{"id":"8c3c0c59-cf10-4c7c-8afb-0b380c820f66","arxiv_id":"2412.01324","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A hierarchical solver, S-SHQP, computes inverse kinematic plans and control commands while autonomously selecting goals, with cost linear in the number of candidate goals.","lead":"This paper presents a nonlinear solver that lets a robot choose which target to reach while simultaneously computing its whole-body joint motion, without first approximating what is reachable. It is a step toward real-time humanoid control in cluttered scenes where the robot must pick among many candidate goals.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The HSF convergence guarantee from the l2 solver [13] is asserted, not proved, for the new l0/log-barrier objective; without a sufficient-decrease argument for the sparse logarithmic cost, the central reliability claim for S-SHQP is open.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing gap: the HSF convergence proof from the l2 least-squares hierarchy is carried over to the l0/log-barrier objective without a proof, while the central claim that S-SHQP is the first to solve hierarchical decision-making with sparse nonlinear constraints depends on the solver's convergence and reliability. My stress-test pass did not find a stronger objection: the algorithmic structure is credible, the NQP complexity argument is plausible, and the experiments support the claimed functionality in the tested regimes. The main risk remains that the filter's sufficient-decrease logic, proven for a least-squares objective, may not hold for the reweighted l1/log objective, especially because the weights are updated at every outer iteration. I also note the Eq. (1)/Eq. (2) objective mismatch, which reinforces the concern that the formal problem being solved is not precisely specified; however, this is likely a typo that should be corrected rather than a separate fatal flaw. The verdict should remain CONDITIONAL until either a convergence proof for the extended filter is supplied or a counterexample is found and addressed.","tokens_in":14586,"tokens_out":10738,"duration_ms":110768,"concrete_test":"Re-derive the global-convergence theorem of [13] for the SHQP model with weights (2) and filter pair (h, phi) = (log(P(|f>=0_Cl|)+epsilon), log(sum |v_i|+epsilon)); check whether the standard HSF acceptance rule guarantees the required sufficient decrease in at least one of h or phi. In parallel, run S-SHQP on a minimal two-level instance with one selection group containing two distinct target constraints and an infeasible lower-priority equality, from 100 random starts; if the solver stalls, cycles, or accepts a step that increases both filter entries, the convergence transfer fails. If the derivation goes through and no counterexample appears, the conditional verdict can be lifted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III adapts the hierarchical step-filter (HSF) and trust-region acceptance logic of [13] to the SHQP with the l0/log-barrier objective, but no convergence theorem is stated or proved for that transfer. In [13], the global-convergence argument rests on the least-squares objective being bounded below, on a compatibility condition between the linearized model and the true objective, and on the filter guaranteeing sufficient decrease in either feasibility or optimality. Here the subproblem model changes at every outer iteration through the reweighted l1 weights in Eq. (2), and the filter's optimality measure for the log objective is never defined beyond the constraint-violation term log(P(|f>=0_Cl|)+epsilon). Without showing that an accepted step decreases either that term or the sparse logarithmic objective, the claim that S-SHQP reliably solves SH-NLP with sparse nonlinear constraints at any priority level is unsupported. The issue is aggravated by an internal inconsistency: Eq. (1) writes log(1^T t + xi), whose derivative is a scalar common to all components, while Eq. (2) gives per-component weights 1/(t_i+xi), which are the derivatives of sum_i log(t_i+xi). The actual problem being solved is therefore ambiguous, and the convergence question cannot be separated from this ambiguity. The experiments in Section VI demonstrate behavior but do not supply the missing proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a sequential sparse hierarchical quadratic programming solver (S-SHQP) for sparse hierarchical nonlinear programs (SH-NLP), targeting robot inverse kinematic planning and control with