{"id":"83902294-4d86-4524-aed8-55b776a0087f","arxiv_id":"2505.11808","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A custom NMPC/WBIC controller slowed the gait and softened foot contacts of a Unitree Go1 quadruped, reducing walking noise by roughly 10 dB and earning better user acceptance than the default gait.","lead":"Researchers built a quieter walking controller for a four-legged guide dog robot and tested it with four blind or low-vision guide dog handlers. The controller cut walking noise by about 10 decibels compared with the robot's default gait, and the handlers preferred it, especially on stairs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 10 dB noise-reduction claim depends on a single uncalibrated smartphone recording at the experimenter's ear, with no trial repeats or error bars; this should be verified with a calibrated, repeated acoustic protocol before the headline number is treated as established.","rationale":"I agree with the reader's weakest_assumption. The central claim is quantitative, and the only objective support for the 10 dB number is a single uncalibrated smartphone recording procedure with no repeats or error bars. This is an evidence gap in a specific measurement, not an internal contradiction or a disagreement with field consensus. The user study and simulation results support the direction and utility of the controller but not the exact acoustic magnitude. Because the concern is addressable and does not invalidate the engineering contribution, CONDITIONAL remains the appropriate verdict. Since the reader's verdict is already CONDITIONAL, my stress test does not change it.","tokens_in":20736,"tokens_out":4314,"duration_ms":48261,"concrete_test":"Run a controlled acoustic evaluation in the same laboratory: use a calibrated Class 1 sound level meter (or a binaural head-and-torso simulator) positioned at the handler's ear height and lateral offset, with a fixed robot path and at least 10 straight-line passes per controller per speed (0.6, 0.8, 1.0, and 1.2 m/s). Report A-weighted equivalent (LAeq) and maximum (LAFmax) sound levels with standard deviations and confidence intervals. If the mean difference between the two controllers is not approximately 10 dB at the relevant speeds, or if the 95% confidence interval includes 6 dB, the 'halving perceived noise' claim should be downgraded to a qualitative improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim — that the controller 'reduces the noise of the robot's walking by 10 dB compared to the Unitree Go1 robot's default controller, effectively halving the perceived noise level' (Conclusion) — rests on the acoustic test in Section V-B. A researcher walked beside the robot holding a smartphone at ear level, and the robot passed through a 2 m recording zone; Fig. 4(c) plots 'average decibel' for the two controllers. No calibration of the phone or its app is reported, no number of repeated trials is given, no error bars or confidence intervals appear, and the recording geometry (distance, orientation, walking-induced phone motion, room acoustics) is uncontrolled. The reported difference is large, and the direction of the effect is corroborated by all four BLV participants' comments, so the qualitative conclusion is plausible. But the exact '10 dB / half loudness' figure is load-bearing for the paper's headline, and the measurement as described cannot support a quantitative claim of that precision. The user study cannot rescue the number: n=4 Likert ratings establish a perceived difference, not a 10 dB acoustic magnitude. The authors' Section VIII limitations do not flag this measurement issue. A calibrated, repeated acoustic comparison is therefore the key missing piece.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a locomotion controller for a Unitree Go1 quadruped intended as a guide dog robot for blind and low-vision (BLV) individuals. The controller replaces a convex MPC with a real-time SQP-based NMPC using SO(3) orientation dynamics, pairs it with WBIC, adds perception-based stair climbing, and is evaluated in simulation and on hardware. The authors report approximately a 10 dB noise reduction relative to the default controller, improved balance under disturbances, and a mixed-methods user study with four BLV guide dog handlers indicating preference for the new controller in flat walking and stair climbing. The paper claims these results validate a human-centered development cycle from stakeholder interviews to system design to user evaluation.","tokens_in":20955,"tokens_out":5155,"duration_ms":52896,"significance":"If the quantitative claims are supported, the