{"id":"b8b7c741-e9f8-4951-afff-cfc45d7faf69","arxiv_id":"2506.23114","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A reinforcement-learning quadruped controller with phase-based foot-impact penalties and a quiet factor achieves about 8 dBA lower walking noise than baseline controllers indoors.","lead":"This paper trains a quadruped robot controller that penalizes fast foot impacts and adds a user-tunable quiet factor to cut walking noise by about 8 dBA indoors. The result matters because hospitals, offices, and homes may accept legged robots only if their footsteps stop being disruptive.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 8 dBA headline may be a speed effect: MUTE β=1 actively trades speed for quietness (Eq. 6, Fig. 6), while Table II reports only commanded 0.5 m/s without actual-speed data.","rationale":"I read the paper as claiming that the control policy itself reduces impact noise, not merely that asking the robot to move more slowly makes it quieter. That requires actual-speed-matched comparison. The paper's own design makes the confound live: β=1 deprioritizes velocity tracking via Eq. (6), and Fig. 6 documents the speed-error cost. Table II reports only commanded speed, so the 8 dBA average is not yet matched-speed evidence. I would not reject the paper—real-robot, multi-surface measurements are a genuine strength, and Fig. 4 attempts an actual-velocity comparison on wood—but the headline number should be conditional on releasing per-trial actual speeds or replotting Table II that way. I also note Eq. (10) labels an unweighted SPL formula as dBA, which is a smaller accuracy issue. The reader's weakest assumption was about foot-impact dominance and microphone placement; my concern overlaps with the speed ambiguity in their rationale, hence partial agreement. The verdict stays conditional.","tokens_in":10967,"tokens_out":7895,"duration_ms":85807,"concrete_test":"Obtain per-trial EKF velocity logs for Table II. Compute the mean and standard deviation of actual speed for each method-surface cell, and compare MUTE β=1 with DreamWaQ and MPC under the 0.5 m/s command. If the mean actual speed differs by more than roughly 0.05–0.1 m/s, rerun (or re-extract from Fig. 4) a matched-speed comparison at equal actual speed and recompute the MNL/PNL gap. A surviving ~8 dBA gap at matched speed would confirm the claim; a shrinking gap would require restating the headline as a speed–noise trade-off.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central claim is ~8 dBA reduction at equivalent locomotion speeds. The strongest source of the reduction, MUTE with β=1, is explicitly designed to sacrifice speed: Eq. (6) scales the velocity-tracking reward by (1−kβ)=0.8 when β=1, and Fig. 6 shows velocity-tracking squared error increasing toward β=1. Table II is the source of the headline average (64.80 vs DreamWaQ 72.58 dBA) and is run at a common 0.5 m/s command; it reports no per-method actual speeds. Contribution 3 asserts 'equivalent locomotion speeds,' but equivalence is only demonstrated later on wood in Fig. 4, where the x-axis is EKF actual velocity—not for the table that produces the 8 dBA average. If MUTE β=1 actually walks slower at the same command, the margin could be a speed trade-off rather than a quieter gait. This is testable from existing logs; if the concern lands, the headline should be re-reported as a tunable speed–noise trade-off with matched-speed data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MUTE, a learned quadruped locomotion controller that combines LSTM-based phase estimation, a phase-dependent foot-velocity penalty in the reward, and a scalar 'quiet factor' β that trades speed against noise. Policies are trained in simulation with PPO and deployed on a Unitree Go1 EDU without further tuning. Real-world experiments measure mean and peak A-weighted sound levels on wood, carpet, and tiles at a 0.5 m/s command, and also include a speed sweep and a long-distance mixed-surface walk. The central claim is that MUTE achieves an average noise reduction of approximately 8 dBA compared with DreamWaQ and built-in MPC baselines at equivalent locomotion speeds.","tokens_in":11184,"tokens_out":7198,"duration_ms":68414,"significance":"If the quantitative claim is substantiated, this is a valuable contribution to an underexplored area: noise-aware locomotion control for legged robots. The paper's strengths are its real-robot evaluation on three surfaces, comparison against two baselines, a tunable trade-off parameter (β), and a long-duration deployment test. The evaluation is not circular: the acoustic measurements are external outcomes and are not used to fit the policy. However, the headline result currently rests on commanded rather than measured speed, the measurement chain is under-specified, and no statistical inference is provided. These issues are fixable but need to be addressed before the claim of an 8 dBA reduction at equivalent speeds can be accepted.","major_comments":[{"comment":"The headline 'equivalent locomotion speeds' is not supported by the reported data. Table II reports only the commanded 0.5 m/s and no measured actual speed per method, while Eq. (6) with k=0.2 explicitly scales the velocity-tracking reward by (1−kβ)=0.8 when β=1, and Fig. 6 