autonomous goal selection. The method reformulates l0-norm selection via a log-barrier approximation and reweighted l1 weights, solves each SHQP subproblem with an interior-point method called NQP, and embeds this in a hierarchical step-filter framework adapted from the authors' earlier l2 NL-HLSP solver. The authors claim linear scaling in the number of sparse constraints and demonstrate the approach on hierarchies of test functions, HRP-2Kai SHIK planning with goal selection, UR3e SHIK control with target selection, and HRP-2Kai SHIK control with continuous selection from 100 goals at 5 ms loop time.","tokens_in":14915,"tokens_out":2325,"duration_ms":21415,"significance":"If the claims hold, the paper makes a useful contribution to robotic hierarchical inverse kinematics: it enables discrete goal selection and whole-body IK/control within a single nonlinear optimization, without reachability approximations, and reports linear scaling in the number of candidate goals. The evaluation against external QP solvers (MOSEK, PIQP variants) and the explicit Theorem 1 on auxiliary-variable activity are strengths, and the reported 5 ms loop time for 100 goals, if reproducible, would support the real-time claim. The central limitation is that the convergence guarantee of the underlying l2 solver is asserted rather than proved for the new l0/log-barrier objective, which leaves the reliability claim for S-SHQP open.","major_comments":[{"comment":"The transfer of the hierarchical step-filter (HSF) convergence property from the l2 solver [13] to the new sparse logarithmic objective is asserted, not proved. In [13] the global-convergence argument relies on sufficient decrease of an l2 least-squares measure or constraint violation; here the filter is updated with log(P(|f>=0_Cl|) + epsilon), and the optimality measure for the reweighted l1/log objective is never defined. Without a sufficient-decrease argument for the new objective, the claim that S-SHQP reliably solves SH-NLP with sparse nonlinear constraints on any priority level is unsupported. This is a load-bearing gap because the experiments demonstrate behavior but do not supply the missing proof.","section":"Section III"},{"comment":"There is an internal inconsistency in the problem definition. Eq. (1) defines the log term as log(1^T t_Cl + xi), whose derivative is a scalar common to all components, while Eq. (2) defines per-component weights 1/(t_Cl,i + xi), which are the derivatives of sum_i log(t_Cl,i + xi). These are different objectives. The actual problem being solved is therefore ambiguous, and the convergence and weighting arguments in Section III cannot be separated from this ambiguity. Please clarify which formulation is intended and adjust the derivation of the SHQP cost and weights accordingly.","section":"Eqs. (1)-(2)"},{"comment":"Table 1 reports low-accuracy sparsity on at least one level for each solver (e.g., level 6 for NQP gives ||v6,1||=1.95 and ||v6,2||=0.05, rather than an exact zero) and the text acknowledges that 'at least one of the levels is solved only to low accuracy sparsity.' Since the central claim includes 'accurate' sparse solutions and exact selection, the paper should quantify the sparsity accuracy achieved in the robot experiments and state whether the nonzero slacks in such cases remain below the selection threshold used in Section IV. Without this, the practical reliability of the autonomous goal selection is not fully demonstrated.","section":"Section VI-A, Table 1"},{"comment":"Several numerical thresholds that are load-bearing for reproducibility are left unspecified: xi (log approximation threshold), epsilon (sparse/activity threshold and switching threshold), chi (step convergence threshold), and the trust-region radius rho in control mode are all introduced without values, and the UR3e joint sparsity weight and object removal distance in Section VI-D are also not given. Please provide the specific values used in each experiment, or state how they were chosen.","section":"Sections III and VI"}],"minor_comments":[{"comment":"The phrase 'It can be shown that the step of SHQP and Newton's method applied to the first-order optimality conditions of (1) are the same' is standard for SQP, but a short derivation or a precise pointer to the relevant result would improve clarity.","section":"Section III"},{"comment":"The right-hand side terms r_El, r_Il, and r_Il-1 in Eqs. (13)-(15) are dense and hard to parse; a brief derivation or a compact notation for the eliminated variables