{"id":"69f7d1ca-d95d-40a0-81c6-4a2b91a3c25c","arxiv_id":"2607.18527","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A ducted-fan spring hopper autonomously switches between hopping and flying in one contact-implicit MPC, measuring about 27% of hovering control effort during hopping.","lead":"DASH is a small robot that hops on a springy leg and flies with a ducted fan, and one controller decides by itself when to do each. The same contact-implicit controller can switch between hopping and flight without predefined mode schedules, and hopping uses about a quarter of the control effort of hovering.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unspecified leg-spring parameters and energy reference undermine the central efficiency and mode-emergence claims.","rationale":"The reader's weakest assumption pinpoints the lack of a defined leg-spring model, which is the most load-bearing concern because it directly affects the dynamics used for contact-implicit optimization. The aerial model ignores the leg, but the terrestrial model includes a spring without specifying its constitutive law; this makes the LCP and gradients incomplete. The efficiency number (26.78%) is computed from commanded thrust, but if the model is wrong, the commanded thrust may not reflect real system efficiency, and the mode-emergence behavior could be an artifact of model mismatch. The proposed concrete test—identify spring parameters and re-simulate—would settle whether the claim generalizes. The motion-capture dependence is a secondary limitation, noted by the reader, but not as directly tied to the core control claim. Since the reader already gave CONDITIONAL, and this concern reinforces the conditionality without demanding rejection, the verdict remains UNCHANGED. The authors should disclose spring parameters, E_ref, and MPC weights to make the claim reproducible.","tokens_in":11608,"tokens_out":6144,"duration_ms":73956,"concrete_test":"Identify the leg spring parameters (stiffness k, damping b, free length l0) via bench tests (e.g., free vibration or drop test). Then re-simulate the same CI-MPC with these identified parameters, keeping all other settings (weights, horizon, relaxation ρ) identical to the paper's description. Compare the simulated hopping cycle, control-effort integral, and mode-transition behavior to the hardware recordings. If the control-effort ratio shifts by more than 20% relative to the reported 26.78%, or the planned contact sequence changes qualitatively, then the unspecified spring model is load-bearing and the original efficiency claim is not reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a single contact-implicit MPC autonomously selects hopping or flying and achieves a 26.78% reduction in injected control effort (Section VI.A). This claim rests on the predictive accuracy of the dynamics model in Section IV.B. In the terrestrial model, the robot is described as containing a passive 'prismatic springy joint' (leg), but the spring stiffness, damping, and free length are never defined; the spring force is subsumed into the generic h(q,v) term in Eq. (7). The aerial model explicitly omits leg dynamics entirely. Consequently, the LCP-based contact dynamics and the analytic gradients (Eqs. 21–29) depend on an unspecified compliance model. Additionally, the MPC cost includes a mechanical energy term ∥E(x_k)−E_ref∥², but E_ref is not defined, so the optimizer may be biased toward a preferred mode. Without these parameters, the 26.78% reduction in commanded thrust and the automatic hopping-to-flying transitions cannot be reproduced or even assessed for robustness. If the real leg spring has different stiffness, damping, or hysteresis, the model may not capture the energy circulation, and the reported efficiency could be an artifact of a tuned but incorrect model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DASH, a ducted-fan aerial-terrestrial robot with a passive spring leg, and proposes a contact-implicit MPC (CI-MPC) that solves a single optimization problem to select between hopping and flying without hand-coded mode schedules. The dynamics are formulated with LCP-based contact conditions, and the MPC is solved with differential dynamic programming using relaxed analytic gradients. Experiments demonstrate in-place hopping, flying with velocity tracking up to 0.5 m/s, and obstacle-triggered transitions between modes. The central quantitative claim is that hopping reduces the injected control effort, defined as ∫|T(t)−T_min|dt, to 26.78% of the flying case, and that this efficiency emerges from the optimizer rather than from explicit switching logic.","tokens_in":11961,"tokens_out":7260,"duration_ms":85286,"significance":"If substantiated, the paper's contribution is of interest to the hybrid aerial-terrestrial robotics community: it combines a mechanically simple platform with a unified contact-implicit control framework, avoiding predefined contact sequences and mode schedules. The hardware