{"id":"a749a4d4-7469-4187-aa51-bdeb2cf83371","arxiv_id":"2505.12231","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A 3-DOF flat-foot hopping robot with a custom MINLP-optimized 3K planetary gearbox performs repetitive hops, a front flip, step-up, and push recovery, serving as an intermediate validation platform for a bipedal robot.","lead":"The authors built a 12.45 kg one-legged robot with a knee, two ankle joints, and a flat foot, and showed it hopping forward and backward, stepping onto a low platform, recovering from pushes, and doing a front flip. The paper is an engineering demonstration that compact high-torque gearboxes and a reinforcement-learned controller can form an intermediate testbed for a future bipedal robot.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) checks out; the load-bearing weakness is that all demonstrations are tethered and cable interaction is never quantified, so the hardware validation may be confounded.","rationale":"The reader's identified weakest assumption, the unproven gear-ratio Eq. (5), is the right kind of load-bearing item, but it is not actually a defect. Using the standard Willis relation for the 3K train with the structural center-distance equations (2) and (3), the input-output ratio simplifies to Eq. (5); plugging ZS=44, ZP1=44, ZP2=32, ZF=132, ZO=120 gives exactly 20. Thus the fabrication ratio and peak torque are internally consistent. The plausible load-bearing weakness is external validity of the hardware experiments: the umbilical tether is the only acknowledged interaction not controlled for. A cable can apply forces during a 0.3-0.5 s hop, and no measurement rules this out. The manuscript also lacks trial counts, success rates, code/CAD/data release, and has minor inconsistencies (840 vs 890 mm, Z_P3), which support the conditional verdict. If a cable-force check shows negligible interaction, the experimental portion should be judged on its current qualitative merits; if not, the central platform-validity claim is materially weakened. Either way, the present evidence does not justify moving from CONDITIONAL to ACCEPT.","tokens_in":10623,"tokens_out":18569,"duration_ms":207151,"concrete_test":"Mount a six-axis load cell at the tether attachment point on the robot (or at the cable anchor) and record cable force/torque during the forward/backward hopping, step-up, front flip, and push-recovery protocols. Report peak and RMS cable force by phase. If peak force exceeds about 5% of body weight (~6 N for the 12.45 kg robot) or is systematically nonzero during flight, repeat the hopping trials with a longer/slack cable or an onboard power source; if the behavior changes materially, the tethered experiments do not validate untethered actuator/controller capability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central claim: the robot's custom 3K gearbox actuators and learned controller generate stable, repeatable hopping and aggressive maneuvers, validating the platform. The reader's Eq. (5) concern does not land: re-deriving the 3K-kinematic ratio from the Willis relation with structural constraints (2) and (3) yields exactly Eq. (5), and the selected tooth counts give 20.0. The most load-bearing unsecured condition is that the experiments are claimed to be \"without any external support except for the umbilical cable,\" yet the cable's mechanical interaction is never measured or bounded. A tether carrying power/communication through the air can exert vertical tension and lateral forces on the base, especially during flight and touchdown, which can assist stabilization, reduce required joint torque, and make a front flip or landing succeed. Because the stated purpose is to validate the hardware as an intermediate platform for an untethered biped, the demonstration is only evidence if the cable force is negligible. No load-cell data, cable management description, or slack/tension analysis is provided; single-demo traces and snapshot videos also do not establish repeatability.