would help readers verify the linear-complexity claim.","section":"Section V, Eq. (9)-(15)"},{"comment":"The column headers mix norms and constraint types (e.g., 'Disk ineq.' and 'Ros. eq.') in a way that requires cross-referencing the text; defining all abbreviations in the caption would improve readability.","section":"Table 1"},{"comment":"The statement that 'in most control instances, at least one of the tracking errors of the two targets is zero' would be more informative with a quantitative fraction and a tolerance threshold, since exact zero is unlikely in floating-point arithmetic.","section":"Section VI-C"},{"comment":"The paper does not state whether the implementation and experiment scripts are available for reproduction; providing code or detailed parameter lists would substantially strengthen the reproducibility of the reported timings and sparsity patterns.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own prior work [12], [13], [19] for the hierarchical step-filter, the nullspace basis computation, and the interior-point method. While those are published and appropriate bases, the missing convergence proof for the new objective is the main technical risk. The self-citation burden is not excessive for a methods paper, but the editor may want to ensure that the novelty claims about 'first to be able to solve such hierarchical decision making problems' are positioned carefully against the related l1/linear sparse programming literature. The Eq. (1)/(2) inconsistency is easily fixable but must be addressed before the paper can be considered sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real attempt at a genuinely useful capability—solving nonlinear hierarchical IK with sparse selection constraints on any priority level, with linear scaling in the number of sparse constraints. The experiments back the basic functionality, and the NQP interior-point derivation (Theorem 1, eqs. 9–15) is coherent. But the convergence argument for the l0/log objective is asserted, not proved, and there is an inconsistency between Eq. (1) and Eq. (2) about which log-objective is actually minimized. Those two issues are the load-bearing soft spots.\n\nWhat is new: they assemble known components—reweighted l1, hierarchical least-squares, interior-point—into a single framework that demonstrably selects among goals in nonlinear IK planning and control. The timing result on 100 goals (4.5 ms) is impressive and beats QP baselines. The selection-constraint bookkeeping (discarding infeasible goals to avoid Hessian activation) is a practical trick that seems to matter. The paper is honest enough to report low-accuracy sparsity on some levels.\n\nSoft spots, in rough order of severity. First, the convergence transfer. The HSF filter is taken from [13] and applied to the sparse logarithmic objective without a sufficient-decrease argument. The filter's optimality measure for the log cost is never defined beyond the constraint-violation term. That is a genuine gap; without it, 'reliably solves' is not supported. Second, the objective mismatch: Eq. (1) writes log(1^T t + ξ) whose derivative is a scalar shared by all components, while Eq. (2) gives per-component weights 1/(t_i+ξ) that are derivatives of sum_i log(t_i+ξ). The actual problem being solved is ambiguous. Third, no code and unspecified thresholds (xi, epsilon, chi, rho) make the linear-scaling claim hard to verify independently. Fourth, the 'first' claim is plausible but depends on a narrow reading of prior work; the paper does cite the relevant alternatives. Self-citation is mild and mostly to published papers, so not a concern.\n\nOverall: the core idea is sound and the experiments are suggestive, but the convergence gap and objective ambiguity need to be fixed before I'd trust the solver beyond demo. This deserves peer review—send it to a serious venue and ask for major revision, with particular attention to the objective definition and a convergence or failure-mode analysis for the filter.","headline":"A credible new formulation for sparse hierarchical IK with autonomous goal selection, but the convergence transfer is unproved and the log-objective is ambiguous; deserves serious refereeing with major revision.","tokens_in":15463,"tokens_out":2801,"would_cite":true,"duration_ms":23514,"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 nonlinear solver can now pick a goal and plan the whole-body motion to reach it.","keywords":["sparse hierarchical nonlinear programming","inverse kinematics","autonomous