{"id":"b097cf32-78d4-4a31-a7d2-535a193e4bfe","arxiv_id":"2504.17647","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"For sampled-data first-order robots with matched safety margins and a linear class-κ function, the complementarity QP and the CBF-QP have the same optimal velocity command.","lead":"This paper proves that two popular robot safety controllers, one based on complementarity constraints and one based on control barrier functions, produce identical velocity commands for velocity-controlled robots in discrete time. The result gives a dictionary for translating between the two frameworks and reusing each other's solvers and guarantees.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The complementarity controller is defined via the false identity J†_{c,i} n_i = (n_i^T J_{c,i})† (Eq. 15), so the controller solved in the numerics may differ from the one Theorem 1 equates to the CBF-QP.","rationale":"The reader's weakest assumption is the scope restriction (linear class-κ, matched margins), which is real but explicitly acknowledged in the paper and does not by itself make the central argument incorrect. The more immediate correctness issue is the false pseudoinverse identity that creates an inconsistency between the controller derivation (15) and the formal object of Theorem 1 (18). The theorem itself is sound for the H=G(A) controller once the extraneous false 'AH is PSD' aside in the proof is removed; the equations defining H and the stationarity condition directly give x* = Hλ' without needing semidefiniteness. The numerical validation can be interpreted as supporting the theorem only if the implementation used (18), which the paper does not state. The fix is straightforward: either define the complementarity controller exclusively via G(A) and delete/correct (15), or prove the equivalence for the J†n parametrization, which would require a different feasible set and would not coincide with the CBF-QP in general. This concern strengthens the conditions for acceptance rather than overturning the verdict, so the conditional recommendation stands.","tokens_in":9707,"tokens_out":26547,"duration_ms":238333,"concrete_test":"Using the 3-DoF planar robot in Section VI, at a sample state q_k where a link-to-obstacle distance constraint is active, compute D1 = J†_{c,i} n_i and D2 = (n_i^T J_{c,i})† for the active link. If D1 ≠ D2, solve three problems: CBF-QP (25) with α(s)=s/τ and δ_CBF=δ_LC; LCQP (19) with the set (18) using H=G(A); and the LCQP variant using u = u_des + D1 λ under the same complementarity constraint (11)-(13). Compare u_CBF with both LCQP solutions over a trajectory. The theorem predicts exact match only for the H=G(A) variant; the (15)-variant will deviate unless the false identity happens to hold. This settles whether the reported numerical equivalence applies to the controller actually derived in (15).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV.A parametrizes the complementarity control input by u = u_des + Σ_i J†_{c,i} n_i λ_i (Eq. 15), justified by the claim J†_{c,i} n_i = (n_i^T J_{c,i})†. This identity is false for a general contact Jacobian; e.g., for J = diag(1,2) and n = (1,1)/√2, J†n = (0.707,0.354) while (n^T J)† = (0.283,0.566). Equations (16)-(18) then define the feasible set U_LC via G(∂h/∂q), whose columns are the row pseudoinverses (n_i^T J_{c,i})†; hence (15) and (18) define different controllers unless the false identity holds. Theorem 1 analyzes the (18) version, so the claimed equivalence is not established for the controller defined by (15). If the MATLAB/fmincon implementation in Section VI follows (15), the reported agreement with the CBF-QP would be coincidental or require special structure; if it follows (18), the paper contains an internal inconsistency in its controller derivation. This is load-bearing because it affects which controller the central equivalence actually applies to.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies two reactive safety filters for velocity-controlled manipulators: a complementarity-constrained QP, in which a nonnegative multiplier activates a correction when the predicted distance falls below a threshold, and a CBF-QP enforcing sampled-data barrier conditions. After rewriting both as abstract QPs over a variable x with data A and b, the paper proves (Theorem 1) that the problem min (1/2)||x||^2 subject to Ax-b >= 0 and the complementarity-parametrized problem min (1/2)||x||^2 subject to x = G(A)lambda, 0 <= lambda perp A G(A)lambda - b >= 0 have identical optimal solutions, where G(A) is the row-pseudoinverse operator. The proof uses KKT conditions and the row-space relation between A and H=G(A). A single-constraint version is given first, and a numerical example on a 3-DoF planar manipulator reports agreement between the two controllers to within solver tolerance.","tokens_in":10048,"tokens_out":18632,"duration_ms":168428,"significance":"The main theorem is a correct and clean equivalence between a linearly constrained QP and an LCP-parametrized QP under the row-pseudoinverse map H=G(A); the KKT proof is sound, and the numerical experiment with a 3-DoF planar robot