paper makes a valuable contribution: it directly addresses an understudied requirement for guide dog robots (acoustic and physical disturbance), provides a concrete NMPC+WBIC formulation that maintains balance at slow gait frequencies, integrates terrain perception for stairs, and reports one of the first user studies of BLV individuals interacting with a quadruped robot during stair climbing. Strengths include real-hardware evaluation, comparison against multiple baseline controllers, a clear description of the design choices grounded in stakeholder feedback, and an explicit acknowledgment of small sample size in the user study. The central '10 dB / half perceived noise' result, however, rests on an acoustic measurement protocol that currently cannot support the precision of the headline claim.","major_comments":[{"comment":"The central quantitative claim of a ~10 dB noise reduction (restated in the Abstract and Conclusion) rests on a single acoustic protocol in which one researcher held a smartphone at ear level while walking beside the robot, with no calibration of the phone or microphone, no stated number of repeated trials, no error bars or confidence intervals, and uncontrolled recording geometry, room acoustics, and background noise. Section VIII does not flag this measurement limitation. The authors should repeat the measurement with a calibrated sound-level meter or a fixed, documented microphone geometry, multiple trials per speed and per controller, and report the mean, spread, and ambient-noise floor; until then, the headline '10 dB / half perceived noise' should either be supported by such data or downgraded to a qualitative observation.","section":"Section V-B, Fig. 4(c), and Conclusion"},{"comment":"The additional acoustic comparisons against cMPC, NL MPC, the RL method of [25], the Anymal-based result in [32], and the wheeled systems in [26] inherit the same uncalibrated protocol and mix different robots, measurement conditions, and published setups. Cross-platform claims such as '50 dB is even lower than the noise level of wheeled systems tested in [26]' are only meaningful under a shared protocol; as written, these comparisons should be presented as indicative, with the differences in measurement conditions stated explicitly.","section":"Section V-B, Figs. 5 and 6(d)"},{"comment":"With n=4 and no inferential statistics, variance information, or individual-level plots, the wording in Section VI-E and the Conclusion ('superior acceptance,' 'lower workload,' 'higher usability') goes beyond what the data can support. The qualitative participant quotes support a directional effect, and Section VIII appropriately mentions the small sample, but the figures need per-participant values or error bars, and the claims should be phrased as observed trends in this small sample rather than as general conclusions.","section":"Section VI, Figs. 8 and 9"},{"comment":"The stair-climbing evaluation uses a custom staircase with a 13 cm rise and 60 cm tread, which the authors note is shorter and wider than standard stairs because of the Go1 robot's leg kinematics. This means the user-study conclusions about 'comfortable stair climbing' (RQ2) cannot yet generalize to standard stair geometry; the paper should state this scope limitation in the conclusions and temper the corresponding claims accordingly.","section":"Section V-B and Section VI-B"}],"minor_comments":[{"comment":"The caption uses 'cMPC (Original)' and 'cMPC (OSQP)' inconsistently with the legend labels in the plots, making it difficult for the reader to map each trace to a controller.","section":"Fig. 3 caption"},{"comment":"The y-axis is labeled 'Angular Yaw Velocity (rad/s)', while the text describes 'maximum stable yaw velocity'; please clarify whether the plotted value is a magnitude or signed value and state the unit consistently.","section":"Fig. 6(c)"},{"comment":"Reference [75] appears to contain a typo in the title ('nmiscar' should likely be 'nonlinear'); please verify and correct.","section":"References"},{"comment":"The hardware section says 'reduces noise by nearly 10 dB' while the Conclusion states a definitive 'reduces the noise ... by 10 dB'; the wording should be unified once the measurement uncertainty is resolved.","section":"Section V-B and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a robotics venue and the technical core is worth publishing after the acoustic measurement is strengthened or the corresponding claims are scaled back. The 10 dB headline and the cross-platform noise comparisons are the main points that need attention; the user study limitations are partly acknowledged but the framing should be more cautious."