{"id":"51c20ce4-d70b-4af0-b68b-a40d30ace4eb","arxiv_id":"2411.12968","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A QP-based thruster-assisted walking controller, validated only in simulation, keeps the Harpy biped tracking a reference trajectory on a 30-degree inclined slope.","lead":"This paper designs a quadratic programming controller that combines a reduced-order inverted pendulum model with thruster forces to make a bipedal robot, Harpy, walk on a 30-degree incline in simulation. The result matters because it offers a control recipe for legged robots that use thrusters to assist locomotion on steep slopes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The QP's reduced-order VLIP model assumes constant CoM height and zero vertical acceleration, but the whole-body controller does not enforce these constraints; the paper provides no evidence that the high-fidelity simulation satisfies them, leaving the QP-thruster force mapping unvalidated.","rationale":"The reader's weakest-assumption identification points to the reduced-order VLIP model with point mass, massless legs, and arbitrary-direction thrusters. My stress-test narrows this to a specific, concrete gap: the VLIP requires constant CoM height and zero vertical acceleration, but the whole-body controller (Eq. 29) does not enforce these conditions. This is genuinely load-bearing because the QP's linearized dynamics (Eq. 18) and the subsequent thruster-force computation (Eq. 12) are only valid under those assumptions. If the high-fidelity Simscape model, with distributed link masses and a compliant ground model, produces vertical CoM motion or nonzero z-acceleration, then the QP's predicted horizontal acceleration is not achieved, and the observed stable walking cannot be confidently attributed to the QP-based force planning. The paper gives no plot of CoM height, no z-error trajectory, no comparison between commanded and actual GRF, and no numerical performance metrics—only qualitative figures. Therefore, the central claim is currently unsupported at its most critical point. The proposed test is a direct computational check: if the VLIP assumptions hold within tight tolerances, the mapping is validated; if not, the paper's claim of QP-driven stability is significantly weakened. I agree with the reader's conditional assessment; no new concern beyond that is needed. The verdict should remain conditional until the authors supply the missing quantitative validation.","tokens_in":8601,"tokens_out":8930,"duration_ms":86598,"concrete_test":"Run the Simscape simulation for at least 10 complete gait cycles on the 30° slope, logging: (1) actual CoM height PB,z(t) and vertical acceleration ̈PB,z(t); (2) actual GRF λ_actual from the compliant ground model; (3) commanded thruster ut and QP-computed λ. Then: (a) compute the maximum deviation of PB,z from the nominal z0 used in Eq. 18 and the peak |̈PB,z| relative to g; (b) evaluate the residual of Eq. 12 at each time step, r = m̈PB,x + mg sin(α) − ut,x + λx_actual, and similarly for z, with the commanded thrust and actual GRF. If the height deviation exceeds about 5% of z0 or the residual magnitude is comparable to the commanded forces, the VLIP assumption is violated and the QP-thruster mapping is not consistent with the high-fidelity dynamics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the QP-derived forces (λx, λz, ut) computed from the reduced-order VLIP model (Sec. II-C, Eqs. 12–14) faithfully predict the high-fidelity Simscape model's dynamics. The VLIP assumes a point mass at constant height (PB,z = z0, ̈PB,z = 0) and massless legs; the QP's linearization (Eq. 18) omits the vertical state entirely. However, the whole-body controller (Eq. 29) does not enforce constant height or the VLIP dynamics; it only enforces contact (Eq. 27) and planner constraints (zero y, roll, yaw accelerations, Eq. 28). The body's vertical position and pitch are free to vary with the compliant ground model and joint dynamics. If the actual CoM height or vertical acceleration deviates from the VLIP assumptions, the thruster force computed by inverting Eq. 12 (Sec. IV-B) will not produce the commanded horizontal acceleration, and the QP's predictions are not reliable. The paper provides no quantitative validation of these assumptions: no CoM height plot, no comparison of QP-commanded vs actual GRF, no tracking-error metrics. The single simulation (Fig. 4) is qualitative, and the conclusion that the QP 'quickly was able to achieve a stable limit cycle' is unsupported by quantitative gait metrics. Without evidence that the reduced-order model is predictive, the claimed stable limit cycle cannot be attributed to the QP controller; it may stem from the swing-leg PID and the passive ground model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a quadratic-programming (QP) based stance controller for the Harpy bipedal robot with thrusters, combined with a whole-body force mapping, to achieve thruster-assisted walking on a 30-degree inclined slope. The QP uses a variable-length inverted pendulum (VLIP) reduced-order model to compute ground reaction forces and thruster forces under friction-cone constraints, and a whole-body controller maps these forces to joint torques. Validation is performed in a single Matlab Simscape simulation, with