goal selection","l0 norm","hierarchical optimization","interior-point method","whole-body control","humanoid robot"],"falsifier":"Run S-SHQP on a small hierarchy with a known unique sparse solution, such as a two-level problem where the sparse level has two candidate targets and the lower level is a Rosenbrock regularization, and check whether the filter can be made to accept steps that monotonically increase the true nonlinear logarithmic objective. If a step sequence cycles or diverges while the filter accepts every step, the convergence assumption fails.","tokens_in":1340,"feed_emoji":"🤖","tokens_out":1776,"duration_ms":35179,"temperature":0.7,"pith_summary":"This paper proposes a nonlinear programming framework that folds discrete decision-making directly into inverse kinematic planning and control. Its central claim is that sparse, nonlinear, priority-ordered constraints — for example, \"choose one of these 100 end-effector targets\" — can be solved on any level of a task hierarchy, with computation that grows linearly in the number of candidate choices rather than cubically. If the claim holds, a robot can simultaneously select a feasible Cartesian goal and compute a complete kinematic plan or control command without reachability approximations, and can do so fast enough for real-time humanoid control.","feed_headline":"Sparse solver picks goals and plans whole-body motion in one pass","feed_subtitle":"A new nonlinear optimizer handles 100 goal candidates at about 5 ms, with no reachability approximations.","key_machinery":"The load-bearing object is the selection constraint group: a vector of least-squares constraints, one per candidate goal, whose ℓ0-norm counts how many candidates remain unsatisfied. The continuous surrogate uses a logarithmic barrier over auxiliary variables, with weights that are re-fixed once the hierarchical step-filter commits to a level's solution. The interior-point solver NQP then forms normal equations in the nullspace of previously active constraints; because the sparse auxiliary variables are eliminated via the one-sided-activity theorem, each sparse constraint contributes only linearly to the factorization cost.","core_discovery":"The paper's central discovery is a sequential sparse hierarchical quadratic programming solver, S-SHQP, that extends the authors' earlier least-squares hierarchical solver to ℓ0-norm selection. The ℓ0 objective is replaced by a continuous logarithmic reformulation, and the solver's interior-point method NQP exploits a key property: for each sparse constraint, the auxiliary bound is always active on exactly one side unless the slack is zero. This lets the auxiliary variables be eliminated, so the per-iteration cost depends linearly on the number of sparse constraints instead of cubically. The paper demonstrates the resulting autonomous goal selection on a humanoid robot, including selecting from 100 passing objects at roughly 5 ms per control cycle.","pith_inferences":["The one-sided-activity argument suggests the method could be extended to other combinatorial selections, such as contact-point choice or grasp selection, where the candidate set enters as a group of least-squares constraints.","Because the per-iteration cost scales linearly in the number of sparse constraints, the practical bottleneck for very large candidate sets will shift from the solver to the nullspace basis computation and the number of S-SHQP iterations.","A straightforward test of the solver's reliability is to compare its selected goal against an exhaustive combinatorial check on small candidate sets, where the true ℓ0 optimum can be computed exactly.","The control mode's constant trust-region acceptance rule means its real-time guarantee depends on the trust-region radius being small enough for the linearized subproblem to stay valid; a harder test would push the robot near singular configurations during selection."],"forward_implications":["Robot planners can select a discrete goal and compute a feasible whole-body trajectory in one optimization, removing the separate reachability approximation step.","The same solver handles both offline planning and instantaneous control, with the control mode accepting every step under a constant trust-region radius.","Real-time autonomous goal selection scales to hundreds of candidate targets: the paper reports 100 goals on a humanoid at roughly 5 ms per solve.","Sparse ℓ0 regularization of joint motion yields visibly reduced joint engagement compared to unregularized least-squares