demonstrations of hopping, flying, and mode transition are valuable, and the use of complementarity-based contact dynamics with analytic gradients follows a modern and reproducible methodology. The claimed 26.78% control-effort reduction is a concrete, falsifiable result. However, the strength of the paper depends on three currently under-reported elements: the leg compliance model, the energy reference in the cost, and the experimental conditions behind the headline efficiency number. These omissions currently prevent reproduction and independent assessment of the central claims.","major_comments":[{"comment":"The headline claim that hopping reduces injected control effort to 26.78% of the flying case is not supported by the reported data. No trial count, duration, standard deviation, or battery/voltage conditions are given. The metric ∫|T(t)−T_min|dt also depends on the unspecified constant T_min. Moreover, the comparison is between different task conditions: the hopping experiment is in-place with zero reference velocity, while the flying experiment includes tracking a reference velocity of up to 0.5 m/s (Fig. 7 and Section VI.A). A fair efficiency comparison should use the same task specification or explicitly report both tasks separately. Please provide repeated trials, error bars, exact T_min, and the experimental protocol.","section":"Section VI.A, Fig. 8"},{"comment":"The terrestrial model in Eq. (7) treats the leg as a 'prismatic springy joint' but never specifies its spring stiffness, damping, or free length; these are hidden inside the generic h(q,v) term. The aerial model, Eqs. (5)–(6), ignores the leg dynamics entirely, with the justification that 'the leg and foot are of small weight.' However, Table I lists the carbon-fiber leg at 542 g, which is about 16.6% of the 3.272 kg total mass. This contradiction is load-bearing: energy circulation and hopping behavior depend directly on the leg compliance, and the relaxed contact gradients in Eqs. (21)–(29) rely on the leg-included dynamics. The paper must report the leg parameters (stiffness, damping, free length, effective mass) and either justify the small-weight assumption with data or include the leg dynamics in the model.","section":"Section IV.B and Table I"},{"comment":"The MPC cost contains the term ∥E(x_k)−E_ref∥²_QE, which is described as enforcing 'coherent energy circulation.' However, E_ref is never defined. It is not stated whether E(x_k) is the total mechanical energy of the floating base, the energy of the leg spring, or a combined quantity; nor is the value of E_ref or its physical meaning specified. Because this term is central to the optimizer's mode selection and the claimed energy efficiency, the omission is a load-bearing reproducibility issue. Please define E(x_k) and E_ref explicitly, and explain how E_ref is chosen (e.g., from the desired hopping height or as an open parameter).","section":"Section V.B, Eq. (20)"},{"comment":"The actuation model in Eqs. (1)–(4) relies on coefficients c1–c4, and the text states they are identified through hardware experiments detailed in Section VI. However, no numerical values, confidence intervals, or identification residuals are reported. Furthermore, the paper enforces Ω1=Ω2, which sets τ_p=0; this makes c3 effectively unused and does not validate the coaxial moment model. Since these coefficients directly enter the MPC dynamics and therefore influence the planned contact modes and the efficiency result, the identified values and validation data should be reported.","section":"Section VI.A, System Identification"}],"minor_comments":[{"comment":"T_p is defined as a vector in Eq. (1), but in Eq. (4) it is used as a scalar magnitude (Tp). Please clarify notation, e.g., use |T_p| or define Tp as the scalar thrust magnitude.","section":"Eq. (4) and Section IV.A"},{"comment":"The expression for A_cc in Eq. (23) has unmatched brackets and the term 'Jns + Jts μ sign(λts)' is ambiguous. Please rewrite with explicit parentheses and define all subscripts (n, t, c, s) in the text near Eq. (21).","section":"Eqs. (23)–(24)"},{"comment":"The paper states the robot has 7 degrees of freedom on the ground (SE(3) plus one prismatic joint), but the aerial model in Eqs. (5)–(6) is a 6-DOF rigid body. The reduction from 7 to 6 DOF when the leg is ignored should be stated explicitly, and the effect on mass/inertia parameters should be quantified.","section":"Section IV.B"},{"comment":"The text gives the total system mass as 'approximately 3.27 kg' and 'height of 4 feet.' Please use SI units throughout and verify the height: 4 feet seems large for a 3.27 kg pogo-style robot. Also, '10-inch radius' should be given in meters or millimeters.","section":"Section III.A and Table I"},{"comment":"The paper states that translational states are obtained from a motion capture system, and later describes 'autonomous mode transitions.' This is controller-level autonomy, not onboard autonomy. A sentence clearly stating this