shows velocity-tracking squared error increasing with β. The average 7.78 dBA reduction of MUTE β=1 over DreamWaQ (64.80 vs 72.58 dBA) could therefore partly reflect slower actual walking. Fig. 4 uses EKF-based actual speed, but it is a separate experiment on wood and is not linked to the Table II trials. Please report per-method actual speeds (or matched-speed runs) for the Table II conditions, or re-scope the claim as a tunable speed–noise trade-off.","section":"§III-D, Table II; Contribution 3"},{"comment":"The numerical support for the trade-off is incomplete. The velocity tracking squared error at β=1 is about 0.012 m²/s², corresponding to an RMS error of roughly 0.11 m/s on a 0.8 m/s command, which is not self-evidently a 'minor compromise.' Moreover, no statistical tests or confidence intervals are provided for any MNL/PNL difference; the five-trial mean ± std in Table II does not establish that the 8 dBA gap is significant or that the ordering across methods is reliable. Please add paired tests or bootstrap intervals and report actual speed or velocity error in physical units.","section":"§III-E, Fig. 6; §III-D"},{"comment":"The measurement chain is under-specified for an absolute dBA claim. No microphone model, calibration procedure, frequency-weighting implementation, or ambient-noise levels during trials are given; the sensor is mounted on the robot and samples at 20 Hz, which is slow for capturing impact transients, so the PNL values may be underestimated. The assertion in §I-C that motor/fan noise is 'relatively minor' is not backed by measurements, and a robot-mounted sensor may pick up such contributions. Raw recordings or a calibration check are needed to rule out systematic bias in the 8 dBA number.","section":"§III-A, §III-B, Eq. (10); §I-C"},{"comment":"The causal mechanism is assumed rather than demonstrated. The reward in Eq. (9) penalizes foot drop velocity, but the paper reports no real-world measurements of foot contact velocity, ground reaction forces, or high-speed motion capture to show that MUTE actually reduces impact velocity. Because β=1 also scales down the speed reward, the relative contribution of the phase reward versus the speed trade-off to the measured noise reduction is not identifiable from the current data. An ablation that fixes actual speed (or reports contact velocity) would make the mechanism claim load-bearing.","section":"§II, Eq. (9); §III-D"}],"minor_comments":[{"comment":"The text states that on surfaces 'such as wood and tiles, MUTE exhibits the most significant reductions,' but Table II shows the largest absolute reduction versus DreamWaQ is on carpet (9.42 dBA), not wood (7.73 dBA) or tiles (6.19 dBA).","section":"§III-D"},{"comment":"Equation (10) defines unweighted SPL but labels it dBA; A-weighting is a frequency-domain correction and should be specified separately, including the filter or meter standard used.","section":"Eq. (10)"},{"comment":"There are typos ('shows a increase') and the phrase 'suggesting a minor compromise' should be replaced with the reported velocity error value and its physical interpretation.","section":"§III-E, Fig. 6"},{"comment":"The statement that 68.25 dBA is below a '70 dBA threshold considered safe for human hearing' and that attenuation makes human-perceived noise lower assumes a particular human distance; please cite a standard and state the assumed distance.","section":"§III-F"},{"comment":"The paper does not state whether the five trials per surface were conducted on different days, whether ambient noise was monitored and subtracted, or whether the sensor was calibrated before each session; please clarify.","section":"§III-A"}],"recommendation":"major_revision","confidential_remarks":"The speed-equivalence issue is testable from existing logs; if actual speeds are available, a simple table would resolve it. If not, the authors should re-scope the abstract and conclusion. The manuscript would also benefit from releasing code and raw acoustic data to support the 8 dBA claim. The paper is within scope for a robotics venue, but the quantitative claim needs stronger empirical support."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper reports a control-only framework that cuts measured dBA by roughly 8 on a Unitree Go1 across wood, carpet, and tile, using a phase-based foot-velocity penalty and a \"quiet factor\" beta that trades speed for noise. That is a genuinely useful engineering result—no hardware changes, real robot, multiple surfaces, and trial-level statistics. The combination of a learned phase estimator, foot-velocity rewards, and a scalar trade-off is not something I've seen in the legged-locomotion noise papers.\n\nThe soft spot is the headline. The 8 dBA average in Table II is at a common 0.5 m/s command, not at matched actual speed. Eq. (6) deliberately scales the velocity-tracking reward down by (1-k*beta)=0.8 at beta=1, and Fig. 6 confirms velocity tracking error climbs with beta. So part of the measured gap likely comes from MUTE beta=1 simply walking slower. The authors do include a matched-actual-speed plot on wood (Fig. 4), which is the right control, but that figure doesn't cover all three surfaces and doesn't produce the 8 dBA average. The contribution statement \"equivalent locomotion speeds\" is therefore stronger than the evidence. The fix is straightforward: report actual speeds from the logs for every trial, or re-run with speed matched, and caveat the averaged claim or re-report as a speed-noise trade-off.