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents the design, fabrication, and experimental evaluation of a 12.45 kg, 3-DOF hopping robot with a flat foot and a human-like lower-limb layout (one knee DOF, two ankle DOFs), positioned as an intermediate platform toward a 13-DOF biped. The hardware contribution centers on a custom knee actuator (D151) containing a 20:1 3K compound planetary gearbox whose tooth counts (ZS=44, ZP1=44, ZP2=32, ZF=132, ZO=120) are selected by a mixed-integer nonlinear program with a hollow shaft for cable routing, plus custom motor drivers and an EtherCAT-CAN converter. The controller is a PPO-trained reinforcement-learning policy with barrier-based style rewards, trained in RaiSim with a closed-loop ankle model and deployed to hardware. Experiments demonstrate forward and backward hopping, an 85 mm step-up, a front flip, and push recovery, all with the robot tethered by an umbilical cable. The paper claims stable, repetitive hopping and concludes the platform is an effective intermediate testbed for bipedal robots.","tokens_in":10852,"tokens_out":6702,"duration_ms":67423,"significance":"If the claims are established, this is a useful contribution: the MINLP gear-tooth selection is concrete and the selected tooth counts verifiably satisfy the stated structural, assembly, and ratio equations, which is a strength; the hardware is real, and the periodic torque/velocity traces and video snapshots give external evidence that the system hops. I also confirmed that Eq. (5), though underived, is correct when derived from the Willis relation and constraints (2)-(3), so the gear-ratio concern raised by the companion stress-test does not land. However, the experimental evidence is anecdotal: there are no repeated-trial statistics, no quantitative performance metrics, and the umbilical cable's mechanical interaction is never measured or bounded. Because the paper's central claim is that the platform validates hardware for an untethered biped, these gaps are load-bearing. The strengths are real, but the validation standard needs to be raised.","major_comments":[{"comment":"Section IV states that all experiments were performed 'without any external support except for the umbilical cable,' but the cable's mechanical interaction is never quantified or bounded. During flight and touchdown, tension in a power/communication tether can exert vertical and lateral forces on the 12.45 kg body, potentially assisting balance, jump height, and flip landing. Since the stated purpose is to validate hardware for an untethered biped (Abstract and Section V), this is a load-bearing gap. Please add a cable-force measurement (e.g., a load cell at the anchor), a slack/tension analysis, or an explicit demonstration that cable forces are negligible relative to body weight (about 122 N) and the measured joint torques. If such data are not available, the conclusions must be correspondingly tempered.","section":"Section IV, first paragraph; Section V"},{"comment":"The claims of 'repetitive hopping' and 'successfully completed a full rotation and landed upright' are supported by one snapshot sequence and one torque/velocity trace per behavior. There are no repeated-trial counts, success rates, or variability metrics such as hop height, hop period, aerial time, foot placement error, or landing success. A single successful demonstration does not establish repeatability or robustness. Please provide N repeated trials with quantitative summaries and, if failures occurred, report the failure rate.","section":"Section IV, Figs. 7 and 8"},{"comment":"Equation (5), the gear-ratio formula for the 3K compound planetary gearbox, is stated without derivation or citation. I re-derived it from the Willis relation together with the structural constraints (2) and (3) and found it to be correct, so this is not a correctness objection; however, the derivation should be included or a source cited, because the selected tooth counts and the claimed 20:1 ratio and knee torque all rest on this formula, and the formula is not obvious.","section":"Section II-D, Eq. (5)"}],"minor_comments":[{"comment":"The Abstract reports 840 mm as the fully extended length, while Section II-A reports 890 mm; these values should be made consistent.","section":"Abstract and Section II-A"},{"comment":"The cost function in Eq. (1) contains Z_{P3}^2, but the optimization variables define only Z_{P1} and Z_{P2}; this should be Z_{P2} to be consistent with the rest of the paper.","section":"Eq. (1)"},{"comment":"Table II reports D151 rotor inertia as 'measured in the three-dimensional design software'; this is a CAD estimate, not a measurement, and should be labeled as such.","section":"Table II"},{"comment":"The controller description refers to [16] for the observation vector details; since [16] is a preprint, include a full list of observations and training hyperparameters for reproducibility.","section":"Section III"},{"comment":"The claim that the knee actuator 'reached both its torque and velocity limits' is not verifiable from Fig. 8 because no limit lines or definitions of the limits are shown; please add them.","section":"Fig. 8"},{"comment":"Reference [26] is an anonymous submission under review; the final version should resolve the attribution and status of this citation according to journal policy.