as a limitation (and distinguishing it from fully onboard operation) would prevent misinterpretation.","section":"Section III.B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope and the platform is a contribution, but the central quantitative claim (26.78% control-effort reduction) is not yet substantiated by sufficient experimental detail, and the model has critical unspecified elements (leg spring, E_ref, c1–c4). These are fixable with additional reporting and analysis, so I recommend major revision rather than rejection. The novelty relative to prior hopping quadrotors (PogoX, Hopcopter) should also be sharpened with a quantitative comparison if possible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is the first hardware demonstration I know of where a single contact-implicit MPC makes a ducted-fan spring-legged robot hop, fly, and switch between the two without a hand-written mode scheduler. The 26.78% control-effort reduction relative to hovering is plausible, but the paper does not yet provide enough experimental detail to treat that number as settled.\n\nWhat is new: the integration itself. CI-MPC with LCP contact dynamics is standard in legged robotics, and ducted fans plus pogo legs exist separately. DASH is the concrete combination, and the obstacle-triggered hop-to-flight transition on hardware is exactly the kind of result that moves the field. The simulation and hardware results align with the claim that contact modes emerge from feasibility and optimality rather than from switching logic. The actuator identification of c1-c4 is standard but appropriate.\n\nWhere it is soft: First, the leg. The terrestrial model has a prismatic springy joint, but the paper never states stiffness, damping, free length, or how the leg mass enters the mass matrix, and the aerial model excludes leg dynamics. The LCP and the analytic gradients therefore depend on a compliance model the reader cannot assess. If the real spring has hysteresis or the leg mass couples with thrust at the 0.03 s planning rate, the planned transitions and the efficiency number may be tuned to this bench rather than general. Second, the energy term in the cost is ||E(x_k)-E_ref||^2, but E_ref is never defined. Without it, you cannot tell whether the optimizer is biased toward one mode and cannot reproduce the trade-off. Third, the 26.78% result appears to be a single hardware comparison with no repeated trials or error bars. For an efficiency claim that is the headline, that is thin. Minor: the related-work sentence saying ducted fans have not been applied to hybrid mobile robots contradicts the paper's own reference [15], SkateDuct; soften that. Also, translational state comes from motion capture, so \"autonomous\" applies to mode selection, not onboard estimation. Those are not fatal, but they should be fixed.\n\nOverall, the central claim holds up in the sense that nothing in the equations reduces to the result by construction, and the hardware is real. A careful reader comes away believing the mode-free controller works, with the details still underspecified.\n\nRecommendation: send it to peer review. A good referee will ask for leg-spring identification, E_ref, repeated efficiency trials, and a corrected novelty statement. With those, this is a strong archival paper.","headline":"First real hardware demo of contact-implicit MPC on a ducted-fan spring hopper, with automatic hop-to-flight transitions and a plausible but under-evidenced 26.78% efficiency claim.","tokens_in":12395,"tokens_out":2400,"would_cite":true,"duration_ms":28398,"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 single contact-implicit model predictive controller lets a ducted-fan spring-legged robot choose between hopping and flying without a predefined mode schedule, cutting injected control effort to about 27% of hovering.","keywords":["aerial-terrestrial robot","ducted fan","contact-implicit MPC","hopping","mode-free locomotion","energy efficiency","complementarity","spring leg"],"falsifier":"Instrument the leg with a force sensor and compare the measured ground-contact times against the contact phases planned by the CI-MPC during an obstacle-crossing run; if the measured and planned contact events deviate by more than one 0.03 s control step, the complementarity-based contact model is not the mechanism producing the behavior.","tokens_in":11527,"feed_emoji":"🤖","tokens_out":4228,"duration_ms":48064,"temperature":0.7,"pith_summary":"This paper claims that a robot with one ducted fan and one springy pogo leg can be controlled by a single contact-implicit MPC, with no hand-coded switching between terrestrial and aerial modes. The controller treats contact through complementarity conditions, so hopping, flying, and transitions emerge from optimizing a cost that includes mechanical energy tracking. The authors demonstrate in-place hopping, flight at 0.5 m/s, and obstacle traversal where the optimizer autonomously leaves the ground and later returns