\n\nOther issues are minor. No code or raw recordings, so it's not independently checkable. The microphone at 30 cm on the trunk captures the near field; the perceptual benefit at ear height is inferred, not measured. No significance tests on the dBA differences, though the std devs are reported. These are all revisable in review.\n\nOverall the core mechanism is plausible and the direction of the result is credible. This is incremental—it doesn't change how legged locomotion is understood—but it's a solid application paper that could be useful to anyone deploying quadrupeds indoors. I'd send it to peer review and ask for the matched-speed data plus a clearer statement of what the 8 dBA does and does not mean.","headline":"A plausible control-only 8 dBA noise reduction on a real quadruped, but the headline is partly confounded by the quiet factor's speed trade-off and needs matched-actual-speed data.","tokens_in":11704,"tokens_out":3075,"would_cite":true,"duration_ms":32545,"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":"A new control policy that slows feet just before ground contact cuts a quadruped's walking noise by about 8 dBA on indoor surfaces, the paper reports.","keywords":["quadruped robots","acoustic noise reduction","quiet locomotion","foot-ground impact","reinforcement learning","gait phase estimation","indoor service robotics","quiet factor"],"falsifier":"Measure the same robot with an independent microphone at a standing person's ear height while it walks on a hard floor, and compare MUTE to the baseline at matched speed; if the ear-height reduction approaches zero even though the trunk-mounted reading still shows roughly 8 dBA, the claimed human-relevant benefit would not survive. A second check would use motion capture of the feet: if two policies with measured impact velocities that are equal still differ by several dBA, then impact velocity alone does not explain the noise difference.","tokens_in":10758,"feed_emoji":"🤫","tokens_out":7283,"duration_ms":70080,"temperature":0.7,"pith_summary":"The paper tries to show that quadruped robots can be made substantially quieter indoors by changing only the control policy, not the hardware. Its central claim is that foot-ground impacts are the dominant noise source and that a learned controller can be trained to let each foot down softly. The proposed method, MUTE, estimates the phase of each leg's gait cycle and adds a reward that penalizes high foot velocity just before contact, with a dialable quiet factor that trades speed for silence. In experiments on wood, carpet, and tile, MUTE with the quiet factor at its maximum measured about 8 dBA less mean noise than the comparison controllers at the same locomotion speed, roughly a 2.5-fold drop in sound pressure. If this stands, quiet operation in hospitals, offices, and homes would not require new motors or padding; it would be a software property of the walk.","feed_headline":"Quiet quadruped: controller cuts walking noise by 8 dBA","feed_subtitle":"Phase-aware reward slows feet before impact, trimming footstep noise on wood, carpet, and tile.","key_machinery":"The load-bearing mechanism is a phase-gated foot-velocity penalty. A network estimates each leg's phase $\\hat{\\varphi}_t$ from the robot's proprioceptive history, telling the policy when a foot is about to strike. The reward term $r_\\varphi = w_d \\sum_i e^{\\hat{\\varphi}_i} v_{f_i,\\text{drop}}^2 + w_r \\sum_i e^{-\\hat{\\varphi}_i} v_{f_i,\\text{raise}}^2$ penalizes downward foot velocity ever more strongly as the phase nears 1 through its negatively weighted first sum, and the quiet factor $\\beta$ in the total reward $r = \\beta r_\\varphi + (1-k\\beta) r_v + r_o$ scales that penalty against the task reward $r_v$, making the noise-speed trade-off explicit and controllable.","core_discovery":"The central discovery, as the authors state it, is that acoustic noise during quadruped locomotion is a controllable output of the gait. Sound amplitude is modeled as proportional to the velocity of the foot at the moment of impact, so the objective becomes shaping the foot trajectory to arrive slowly. MUTE learns a policy in simulation using an estimated phase variable $\\hat{\\varphi}_t$ for each leg; as the phase approaches 1 (contact), a phase reward term imposes a growing penalty on squared drop-foot velocity, while a small reward encourages consistent foot lift. A scalar quiet factor $\\beta \\in [0,1]$ blends this phase reward with the usual velocity-tracking reward, letting the operator choose between speed and quiet. On wood, carpet, and tile, the method reports an average reduction of approximately 8 dBA compared with the baselines at matched speed, which corresponds to about 2.5 times lower sound pressure, and the reduction persists even when $\\beta = 0$.","pith_inferences":["A testable extension the paper does not explore is closed-loop $\\beta$ control, where a higher-level planner raises the quiet factor in a hospital corridor at night and lowers it outdoors, treating the quiet factor as a command