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's novelty relative to prior KAIST Hound work [13] and the authors' own RL framework [16] should be checked by the editor; the contribution is primarily hardware integration with a learned controller, and the experimental evaluation is currently too anecdotal to fully support the stated conclusions. Also, reference [26] is cited as an anonymous under-review submission; if this paper appears first, that citation may compromise double-blind review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead arXiv:2505.12231. The headline: this is a real hardware platform paper with a genuinely new demonstration—a 3-DOF flat-foot hopper doing repetitive hopping, front flip, step-up, and push recovery—and the gear-design optimization is verifiable. The selected teeth satisfy Eq. (5), and re-deriving the 3K ratio from the Willis relation gives the same formula, so the reader's Eq. (5) worry does not land. The paper earns credit for shipping a working 12.45 kg robot with custom hollow-shaft actuators, custom drivers, and an RL controller trained in RaiSim with closed-loop ankle kinematics. That is nontrivial engineering.\n\nThe soft spots are real but not fatal. First, all demos are tethered; the umbilical cable is the only stated support, but its mechanical interaction is never measured or bounded. For a hopping/front-flip validation, cable tension can assist stabilization or change touchdown loads, so the hardware claim is not fully clean. Second, there are no repeated-trial statistics or quantitative success rates; the evidence is snapshots and a few torque/velocity traces. That is typical for this kind of paper but still limits strength. Third, internal inconsistencies: abstract says 840 mm, body says 890 mm; Eq. (1) uses Z_P3 but the variable list defines only Z_P1 and Z_P2. Minor but should be fixed. No code, CAD, or data is released, which makes the RL and optimization hard to reproduce.\n\nThe citation pattern is okay: the MINLP gear optimization and barrier-RL are cited to prior work, and the flip policy is from a submitted anonymous paper—that citation is appropriate. Self-citation is not a problem here because the cited methods are actually used.\n\nWho is this for? Legged-robotics hardware people, especially those building humanoid-scale joints. It is a credible intermediate platform report, not a breakthrough in control or design theory. It deserves a serious referee: the experiments should be repeated with measured cable force, and the paper needs trial counts and a cable-force bound. I would send it to review with a request for major revision along those lines, not desk-reject.\n\nReading group maybe; I'd bring it if the group cares about hardware platforms.","headline":"Real hardware platform with a new flat-foot hopping demo and verifiable gear design; the main gap is unquantified cable interaction, not the gear-ratio formula.","tokens_in":11450,"tokens_out":2157,"would_cite":true,"duration_ms":20091,"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 flat-footed 3-DOF leg, with an optimized planetary gearbox and a learned controller, hops, flips, climbs, and recovers—the authors' case that the platform can grow into a bipedal robot.","keywords":["hopping robot","3K compound planetary gearbox","mixed-integer nonlinear programming","gear teeth optimization","reinforcement learning control","flat-foot legged locomotion","bipedal robot development","sim-to-real transfer"],"falsifier":"Instrument the fabricated D151 gearbox: drive the sun gear at a known speed, measure the output ring speed, and compare the ratio to the 20:1 target; a large deviation would refute the gearbox claim. Independently, re-derive the ratio for the published tooth counts ($Z_S=44$, $Z_{P1}=44$, $Z_{P2}=32$, $Z_F=132$, $Z_O=120$) from standard compound-planetary kinematics and check it against Eq. (5); any mismatch would refute the optimization's central constraint rather than the prototype.","tokens_in":10391,"feed_emoji":"🤖","tokens_out":12681,"duration_ms":108739,"temperature":0.7,"pith_summary":"This paper establishes that a single 3-DOF leg with a flat foot, arranged like a human lower limb with one knee and two ankle joints, can hop repeatedly, flip forward, climb steps, and recover from pushes when driven by actuators whose gearboxes are designed by mixed-integer nonlinear programming. Hopping