to hopping. If this holds, it would mean hybrid locomotion can be unified in one optimization rather than layered mode logic, and that energy-efficient hopping can be discovered automatically.","feed_headline":"One optimizer makes this robot hop, fly, and switch by itself","feed_subtitle":"Contact-implicit control lets a ducted-fan spring-legged robot choose its gait, cutting hop thrust to about 27% of hovering.","key_machinery":"The load-bearing object is the contact-implicit MPC: a model predictive controller built on time-stepping rigid-body dynamics with complementarity conditions for non-penetration and maximum dissipation, solved as a linear complementarity problem (LCP). Analytical gradients of the contact impulse, relaxed at contact boundaries, let a differential dynamic programming solver optimize the thrust, moment, and contact schedule concurrently. The cost function tracks the reference state plus mechanical energy, which encourages passive energy circulation through the spring leg.","core_discovery":"The central discovery is that a single optimization over a linear-complementarity contact model automatically picks the locomotion mode. When ground contact is feasible, the optimizer chooses hopping and intermittently stores and releases elastic energy; when an obstacle makes contact infeasible, the same optimizer transitions to flight. In hardware, hopping reduced the injected control effort, defined as the time integral of thrust minus the minimal motor-idling thrust, to 26.78% of the hovering case.","pith_inferences":["If the efficiency result is robust, the energy-tracking term in the cost function is likely the key; a natural ablation study would remove that term and test whether the controller reverts to hovering or loses the 26.78% reduction.","The hardware experiments rely on motion capture for translational state, so the claimed autonomy still needs demonstration with onboard visual–inertial estimation before outdoor operation is credible.","The leg spring parameters are not reported; a sensitivity analysis varying stiffness and damping would tell whether the automatic hopping–flight transitions survive across leg designs.","The authors mention swimming as future work; the same contact-implicit formulation could be extended to a fluid-contact regime, but that would require a different contact model."],"forward_implications":["Mode transitions in hybrid robots need not be scheduled in advance: feasibility and optimality in a single MPC are sufficient to switch between hopping and flying.","Spring-legged hopping can be selected automatically over hovering when ground contact is available, cutting injected thrust to about 27% of the flying case.","The same controller can be applied to other ducted-fan legged robots without rewriting switching logic, since contact behavior is determined by the optimization.","Energy tracking in the cost function is a mechanism that makes the optimizer prefer intermittent contact over sustained thrust; this may generalize to other energy-storing robots."],"fun_headline_variants":["Meet DASH: a ducted-fan spring-hop robot that chooses its own moves","Single controller lets DASH hop or fly with no mode switching code","Contact-implicit control cuts hop thrust to ~27% of hovering","DASH robot uses one optimizer to hop and fly, saving energy","Spring-legged flying robot picks its gait via contact-implicit MPC"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the leg behaves as an unmodeled passive spring while motion capture provides perfect translation state; if the spring's stiffness, damping, or leg-swing dynamics couple with thrust at the 0.03 s planning rate, or if state estimation becomes imperfect outdoors, the automatic mode-switching and efficiency gain may fail to generalize.","fun_headline_variants_meta":{"raw":{"variants":["Meet DASH: a ducted-fan spring-hop robot that chooses its own moves","Single controller lets DASH hop or fly with no mode switching code","Contact-implicit control cuts hop thrust to ~27% of hovering","DASH robot uses one optimizer to hop and fly, saving energy","Spring-legged flying robot picks its gait via contact-implicit MPC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1087,"prompt_tokens":653,"completion_tokens":434,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":397,"completion_tokens_details":{"reasoning_tokens":340}},"tokens_in":397,"tokens_out":434,"duration_ms":5909,"temperature":1.0,"reasoning_tokens":340,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T15:08:14.595524+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Instrument the leg with a force sensor and compare the measured ground-contact times against the contact phases planned by the CI-MPC during an obstacle-crossing run; if the measured and planned contact events deviate by more than one 0.03 s control step, the complementarity-based contact model is not the mechanism producing the behavior.","supporting_citations":[],"review_version":1}