supports the predicted agreement. If the scope restrictions are stated accurately, the result is a useful theoretical bridge between the complementarity and CBF communities, with concrete downstream benefits such as transferring sampled-data margin guarantees. The main caveats are that the equivalence is established only for the G-parametrized complementarity controller, not for the J† n parametrization introduced in Eq. (15), and only for the linear class-kappa function alpha(s)=s/tau with matched thresholds. The paper's broad title and framing should be tempered to reflect this scope.","major_comments":[{"comment":"The asserted identity J†_{c,i} n_i = (n_i^T J_{c,i})† is false for a general contact Jacobian. For example, with J=diag(1,2) and n=(1,1)/sqrt(2), J† n=(0.707,0.354)^T while (n^T J)†=(0.283,0.566)^T. Consequently Eq. (15) and Eq. (17) define different control laws, and Theorem 1 applies to the G-parametrized set (18), not to the controller defined by Eq. (15). The problem statement (19) uses U_LC from (18), so the reported numerical experiment should be interpreted as validating the G-parametrized controller. The paper must either define the complementarity controller directly via G(partial h/partial q) and remove the false identity, or explicitly restrict the unification claim to that parametrization and state the limitation.","section":"Sec. IV.A, Eq. (15)"},{"comment":"The reduction lambda* = argmin_lambda ||J†_{c} n lambda||^2 = argmin_lambda ||lambda||^2 is valid only when ||J†_{c} n||=1, which does not hold for a general full-row-rank contact Jacobian. For the G-parametrization x=(n^T J_c)† lambda used in Theorem 1, one has ||x||^2=lambda^2/||n^T J_c||^2, so minimizing over lambda is equivalent to minimizing lambda^2; this is the correct version of the single-constraint argument. As written, the single-constraint proof inherits the error from Eq. (15) and should be rewritten around the row-pseudoinverse parametrization.","section":"Sec. V.A, Eqs. (28)-(32)"}],"minor_comments":[{"comment":"The claimed convexity of U_LC does not follow from the statement that A_LC G(A_LC) is positive semidefinite: A G(A) is not symmetric in general and need not satisfy z^T A G(A) z >= 0 for all z. For example, A=[[1,0],[0.1,0.1]] gives A G(A)=[[1,5],[0.1,1]], and z=(1,-1) yields z^T A G(A) z = -3.1. The corollary needs a proof or should be removed.","section":"Sec. V.C, Corollary 1"},{"comment":"The theorem statement should include the assumption that X0 is nonempty, since the KKT argument applies only to feasible problems; the control application presumably guarantees this, but it should be stated explicitly.","section":"Sec. V, Theorem 1"},{"comment":"Please clarify the domain of the class-kappa function alpha_i when h_i - delta_CBF,i is negative, since class-kappa functions are usually defined on [0,a) and the sampled-data condition in Eq. (23) evaluates them at possibly negative arguments.","section":"Sec. IV.B, Eq. (23)"},{"comment":"Please state explicitly whether fmincon solves the complementarity problem in the form (19)-(18) with G(partial h/partial q), or a formulation based on Eq. (15); the numerical result reported in Eq. (48) is otherwise not tied to the controller analyzed in Theorem 1.","section":"Sec. VI.A"},{"comment":"There are minor language and reference typos, e.g., 'casted' in Sec. II and 'nommonotone' in Ref. [19]; these should be corrected in the final version.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The central theorem is sound for the G-parametrized complementarity controller and the numerical test supports it, but the false identity in Eq. (15) currently overstates the scope of the unification. If the authors intend to equate their controller with the J† n parametrization used in prior complementarity-based work, the equivalence is generally false; if they are willing to explicitly frame the contribution around the row-pseudoinverse parametrization, the revision is local. I would also ask the editor to ensure the convexity claim in Corollary 1 is either proved or withdrawn, as it appears incorrect as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main result is real but narrower than the packaging. The paper is the first, as far as the cited literature goes, to state the equivalence between the complementarity LCQP controller and the CBF-QP controller for first-order sampled-data systems, and the core theorem is correct. Theorem 1 is essentially the KKT/LCP duality for a convex QP, but the identification of the two controller families with the specific parameter mapping in Eq. (41) is new and useful. The proof is clean, and the small numerical check (max error 1.5e-6) matches what the theorem predicts. Credit where due: this gives two communities a concrete dictionary between solvers and robustness results.\n\nThe load-bearing problem is Eq. (15). It asserts J†_{c,i} n_i = (n_i^T J_{c,i})†, which is false in general (a 2x2 diagonal example shows it). The controller in Eq. (15) is therefore not the same as the controller in Eq. (17)-(18), whose columns are the row pseudoinverses. Theorem 1 analyzes the latter. The error is in the derivation, not in the theorem: you can define the complementarity controller directly with G(∂h/∂q), no pseudoinverse identity needed. But as written, the paper gives two different parametrizations and calls them equal. This has to be fixed, and the numerics need to state which one fmincon actually solved.