matters as a test because it forces a leg to deliver high impact torque, stay agile, and manage intermittent ground contact—conditions a full bipedal robot will also face. Because the motor, the 3K compound planetary gearbox, the motor driver, and the communication interface were all custom-built to fit the joint geometry, the hardware experiments validate the actuation and structure as much as the controller. The authors conclude that the platform, weighing 12.45 kg, is an effective intermediate step toward a 13-DOF bipedal robot, with the gearbox optimization and the learned-control pipeline carrying over directly.","feed_headline":"A custom gearbox makes a single leg hop, flip, and recover","feed_subtitle":"The 12.45 kg flat-footed leg, controlled by a learned policy, is the authors' stepping stone to a full biped.","key_machinery":"The load-bearing mechanism is the 3K compound planetary gearbox of the knee actuator, a compact reducer in which two planet gears and two ring gears share one carrier to reach a high ratio within a small diameter. Its tooth counts are the outputs of a mixed-integer nonlinear program whose target is the ratio identity $G_{\\text{target}} = \\frac{2Z_{P1}(Z_F - Z_{P1} + Z_{P2})}{(Z_F - 2Z_{P1})(Z_{P1} - Z_{P2})}$, constrained by the structural equalities $Z_F = Z_S + 2Z_{P1}$ and $Z_O = Z_S + Z_{P1} + Z_{P2}$, by assembly conditions, and by carrier-interference and gear-size inequalities. This identity is what ties the chosen integers to the claimed 20:1 reduction, the hollow shaft, and the torque rating, and it lets the gearbox sit inside the motor rotor so the actuator stays compact. The complementary mechanism is the closed-loop parallel ankle linkage, with a universal joint at the foot and ball joints at the parallel actuator outputs, which the training simulation represents as a pin-constrained kinematic chain so the learned policy is aware of the coupling between the two ankle actuators.","core_discovery":"The paper's central claim is that dynamic, repetitive flat-foot hopping with a human-like leg layout is achievable on a 3-DOF prototype, and that this validates the hardware design as the foundation for a bipedal robot. In the authors' account, the knee actuator reaches a 320 N·m peak torque and a 20:1 reduction through a 3K compound planetary gearbox whose tooth counts ($Z_S = 44$, $Z_{P1} = 44$, $Z_{P2} = 32$, $Z_F = 132$, $Z_O = 120$) were selected by solving a mixed-integer nonlinear program that maximizes the hollow-shaft diameter for cable routing, minimizes mass, and enforces assembly, interference, and size constraints. The same custom-design approach extends to the motor drivers and the EtherCAT–CAN converter, which fit within the compact joint envelopes. On the control side, a reinforcement-learning policy trained in a simulation that models the closed-loop ankle linkage as a kinematic chain with a pin constraint transfers to the real robot without external support and produces the demonstrated behaviors: repetitive forward and backward hopping at 0.8 m/s, a complete front flip landing upright, an 85 mm step-up, and push recovery. The authors read these results as evidence that the actuators can deliver both the torque and the velocity that aggressive maneuvers demand, and that the platform can absorb the resulting impact loads.","pith_inferences":["A standalone check of the uncited ratio formula, Eq. (5), is the fastest way to confirm the gearbox claim independently: re-deriving it from standard compound-planetary kinematics and comparing with the published tooth counts takes a short calculation.","Because the same controller handled an 85 mm step with no retraining, its terrain margin is probably larger than demonstrated; a sweep of increasing step height would quantify how much headroom the sim-to-real training left on the table.","The experiments run tethered, so the onboard battery and the custom converter have not been stressed by repeated impacts; running the same motions on battery power would test whether the power and communication chain, not just the actuators, survives dynamic loads.","A measured push-impulse limit for the single-leg policy would double as a lower-bound disturbance-rejection benchmark for the biped, since balancing on one flat foot is harder than on two."],"forward_implications":["The mixed-integer nonlinear gearbox-design procedure—ratio, assembly, interference, and size constraints over integer tooth counts—applies directly to the joints of the planned 13-DOF biped, which the authors state is being built on the same electrical