qualitative plots of body states, joint angles, ground reaction forces, and thruster forces. The paper claims that the QP computes contact forces in 0.3 ms and that the controller quickly reaches a stable limit cycle.","tokens_in":8825,"tokens_out":9135,"duration_ms":89892,"significance":"If the simulation claims are quantitatively verified, the work would be a useful demonstration of optimization-based control for legged-aerial systems, extending standard VLIP and whole-body control ideas to thruster-assisted slope walking. The architecture is clearly relevant to bio-inspired wing-assisted incline running. The paper also provides a complete model derivation and reports a favorable QP solve time. However, the current evidence is mostly qualitative, and the central claim of stable limit-cycle walking rests on unverified reduced-order assumptions and an unquantified match between commanded and simulated contact forces.","major_comments":[{"comment":"The central claim that the QP controller 'quickly was able to achieve a stable limit cycle' is not supported by the evidence presented. Please report quantitative tracking-error norms for body position and velocity, a comparison of the QP-commanded ground reaction forces with the forces produced by the compliant ground model (Eq. 9), and a gait-cycle metric such as periodicity, orbital stability, or CoM height excursion. Without these, the simulation demonstration in Fig. 4 remains qualitative and the attributed effectiveness of the QP is not established.","section":"Section V, Figs. 4–8"},{"comment":"The QP relies on VLIP assumptions of constant CoM height and zero vertical acceleration (PB,z = z0, \\ddot PB,z = 0), but the whole-body controller in Eq. (29) does not enforce these conditions; body height, pitch, and vertical acceleration are free to vary in the high-fidelity model. Please provide plots of PB,z, pitch, and \\ddot PB,z over the gait and verify that the reduced-order model assumptions are actually satisfied. If they are not, the inversion of Eq. (12) to compute thruster forces may produce incorrect commands, so this verification is load-bearing for the paper's central claim.","section":"Section II-C and Section IV-A/B"},{"comment":"The friction-cone constraints are stated incompletely: 'λz > λmin and λx < μ|λz|' omits the lower bound on λx (λx ≥ −μλz) and does not define an upper bound on λz, and the inequality in Eq. (24) is written as strict ('Ain u < Bin') although a QP requires non-strict inequalities. Please state the complete cone as implemented in Bin, specify which μ from the Stribeck model in Eq. (9) is used in the QP, and reconcile the strict inequality notation.","section":"Section IV-A, Eqs. (23)–(25)"},{"comment":"The whole-body controller imposes a rigid-contact constraint Js \\ddot q = −˙Js ˙q, while the simulation uses a compliant ground model with Stribeck friction (Eq. 9). The paper does not demonstrate that the actual ground reaction forces in the simulation track the QP-commanded GRFs, which is necessary to claim that the controller applies the QP solution to the high-fidelity model. Please add such a comparison and report any foot slip or contact-model mismatch.","section":"Section IV-B, Eq. (29)"}],"minor_comments":[{"comment":"The term 'cosQP(α)' appears to be a typo and should read 'cos(α)'.","section":"Section II-C, Eq. (12)"},{"comment":"The sentence 'Harpy’s height measures 600 cm' is presumably a typo for 60 cm; please correct the value and verify all physical dimensions.","section":"Section I"},{"comment":"The PID gains used in Eq. (16) are not reported; please list Kp, Ki, and Kd so that the simulation is reproducible.","section":"Section V-A"},{"comment":"The statement that 'even when QP optimization is not running robot never violates friction cone condition' is unclear; please clarify whether it refers to the intervals between 100 Hz QP updates and define how constraint satisfaction is monitored.","section":"Section V-B"},{"comment":"The caption of Fig. 8 states that the figure shows λx, λz and thruster forces, while the text refers to it as ground reaction forces from the QP solver; please align the caption with the text and label the axes clearly.","section":"Section V-B, Fig. 8"},{"comment":"The linearization in Eq. (18) uses λz in the state matrix A while λz is also treated as an optimization variable in the input vector u; please state the operating point used for this linearization and justify the resulting time-invariant prediction over the horizon.","section":"Section IV-A, Eq. (18)"}],"recommendation":"major_revision","confidential_remarks":"The novelty relative to the authors' earlier Harpy capture-point control paper [13,18] should be clarified; the current work appears to be an incremental extension to a QP-based formulation, and a clear statement of the new contribution would help the editor assess fit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper combines a VLIP reduced-order model, a condensed QP with friction-cone constraints, and a whole-body mapping to produce thruster-assisted slope walking for the Harpy biped in simulation. The new bit is the specific combination; the prior Harpy controller used capture-point control, and the thruster-augmented VLIP formulation on a 30-degree incline is a reasonable next step. The derivation of the full-body dynamics is detailed, and the use of a compliant ground model with Stribeck friction is more realistic than the rigid contacts in many similar studies. The simulation does show stable-looking walking, and the QP's 0.3 ms solve time is worth reporting.