\n\nOther soft spots are minor. Corollary 1 says ALC G(ALC) is positive semi-definite; the matrix is nonsymmetric. It is PSD in the monotone-LCP sense (AG(A)=AA^T D with D positive), so the claim is salvageable, but needs a definition or reference. The scope restriction is more important than the prose admits: the mapping (41) only works for linear class-κ α(s)=s/τ with matched margins. They do state it, but the abstract and conclusion lean on 'unifying' the methodologies. For nonlinear class-κ the two QPs are not equivalent, and the paper should say that in the headline, not just in the proof.\n\nMy take: this deserves a serious referee and can be accepted after revision. The core equivalence is correct and useful; the fix is to remove the false identity, clarify which controller is being solved, and tighten the scope claims.","headline":"Correct core equivalence between LCQP and CBF-QP, but the paper asserts a false pseudoinverse identity that blurs which controller is actually being compared; needs a focused revision.","tokens_in":10574,"tokens_out":6794,"would_cite":true,"duration_ms":64154,"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":"This paper proves that complementarity-based and control-barrier-function quadratic programs produce the same joint-velocity commands for sampled-data first-order whole-body robot control when margins and linear barrier rates are matched.","keywords":["complementarity constraints","control barrier functions","whole-body robot control","collision avoidance","quadratic programming","sampled-data control","convex optimization","equivalence proof"],"falsifier":"Run the two controllers on a sampled-data velocity-controlled robot with a nonlinear class-$\\kappa$ function, for instance $\\alpha(s)=2s$; if the joint-velocity commands still coincide exactly, the linear-$\\alpha$ condition is not actually necessary, while if they differ, the unification holds only inside the stated scope. A more direct check is to solve the two general QPs of Theorem 1 for a small matrix $A$ with nonzero rows, such as $A=\\begin{bmatrix}1&1\\\\1&-1\\end{bmatrix}$ and $b=[1,\\,1]^T$, to machine precision and compare the minimizers, which the theorem predicts are identical.","tokens_in":9556,"feed_emoji":"🤖","tokens_out":10654,"duration_ms":93926,"temperature":0.7,"pith_summary":"This paper proves that two established ways of making a velocity-controlled robot avoid obstacles—linear-complementarity constrained quadratic programs and control-barrier-function quadratic programs—produce the same joint-velocity command at every sample time, provided their safety margins are matched and the barrier decay rate is the linear function $\\alpha(s)=s/\\tau$. The two frameworks grew up separately, yet for sampled-data first-order systems they are not competing designs: each is a minimum-norm correction to a desired velocity, projected onto the same linear safety constraints. The proof goes through a general theorem about quadratic programs: minimizing $\\tfrac{1}{2}\\|x\\|^2$ over $Ax-b\\ge 0$ has the same optimum as minimizing it over the complementarity set $x=G(A)\\lambda$, $0\\le\\lambda\\perp AH\\lambda-b\\ge 0$, when the rows of $A$ are nonzero. A simulation of a 3-DoF planar arm confirms the two controllers trace identical paths to solver precision. If the proof holds, safety margins, sampling-time conditions, and algorithmic improvements can be transferred in both directions between the two frameworks.","feed_headline":"Two robot safety controllers proven to be the same","feed_subtitle":"With matched margins and a linear barrier-decay rate, both QPs issue identical joint-velocity commands at every step.","key_machinery":"The load-bearing object is the row-to-column operator $G$, defined by $G(A)=[a_1^{\\dagger}\\ \\dots\\ a_m^{\\dagger}]$, where $a_i^{\\dagger}=a_i^T/\\|a_i\\|^2$ is the Moore–Penrose inverse of the $i$-th row of $A$. Because these row pseudo-inverses span the row space of $A$, the matrix $H=G(A)$ lets the complementarity variable $\\lambda$ reproduce exactly the direction any KKT multiplier would take, turning the perpendicularity condition into the same complementary-slackness condition as the convex CBF QP. The equality of the two formulations is completed by the parameter mapping $\\alpha_i(x)=x/\\tau$ and $\\delta_{\\mathrm{LC}}=\\delta_{\\mathrm{CBF}}$, which makes the affine constraint $Ax-b\\ge 0$ identical in both formulations; equation (41) states this condition.","core_discovery":"On the paper's own terms, the central discovery is an exact equivalence theorem. For any matrix $A\\in\\mathbb{R}^{m\\times n}$ with nonzero rows and any offset $b$, the convex QP $\\min \\tfrac{1}{2}\\|x\\|^2$ s.t. $Ax-b\\ge 0$ and the complementarity-constrained QP $\\min \\tfrac{1}{2}\\|x\\|^2$ s.t. $x=H\\lambda$, $0\\le\\lambda\\perp