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you care about assistive legged robots or the gap between claimed and measured acoustic performance. The paper is a genuine engineering effort: the authors build on NMPC+WBIC with an SO(3) formulation and real-time SQP, slow the stepping frequency, lengthen the downward swing, and tune gains to reduce mechanical noise. They also add perception-based stair climbing with a pause-on-first-step behavior that mimics trained guide dogs, and they test robustness on slippery surfaces and uneven terrain. The human-centered loop – interviews, observation, then a user study with four experienced BLU guide-dog handlers – is a real strength and, as far as I know, the first stair-climbing evaluation of a quadruped guide robot with blind users. That part deserves credit.\n\nThe soft spots are exactly where the reader puts them. The 10 dB claim rests on a single uncalibrated smartphone held at the experimenter's ear, with no trial counts, calibration, or error bars. That is not enough to support a precise quantitative headline. I agree with the stress-test note: this is load-bearing because the paper's own summary says it \"effectively halves\" perceived loudness. The right fix is a calibrated sound-level meter, several repeated passes at each speed, fixed geometry, and reported variance. The user study has n=4 and no inferential statistics, but the authors acknowledge that in Section VIII, and the qualitative agreement across all four participants is still informative. The staircase is non-standard (13 cm rise, 60 cm tread), which is a minor limitation, and no code or data are released, which is a shame for reproducibility.\n\nOne thing the stress-test note gets right: Section VIII lists several limitations but does not flag the acoustic measurement methodology itself. That omission should be corrected.\n\nBottom line: I would send this out for peer review rather than desk-reject it. The controller is thoughtfully constructed, the user-centered process is exemplary, and the qualitative findings are valuable. The 10 dB number needs verification, not dismantling. I would cite it for the controller design and the user-study methodology, but not for the decibel figure until it is re-measured properly. Bring it to reading group if you want a conversation about what counts as evidence in assistive robotics.","headline":"A solid engineering contribution to guide-dog robots whose 10 dB noise-reduction headline outruns the acoustic evidence; the qualitative user study and controller design are the real strengths.","tokens_in":21554,"tokens_out":1781,"would_cite":true,"duration_ms":21141,"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":"A quadruped guide-dog robot can walk at a natural human pace while cutting its noise by about 10 dB — half the perceived loudness of the default controller — without losing balance, and four guide-dog handlers preferred it.","keywords":["guide dog robot","quadruped locomotion","noise suppression","model predictive control","blind and low-vision users","stair climbing","assistive mobility","user-centered design"],"falsifier":"Re-measure both controllers with a fixed, calibrated microphone at the handler's ear height for many passes at each speed in the same room; if the mean difference is not close to 10 dB or the error bars overlap, the headline noise claim is not supported. Equally, a listening test in which blindfolded participants must detect a quiet sound source (such as a recorded car or a wall echo) while the robot walks would show whether the quieter gait restores the acoustic awareness the paper says it preserves.","tokens_in":20500,"feed_emoji":"🦮","tokens_out":7867,"duration_ms":76341,"temperature":0.7,"pith_summary":"Most quadruped guide-dog research treats navigation as the hard problem; this paper argues that the way the robot walks is the gate. Off-the-shelf trot controllers are noisy and jerky enough to mask the environmental sounds blind and low-vision users navigate by, so the authors build a controller that steps slowly, lands softly, and keeps balance while moving at the fast walking speeds experienced handlers actually use. They report that it cuts the Unitree Go1's walking noise by about 10 dB — half the perceived loudness of the default controller — and that four guide-dog handlers recognized the difference, rated it more compliant and satisfying, and found stair climbing similar to working with a dog. The paper's contribution is a complete human-centered loop: interviews and observations set the target, a nonlinear MPC controller