architecture and gearbox methodology.","The reinforcement-learned controller transfers from a simulation that explicitly models the closed-loop ankle, so the same training pipeline should extend to the more complex kinematics of a full leg without a new method.","The knee actuator reaching both its torque and velocity limits during the front flip means the D151 is sized at the envelope for aggressive maneuvers, giving a measured operating point for specifying the biped's actuators.","Flat-foot standing and push recovery are demonstrated on a single leg with a small support polygon; the biped's larger support polygon and second leg should make the same learned behaviors easier to maintain."],"supporting_citations":[{"why":"Supplies the mixed-integer nonlinear optimization method for gear-train design that the paper extends to the 3K compound planetary gearbox.","marker":"[13]"},{"why":"Provides the closed-loop parallel mechanism adopted for the ankle so two parallel actuators drive pitch and roll degrees of freedom.","marker":"[14]"},{"why":"Supplies the assembly-feasibility conditions (Eq. 4) that the tooth-count optimization must satisfy.","marker":"[15]"},{"why":"The barrier-based style-reward reinforcement learning framework that the hopping controller extends, including gait-phase and foot-clearance priors.","marker":"[16]"},{"why":"Supplies the proximal policy optimization algorithm used to train the control policy.","marker":"[21]"},{"why":"Provides the contact dynamics solver used to build the closed-loop ankle simulation in which the policy is trained.","marker":"[22]"},{"why":"Provides the centroidal-velocity-reward and sim-to-real technique on which the front-flip policy is based.","marker":"[26]"}],"fun_headline_variants":["3-DOF leg hops via RL, 0.8 m/s, flips","Custom gearbox leg hops, flips, recovers","Learned flat-foot hopping leg for bipeds","Optimized gearbox enables 3-DOF leg hops","RL-driven leg hops, flips, step-up, recover"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire actuator design rests on an unstated gear-ratio formula, Eq. (5): if that formula mis-models how the two planet gears and two ring gears interlock, the knee joint is not actually a 20:1 reducer and the hopping and flipping experiments do not validate the gearbox as designed.","fun_headline_variants_meta":{"raw":{"variants":["3-DOF leg hops via RL, 0.8 m/s, flips","Custom gearbox leg hops, flips, recovers","Learned flat-foot hopping leg for bipeds","Optimized gearbox enables 3-DOF leg hops","RL-driven leg hops, flips, step-up, recover"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1317,"prompt_tokens":1004,"completion_tokens":313,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":227}},"tokens_in":620,"tokens_out":313,"duration_ms":3954,"temperature":1.0,"reasoning_tokens":227,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:39:09.673685+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Instrument the fabricated D151 gearbox: drive the sun gear at a known speed, measure the output ring speed, and compare the ratio to the 20:1 target; a large deviation would refute the gearbox claim. Independently, re-derive the ratio for the published tooth counts ($Z_S=44$, $Z_{P1}=44$, $Z_{P2}=32$, $Z_F=132$, $Z_O=120$) from standard compound-planetary kinematics and check it against Eq. (5); any mismatch would refute the optimization's central constraint rather than the prototype.","supporting_citations":[{"cited_title":"Design of kaist hound, a quadruped robot platform for fast and efficient locomotion with mixed- integer nonlinear optimization of a gear train,","cited_arxiv_id":null,"evidence_quote":"Supplies the mixed-integer nonlinear optimization method for gear-train design that the paper extends to the 3K compound planetary gearbox."},{"cited_title":"Design and control of the rapid legged platform gazelle,","cited_arxiv_id":null,"evidence_quote":"Provides the closed-loop parallel mechanism adopted for the ankle so two parallel actuators drive pitch and roll degrees of freedom."},{"cited_title":"Single planetary mechanism teeth matching conditions,","cited_arxiv_id":null,"evidence_quote":"Supplies the assembly-feasibility conditions (Eq. 4) that the tooth-count optimization must satisfy."},{"cited_title":"Learning impact-rich rotational maneuvers via cen- troidal velocity rewards and sim-to-real techniques: A one-leg hop- per flip case study,","cited_arxiv_id":null,"evidence_quote":"Provides the centroidal-velocity-reward and sim-to-real technique on which the front-flip policy is based."}],"review_version":1}