\n\nThe soft spots are mostly about evidence, not about the core idea. There is one simulation, no quantitative tracking errors, no baseline comparison (e.g., without thrusters or against the prior capture-point controller), and no hardware. The PID gains for the swing leg are not given, and the Q/R weights are hand-tuned, so the results are hard to reproduce. There are also typos (\"cosQP\", \"600 cm\" tall) and the VLIP derivation is compressed to the point of being hard to follow.\n\nThe stress-test concern about the VLIP assumptions deserves a direct answer. The whole-body controller does not explicitly enforce constant CoM height or zero vertical acceleration, and the QP's linearized model omits the vertical state. However, the paper states that the thruster force is computed by inverting Eq. (12) using the body's actual acceleration, so the force balance at the body is not purely open-loop. The remaining risk is that the QP's GRF predictions are based on a model that may not match the full simulation if the CoM height or vertical acceleration deviate significantly. The paper should include traces of the CoM height and vertical acceleration, and ideally a comparison of QP-commanded versus actual GRFs. Without those, the claim of a stable limit cycle is not quantitatively supported.\n\nThe friction-cone statement \"even when QP optimization is not running robot never violates friction cone condition\" is confusing and should be clarified. The citation pattern is fine; the self-citations are to prior Harpy work and are appropriate, though a direct comparison to ref. [13] would strengthen the paper.\n\nThis is a subfield-contribution paper, not a breakthrough. It deserves a serious referee, but my advice is to require major revisions: quantitative metrics, a baseline, full controller parameters, and explicit validation of the reduced-order model against the full simulation. As it stands, the paper is a decent workshop-level result that needs more work to be truly convincing.","headline":"Simulation-only QP thruster-assisted walking for the Harpy biped: a plausible extension of established legged-control components, but the evidence is thin and the ROM-to-full-model transfer is not quantitatively validated.","tokens_in":9510,"tokens_out":2643,"would_cite":false,"duration_ms":60914,"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":"The paper claims that a quadratic-programming controller using a virtual inverted-pendulum model, friction-cone constraints, and whole-body force mapping can produce stable thruster-assisted walking on a 30-degree inclined slope in…","keywords":["quadratic programming","bipedal locomotion","thruster-assisted walking","reduced-order model","VLIP","friction cone","whole-body control","inclined slope"],"falsifier":"Put the same controller on the physical Harpy robot on a 30-degree slope and see whether the gait converges to the reference; if the real thruster mounting directions and distributed body inertia invalidate the simplified model, the walk will not stabilize. A cheaper check is to rerun the simulation with the thrusters restricted to Harpy's actual force directions instead of the idealized any-direction assumption and observe whether the stable limit cycle survives.","tokens_in":8296,"feed_emoji":"🤖","tokens_out":14625,"duration_ms":136259,"temperature":0.7,"pith_summary":"The paper aims to show that a bipedal robot with body-mounted thrusters can walk up a steep 30-degree slope using a quadratic-programming (QP) controller rather than hand-tuned heuristics. The controller plans ground reaction forces with a simplified virtual linear inverted pendulum model, enforces friction-cone constraints so the feet do not slip, and maps the planned forces through a whole-body controller to joint torques and thruster forces. In simulation of the Harpy robot, the gait tracks the reference, settles into a stable limit cycle, and the QP solves in 0.3 milliseconds. This matters because thrusters could let legged robots climb inclines that exceed what leg friction alone can support, just as Chukar birds use their wings to run up steep walls.","feed_headline":"Thruster-assisted biped walks up a 30-degree slope in simulation","feed_subtitle":"A quadratic-programming controller keeps the gait stable and slip-free, pointing legged robots toward slopes friction alone cannot handle.","key_machinery":"The load-bearing object is the virtual linear inverted pendulum (VLIP) reduced-order model projected onto the sagittal plane, in which the robot is a point mass on massless legs with thrusters able to push in any direction. This model is linearized and discretized into a state-space form, and a condensed quadratic program solves for the ground reaction forces $[\\lambda_x,\\lambda_z]^\\top$ that minimize tracking error while respecting friction-cone inequalities. A whole-body mapping, built from the full model's dynamics with contact and planner constraint Jacobians, then turns those reduced-order forces into stance-leg