input rather than a fixed setting.","The paper leaves the human-relevance question open: its microphone rides on the robot near the feet, and the claim that ear-height noise is lower relies on spherical spreading, which reflective floors and A-weighting could weaken in practice.","The same phase-gated contact-velocity reward could transfer to other impact-noise systems, such as bipedal walkers, drone landing controllers, or manipulators setting objects down, whenever contact velocity is a controllable output.","The 8 dBA headline is an average over surfaces and baselines; the paper's surface-by-surface table shows variation of several dBA, so a real deployment would likely need surface-specific tuning to guarantee the advertised reduction."],"forward_implications":["At matched commanded speed, MUTE lowers mean and peak noise on wood, carpet, and tile, with the largest measured reductions on hard surfaces.","Dialing the quiet factor from 0 to 1 monotonically reduces noise while raising velocity-tracking error only slightly, giving a field-adjustable noise-speed knob.","Because MUTE is quieter than the baselines even at $\\beta = 0$, the phase-aware reward appears to improve gait smoothness beyond explicitly trading away speed.","On a 91.7-meter mixed-surface indoor walk, the measured mean noise of 68.25 dBA stays below the 70 dBA hearing-safety threshold, and the paper argues the human-perceived level at ear height would be lower still."],"supporting_citations":[{"why":"This reference identifies foot-ground impact as the primary noise source in legged robots, motivating the paper's control-level approach.","marker":"[17]"},{"why":"This reference supplies the physical model that sound amplitude is proportional to the square root of impact energy, i.e., proportional to foot velocity at contact.","marker":"[19]"},{"why":"This reference is the learned locomotion baseline whose noise performance MUTE is compared against, and it is also cited for velocity estimation from observation history.","marker":"[21]"},{"why":"This reference provides the asymmetric actor-critic training architecture that lets the critic use privileged state while the actor uses only observations.","marker":"[25]"},{"why":"This reference is the proximal policy optimization algorithm used to train the policy and value networks in simulation.","marker":"[26]"},{"why":"This reference supplies the massively parallel simulation approach that makes training thousands of agents in about eight hours practical.","marker":"[27]"},{"why":"This reference defines the A-weighting standard used to justify reporting sound levels in dBA for human auditory relevance.","marker":"[29]"}],"fun_headline_variants":["Quadruped robot whispers: 8 dBA quieter steps","Slower footfalls cut robot noise by 8 dBA","Mute quadruped: AI-driven gait cuts noise 8 dBA","8 dBA quieter: control strategy silences robot steps","Gait control dampens footsteps: 8 dBA quieter indoors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reduction depends on foot-ground impact velocity being the dominant and controllable source of the robot's noise, and on the 30-cm-above-ground, robot-mounted microphone capturing the sound field that people actually perceive.","fun_headline_variants_meta":{"raw":{"variants":["Quadruped robot whispers: 8 dBA quieter steps","Slower footfalls cut robot noise by 8 dBA","Mute quadruped: AI-driven gait cuts noise 8 dBA","8 dBA quieter: control strategy silences robot steps","Gait control dampens footsteps: 8 dBA quieter indoors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000572,"raw_usage":{"total_tokens":2672,"prompt_tokens":882,"completion_tokens":1790,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":1702}},"tokens_in":498,"tokens_out":1790,"duration_ms":12219,"temperature":1.0,"reasoning_tokens":1702,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:48:46.248631+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the same robot with an independent microphone at a standing person's ear height while it walks on a hard floor, and compare MUTE to the baseline at matched speed; if the ear-height reduction approaches zero even though the trunk-mounted reading still shows roughly 8 dBA, the claimed human-relevant benefit would not survive. A second check would use motion capture of the feet: if two policies with measured impact velocities that are equal still differ by several dBA, then impact velocity alone does not explain the noise difference.","supporting_citations":[{"cited_title":"Acoustics based terrain classification for legged robots,","cited_arxiv_id":null,"evidence_quote":"This reference identifies foot-ground impact as the primary noise source in legged robots, motivating the paper's control-level approach."},{"cited_title":"The sounds of physical shapes,","cited_arxiv_id":null,"evidence_quote":"This reference supplies the physical model that sound amplitude is proportional to the square root of impact energy, i.e., proportional to foot velocity at contact."},{"cited_title":"Electroacoustics—sound level meters—part 1: Specifications (iec 61672-1),","cited_arxiv_id":null,"evidence_quote":"This reference defines the A-weighting standard used to justify reporting sound levels in dBA for human auditory relevance."}],"review_version":1}