meets it, and user trials check the result.","feed_headline":"Guide-dog robot cuts walking noise by 10 decibels","feed_subtitle":"Four guide-dog handlers preferred the quieter, smoother gait, and stair climbs felt like working with a dog.","key_machinery":"The carrying object is the controller stack: a real-time SQP-based nonlinear MPC paired with whole-body impulse control. The NMPC keeps the robot's full orientation dynamics through an SO(3) representation instead of simplified Euler angles, and solves each step as one SQP iteration with a warm start, reaching 400–500 Hz updates; this lets the robot use a slow trot (swing times around 0.25–0.3 s) while staying balanced under pulls and impacts. The companion whole-body impulse controller tracks the MPC's reaction forces and shapes the swing-leg trajectory, with a low derivative gain ($K_d = 0.5$) to avoid motor whine and a touchdown phase that takes 65% of the swing time to lower the foot. For stairs, a depth camera feeds a 2.5D height map that adjusts landing height and location, and the controller pauses when the front feet contact the first step. The mechanism's work is to decouple stepping rate from walking speed: fewer, longer, softer steps produce the measured 10 dB noise reduction while the body still advances at human pace.","core_discovery":"The central claim is that a quadruped can serve as a guide dog only if its locomotion is quiet and smooth enough to preserve the user's acoustic awareness, and that this is achievable without sacrificing speed or stability. The paper's controller replaces the default convex MPC with a nonlinear MPC that keeps full SO(3) orientation dynamics, solved by real-time SQP at 400–500 Hz, and pairs it with whole-body impulse control; slower swing times, a low derivative gain, and a swing trajectory that spends 65% of its time lowering the foot make touchdown gentle. In hardware tests on the Unitree Go1 the robot walked at 0.6–1.2 m/s with roughly 10 dB less noise than the default controller, kept its body velocity and orientation steadier, stayed balanced under a 25 N handler pull and a 100 N impulse, and traversed slippery, uneven, and cluttered surfaces. With a depth camera and a 2.5D height map, the same controller climbs stairs, pausing with front feet on the first step to cue the handler exactly as a trained guide dog does. In the user study, all four blind and low-vision handlers reported lower noise, and stair climbing drew workload ratings comparable to their dogs, with high usability scores.","pith_inferences":["An implication the authors leave implicit is that the same slow-stepping controller recipe could make other legged platforms, including humanoids and delivery robots, acceptable in human environments where noise is the limiting factor.","The paper does not test whether the 10 dB difference actually restores a user's ability to detect a specific environmental sound, such as an approaching car; a perceptual shadowing test would convert the acoustic claim into a safety metric.","Since the height map is the only terrain cue, a direct comparison of the same controller with and without the perception system on stairs would separate the locomotion contribution from the perception contribution to the quiet stair-climbing behavior."],"forward_implications":["A quiet, slow-stepping gait can preserve a blind or low-vision user's ability to hear traffic, wall echoes, and other navigation cues while the robot moves at the user's preferred speed.","The controller keeps the robot upright under a 25 N backward pull and a 100 N impulse, so the noise reduction does not come at the cost of balance under real handler forces.","The pre-stop on the first stair step gives the same cue a trained guide dog gives, and users in the study could command the climb when ready.","The measured walking noise (about 50 dB) falls below the 65 dB reported for wheeled guide robots, addressing the reason earlier BLV participants chose wheeled systems."],"supporting_citations":[{"why":"Supplies the whole-body impulse control and convex MPC formulation that the paper modifies and uses as the primary stability baseline.","marker":"[33]"},{"why":"Provides the real-time SQP iteration scheme that lets the controller update at 400–500 Hz without waiting for full convergence.","marker":"[36]"},{"why":"Earlier RL-based guide-dog controller and user-expectation study; provides the noise-reward baseline and the Likert-scale survey items adapted here.","marker":"[25]"},{"why":"Reports 65 dB noise for wheeled guide robots and BLV users' preference for wheeled over quadruped systems, setting the noise