torques and thruster forces for the high-fidelity simulation model.","core_discovery":"The paper's central claim is that adding thrust vectoring to a bipedal walking controller makes the reduced-order stance dynamics fully actuated, and that a QP formulation can exploit this to produce stable slope walking. The QP minimizes tracking error in the linearized discrete dynamics with state $x = [P_{B,x}, \\dot{P}_{B,x}]^\\top$ and inputs $u = [\\lambda_x, \\lambda_z]^\\top$, subject to friction-cone bounds $\\lambda_z > \\lambda_{\\min}$ and $\\lambda_x < \\mu|\\lambda_z|$. Its outputs are ground reaction forces; the body acceleration equation then gives the thruster forces, and a whole-body mapping converts these, together with a polynomial swing-leg trajectory, into stance-leg torques. On the simulated Harpy model at a 30-degree slope, the body position and velocity track the QP reference, the planned contact forces stay inside the friction cone, and the walk converges to a stable limit cycle.","pith_inferences":["A direct test of the thruster's role would be to raise the slope until a no-thruster controller slips and then show this controller continues walking; the gap would quantify how much thrust adds beyond friction-limited locomotion.","The planner constraint that zeros yaw, roll, and lateral acceleration confines the demonstrated result to the sagittal plane, so a 3D extension would need the QP to also manage lateral foot placement; the paper identifies 3D motion as its next step.","Because the thrusters are idealized as massless and omni-directional, the simulation margin may shrink when real actuator mounting angles and response delays are added; adding those constraints to the QP is a testable refinement."],"forward_implications":["The controller tracks the reference trajectory in simulation, so the QP and whole-body pipeline is a working control pattern for thruster-assisted slope walking, not just a planning abstraction.","Because the QP enforces friction-cone bounds, the planned gait is slip-free by construction, and the simulated ground forces remain inside the cone throughout the walk.","The 0.3 ms solve time at a 100 Hz update rate leaves ample computation margin, supporting real-time use on the physical robot.","Thrusters make the simplified stance dynamics fully actuated, so the controller can track a desired velocity profile instead of only stabilizing an underactuated gait.","The repeated stable limit cycle in simulation indicates that the swing-leg trajectory and the stance QP form a complete, repeatable gait cycle."],"supporting_citations":[{"why":"Supplies the real-time sparse QP solver used to compute contact forces in 0.3 ms at the 100 Hz update rate.","marker":"[21]"},{"why":"Establishes prior controller design for the same Harpy platform that this QP formulation builds on.","marker":"[17]"},{"why":"Gives the earlier capture-point controller for thruster-assisted Harpy locomotion that this work extends.","marker":"[13]"},{"why":"Documents wing-assisted incline running in Chukar birds, the biological analogue for thruster-assisted slope walking.","marker":"[14]"},{"why":"Provides aerodynamic analysis of wing-assisted incline running, supporting the idea that thrust augments contact forces on steep slopes.","marker":"[15]"},{"why":"Demonstrates an earlier robot combining posture manipulation and thrust vectoring, defining the design space this controller targets.","marker":"[7]"}],"fun_headline_variants":["Biped robot climbs 30° slope with thruster-boosted gait","Thrusters make bipedal slope walking fully actuated","QP controller drives thruster-assisted biped up slopes","Simulated Harpy robot uses thrusters to master 30° slopes","Thruster-assisted biped walks steep slopes via quadratic programming"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the robot's body and legs behave like the simplified pendulum model used by the controller -- a point mass on massless legs with thrusters that can push in any direction -- so forces computed on that model remain correct for the full machine.","fun_headline_variants_meta":{"raw":{"variants":["Biped robot climbs 30° slope with thruster-boosted gait","Thrusters make bipedal slope walking fully actuated","QP controller drives thruster-assisted biped up slopes","Simulated Harpy robot uses thrusters to master 30° slopes","Thruster-assisted biped walks steep slopes via quadratic programming"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1255,"prompt_tokens":918,"completion_tokens":337,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":248}},"tokens_in":534,"tokens_out":337,"duration_ms":3940,"temperature":1.0,"reasoning_tokens":248,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:58:56.305561+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Put the same controller on the physical Harpy robot on a 30-degree slope and see whether the gait converges to the reference; if the real thruster mounting directions and distributed body inertia invalidate the simplified model, the walk will not stabilize. A cheaper check is to rerun the simulation with the thrusters restricted to Harpy's actual force directions instead of the idealized any-direction assumption and observe whether the stable limit cycle survives.","supporting_citations":[],"review_version":1}