control.","The approach is positioned as a step toward simultaneous contact planning and optimal control from many candidate contacts."],"supporting_citations":[{"why":"Supplies the sequential hierarchical least-squares framework, including the hierarchical step-filter and trust-region constraint that S-SHQP adapts to the sparse logarithmic objective.","marker":"[13]"},{"why":"Provides the reweighted-ℓ1 continuous approximation of ℓ0-norm minimization that underlies the logarithmic surrogate in the new formulation.","marker":"[6]"},{"why":"Supplies the interior-point method for hierarchical least-squares programs, including the nullspace projection strategy that NQP extends and specializes to sparse constraints.","marker":"[19]"},{"why":"Provides the filter-SQP convergence framework the hierarchical step-filter is built on, which the paper assumes carries over to the sparse setting.","marker":"[9]"},{"why":"Establishes the hierarchical Newton's method and the real-time feed-forward inverse kinematic control loop in which the 5 ms performance is judged sufficient.","marker":"[12]"},{"why":"Acts as a reference QP solver baseline (PIQP) against which the proposed NQP solver's speed is compared.","marker":"[14]"},{"why":"Acts as a reference QP solver baseline (MOSEK) against which the proposed NQP solver's speed is compared.","marker":"[15]"}],"fun_headline_variants":["L0-norm solver picks end-effector targets in milliseconds","Hierarchical solver unifies IK planning and goal selection","5 ms sparse solver chooses among 100 grasp targets","Nonlinear optimizer tackles discrete choices in IK at 5 ms","Sparse hierarchical solver: one pass for goals and motion"],"cache_read_input_tokens":17536,"weakest_assumption_plain":"The proof of convergence assumes that the hierarchical step-filter, which was designed for least-squares objectives, still guarantees progress when applied to the new sparse logarithmic cost.","fun_headline_variants_meta":{"raw":{"variants":["L0-norm solver picks end-effector targets in milliseconds","Hierarchical solver unifies IK planning and goal selection","5 ms sparse solver chooses among 100 grasp targets","Nonlinear optimizer tackles discrete choices in IK at 5 ms","Sparse hierarchical solver: one pass for goals and motion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000622,"raw_usage":{"total_tokens":2846,"prompt_tokens":872,"completion_tokens":1974,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":1893}},"tokens_in":488,"tokens_out":1974,"duration_ms":12760,"temperature":1.0,"reasoning_tokens":1893,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:29:19.545698+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run S-SHQP on a small hierarchy with a known unique sparse solution, such as a two-level problem where the sparse level has two candidate targets and the lower level is a Rosenbrock regularization, and check whether the filter can be made to accept steps that monotonically increase the true nonlinear logarithmic objective. If a step sequence cycles or diverges while the filter accepts every step, the convergence assumption fails.","supporting_citations":[{"cited_title":"Sequential hierarchical least-squares programming for prioritized non-linear optimal control,","cited_arxiv_id":null,"evidence_quote":"Supplies the sequential hierarchical least-squares framework, including the hierarchical step-filter and trust-region constraint that S-SHQP adapts to the sparse logarithmic objective."},{"cited_title":"Enhancing sparsity by reweighted ℓ1 minimization,","cited_arxiv_id":null,"evidence_quote":"Provides the reweighted-ℓ1 continuous approximation of ℓ0-norm minimization that underlies the logarithmic surrogate in the new formulation."},{"cited_title":"On the global convergence of a filter–sqp algorithm,","cited_arxiv_id":null,"evidence_quote":"Provides the filter-SQP convergence framework the hierarchical step-filter is built on, which the paper assumes carries over to the sparse setting."},{"cited_title":"The hier- archical newton’s method for numerically stable prioritized dynamic control,","cited_arxiv_id":null,"evidence_quote":"Establishes the hierarchical Newton's method and the real-time feed-forward inverse kinematic control loop in which the 5 ms performance is judged sufficient."},{"cited_title":"ApS, MOSEK Fusion API for C++ 10.1.12 , 2019","cited_arxiv_id":null,"evidence_quote":"Acts as a reference QP solver baseline (MOSEK) against which the proposed NQP solver's speed is compared."}],"review_version":2}