target.","marker":"[26]"},{"why":"Presents an RL controller for quiet energy-efficient walking on a heavier quadruped, used as a noise-level comparison.","marker":"[32]"},{"why":"Reports measured guide-dog pulling forces that justify the 25 N dragging-force robustness test.","marker":"[27]"},{"why":"Provides the height-map-based landing location adjustment that the stair controller adapts for gentle touchdown.","marker":"[78]"},{"why":"Provides the 2.5D elevation mapping used to represent terrain for stair climbing.","marker":"[85]"},{"why":"Shows the SO(3)-based full-orientation state formulation used in the paper's nonlinear MPC.","marker":"[74]"},{"why":"The quadratic-programming solver that makes each SQP iteration fast enough for the high update rate.","marker":"[80]"}],"fun_headline_variants":["Robot guide dog: half the noise, same gait speed","Guide-dog robot with quiet, smooth gait preferred by all users","Steady quiet robot dog climbs stairs like a real guide","Robot guide dog: 10 dB quieter, handlers approve","Quadruped guide dog: quiet gait, smooth stairs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the robot's noise was measured the way a user would actually hear it — a researcher holding an uncalibrated smartphone at ear level while walking beside the robot, with no repeated trials or error bars reported for the headline 10 dB figure.","fun_headline_variants_meta":{"raw":{"variants":["Robot guide dog: half the noise, same gait speed","Guide-dog robot with quiet, smooth gait preferred by all users","Steady quiet robot dog climbs stairs like a real guide","Robot guide dog: 10 dB quieter, handlers approve","Quadruped guide dog: quiet gait, smooth stairs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000652,"raw_usage":{"total_tokens":3050,"prompt_tokens":1064,"completion_tokens":1986,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":1903}},"tokens_in":680,"tokens_out":1986,"duration_ms":12567,"temperature":1.0,"reasoning_tokens":1903,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:47:20.724841+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-measure both controllers with a fixed, calibrated microphone at the handler's ear height for many passes at each speed in the same room; if the mean difference is not close to 10 dB or the error bars overlap, the headline noise claim is not supported. Equally, a listening test in which blindfolded participants must detect a quiet sound source (such as a recorded car or a wall echo) while the robot walks would show whether the quieter gait restores the acoustic awareness the paper says it preserves.","supporting_citations":[{"cited_title":"A real-time iteration scheme for nonlinear optimization in optimal feedback control","cited_arxiv_id":null,"evidence_quote":"Provides the real-time SQP iteration scheme that lets the controller update at 400–500 Hz without waiting for full convergence."},{"cited_title":"Can quadruped guide robots be used as guide dogs? In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 4094–4100","cited_arxiv_id":null,"evidence_quote":"Reports 65 dB noise for wheeled guide robots and BLV users' preference for wheeled over quadruped systems, setting the noise target."},{"cited_title":"Accurate power consumption estimation method makes walking robots energy efficient and quiet","cited_arxiv_id":null,"evidence_quote":"Presents an RL controller for quiet energy-efficient walking on a heavier quadruped, used as a noise-level comparison."},{"cited_title":"System configuration and navigation of a guide dog robot: Toward animal guide dog-level guiding work","cited_arxiv_id":null,"evidence_quote":"Reports measured guide-dog pulling forces that justify the 25 N dragging-force robustness test."},{"cited_title":"Vision aided dynamic exploration of unstructured terrain with a small-scale quadruped robot","cited_arxiv_id":null,"evidence_quote":"Provides the height-map-based landing location adjustment that the stair controller adapts for gentle touchdown."},{"cited_title":"Probabilistic terrain mapping for mobile robots with uncertain localization","cited_arxiv_id":null,"evidence_quote":"Provides the 2.5D elevation mapping used to represent terrain for stair climbing."},{"cited_title":"Learning generic and dynamic loco- motion of humanoids across discrete terrains","cited_arxiv_id":null,"evidence_quote":"Shows the SO(3)-based full-orientation state formulation used in the paper's nonlinear MPC."},{"cited_title":"Stellato, G","cited_arxiv_id":null,"evidence_quote":"The quadratic-programming solver that makes each SQP iteration fast enough for the high update rate."}],"review_version":1}