1 parameter fit removes dynamics randomization and cuts ANYmal's cost of transport by 32 percent.","key_machinery":"The engine of the argument is the per-joint closed-loop model in Eq. (6): each actuator is treated as a constant-inertia, constant-damping second-order system driven by saturated PD torque, with an added joint-position bias and a global command delay. The parameter vector $p=(I_a, d, \\tau_f, \\tilde{q}_b, T_d) \\in \\mathbb{R}^{4n+1}$ is fitted by minimizing the time-averaged squared joint-position error between real and simulated chirp replays, using CMA-ES over thousands of parallel environments. Once fitted, this model carries the sim-to-real transfer: the policy is trained in the fitted simulator with a compact reward whose energetic term is the PMSM electrical-plus-mechanical loss model, and it is deployed without dynamics randomization.","core_discovery":"The paper's central discovery is a bottom-up identification recipe, called PACE: suspend the robot, drive all joints with chirp signals, replay the commands in a massively parallel simulator, and use evolutionary search to fit per-joint armature inertia, viscous damping, Coulomb friction, and joint bias, together with one global command delay. The fitted simulator reproduces in-air joint trajectories nearly exactly, generalizes to unseen PD gains and trajectories, and then serves as the training environment for a blind policy with a four-term reward built around a physics-grounded PMSM loss model. Deployed zero-shot on hardware, the policy walks, runs, balances on two legs, and climbs stairs; on ANYmal it lowers the cost of transport from 1.86 to 1.27 and extends a full-charge run from 2.64 km to 4.12 km. The same parameterization is reported to fit both open and closed platforms.","pith_inferences":["The paper leaves implicit that the constant-parameter model may saturate at higher step frequencies or on strongly varying terrain, since the effective inertia at each joint changes with stance posture; a testable extension is to add a posture-dependent inertia term and see whether it widens the no-randomization envelope.","The identification recipe is not obviously limited to quadrupeds: any articulated robot with PMSM drives and encoder-only joint sensing could in principle be fit from fixed-base chirps, so the claim suggests a general sim-to-real shortcut for arms and humanoids.","A head-to-head study of the same policy with and without dynamics randomization on identical terrain could separate how much of the 32 percent cost-of-transport gain comes from removing randomization versus from the energy reward.","Temperature and wear will shift the fitted parameters over time, so a lightweight online update of the global scale of the inertia and damping terms is a natural follow-up that the paper names as future work."],"forward_implications":["Robots without joint-torque sensors can obtain deployable locomotion policies from about 20 seconds of encoder-only, in-air data, removing a practical barrier for small or closed platforms.","Because dynamics randomization is no longer needed, policies trained in the fitted simulator can exploit the machine's actual dynamics, producing the reported straighter-knee gaits and the ANYmal cost-of-transport drop to 1.27.","The same 4n+1 parameterization is reported to transfer across platforms, from a 52 kg quadruped to a 4 kg largely 3D-printed one, suggesting the identification is not tied to one drive design.","With the model fitted, the remaining ceiling on running speed is hardware, the battery current limit, rather than model error, which shifts the practical bottleneck to power electronics and sensing."],"supporting_citations":[{"why":"Supplies the torque-sensor actuator-network baseline whose in-air fidelity and ANYmal locomotion performance PACE is compared against.","marker":"(20)"},{"why":"Contributes the velocity-tracking reward form and teacher-student-style training ideas that the compact reward adapts.","marker":"(32)"},{"why":"Provides the asymmetric actor-critic setup that lets the policy train on proprioception while the critic sees privileged state.","marker":"(44)"},{"why":"Supplies the massively parallel GPU training paradigm that makes the evolutionary identification and policy learning practical.","marker":"(45)"},{"why":"Provides the CMA-ES optimizer used in the evolutionary parameter identification.","marker":"(57)"},{"why":"Gives the actuator-aware electrical and mechanical loss formulations on which the PMSM energy model is built.","marker":"(52, 53)"},{"why":"Names the open hardware platform behind TYTAN, used for the single-drive and full-robot identification studies.","marker":"(61)"},{"why":"Describes ANYmal, the closed-source platform used for the main sim-to-real and cost-of-transport comparisons.","marker":"(62)"},{"why":"Defines the 400 m track protocol and the baseline cost-of-transport figures against which the 32 percent improvement is measured.","marker":"(71)"},{"why":"Supplies the biological dog cost-of-transport comparison used to contextualize the measured robot efficiency.","marker":"(72)"}],"fun_headline_variants":["Sim-to-real robot control: 20s of in-air data, no randomization","PACE: 4n+1 parameters fit legged robots, cut energy 32%","No dynamics randomization: PACE transfers policies to 13 robots","In-air chirp fit trains blind loco policies, 32% lower cost","20s in-air data beats dynamics randomization for legged robots"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a constant inertia and damping measured while the legs swing freely in air remain the right joint dynamics when the feet are on the ground, even though the effective inertia at each joint is then dominated by the body mass and varies with posture.","fun_headline_variants_meta":{"raw":{"variants":["Sim-to-real robot control: 20s of in-air data, no randomization","PACE: 4n+1 parameters fit legged robots, cut energy 32%","No dynamics randomization: PACE transfers policies to 13 robots","In-air chirp fit trains blind loco policies, 32% lower cost","20s in-air data beats dynamics randomization for legged robots"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00029,"raw_usage":{"total_tokens":1682,"prompt_tokens":918,"completion_tokens":764,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":663}},"tokens_in":534,"tokens_out":764,"duration_ms":6160,"temperature":1.0,"reasoning_tokens":663,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:17:50.967034+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the 4n+1 parameters from in-air chirps, then stand the robot on a force plate and inject a known PD torque perturbation during stance; if the simulated joint response diverges from the measured one as posture changes while the in-air fit is held fixed, the claim that in-air data suffices for contact tasks fails. A second decisive check is to train the PACE policy with no dynamics randomization and find any terrain or speed at which it fails while a randomization-trained policy succeeds.","supporting_citations":[],"review_version":2}