AH\\lambda-b\\ge 0$, with $H=G(A)$ built from Moore–Penrose inverses of the rows of $A$, have the same optimal solutions. The proof shows $X_1\\subseteq X_0$, then uses KKT stationarity to show every optimizer of the first problem lies in the row space of $A$, and since $H$ maps $\\mathbb{R}^m$ onto that row space, that optimizer is feasible for the complementarity problem; the complementary-slackness condition lines up with the perpendicularity condition. When the CBF class-$\\kappa$ function is chosen as $\\alpha_i(s)=s/\\tau$ and the margins are equal, the CBF-QP and the LCQP are exactly these two problems, so their solutions coincide. The paper frames the result as a unification: one geometric statement covers both methods, and the active multiplier in each formulation plays the role of the push needed to keep the control inside the safe polyhedron.","pith_inferences":["The KKT proof suggests an extension the paper leaves open: any monotone barrier decay function may correspond to a modified complementarity condition with $\\lambda$ weighted by a nonlinearity, so the unification could survive beyond linear $\\alpha$.","If the equivalence extends to acceleration-level control, complementarity formulations of contact-rich manipulation and higher-order CBFs may turn out to be the same projection with different constraint derivatives; a numerical comparison on a torque-controlled manipulator would test this.","The convexity result implies that reactive whole-body avoidance has a convex solution set under velocity control, which may simplify real-time implementations on resource-constrained robots, though the paper does not test deployment."],"forward_implications":["A complementarity-based controller inherits the sampled-data safety guarantees and Lipschitz-based margin design developed for CBFs, so its physical safety margin and time step can be certified rather than tuned ad hoc.","A CBF-QP gains access to complementarity solvers and to planning formulations that treat avoidance as switched contact-like constraints, which may simplify whole-body planning.","In the matched-parameter setting the two controllers can be swapped freely in deployment: the numerical experiment shows command differences at solver tolerance, with a maximum error of $1.5\\times10^{-6}$.","The convexity corollary means the linear complementarity safe-control problem is convex despite being stated with complementarity constraints, so uniqueness and sufficiency results carry over.","Both frameworks are minimum-norm projections onto one polyhedron, so any algorithmic acceleration for one projection formulation transfers to the other."],"supporting_citations":[{"why":"Supplies the sampled-data control-barrier margin formulation and the discrete invariance condition used in constraint (23).","marker":"[10]"},{"why":"Establishes the control-barrier-function quadratic program that the paper compares with the complementarity controller.","marker":"[8]"},{"why":"States the class-$\\kappa$ condition (21) that renders the safe set forward invariant.","marker":"[9]"},{"why":"Presents the complementarity-based whole-body velocity controller whose solution is being unified.","marker":"[7]"},{"why":"Provides the complementarity switching formulation (11) for velocity-level obstacle avoidance.","marker":"[6]"},{"why":"Gives the complementarity-based kinodynamic planning direction that the unification links to CBF planning.","marker":"[4]"},{"why":"Supplies the KKT conditions used in the proof of Theorem 1.","marker":"[36]"},{"why":"Supports the convexity claim for the linear complementarity problem in Corollary 1.","marker":"[37]"}],"fun_headline_variants":["Two safety controllers, one proof","Same velocities from two safety QPs","One theorem for two safety controllers","CBF equals complementarity for robot safety","Identical commands: two safety methods unified"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The equivalence rests on the control-barrier 'class-$\\kappa$' function—the function that sets how fast the safety margin is enforced—being exactly the linear map $\\alpha(s)=s/\\tau$, and on the two safety margins being equal; with the nonlinear decay functions common in CBF practice, the two QPs are no longer the same problem and the theorem does not cover their commands.","fun_headline_variants_meta":{"raw":{"variants":["Two safety controllers, one proof","Same velocities from two safety QPs","One theorem for two safety controllers","CBF equals complementarity for robot safety","Identical commands: two safety methods unified"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001311,"raw_usage":{"total_tokens":5356,"prompt_tokens":971,"completion_tokens":4385,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":4324}},"tokens_in":587,"tokens_out":4385,"duration_ms":25374,"temperature":1.0,"reasoning_tokens":4324,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:36:16.374811+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two controllers on a sampled-data velocity-controlled robot with a nonlinear class-$\\kappa$ function, for instance $\\alpha(s)=2s$; if the joint-velocity commands still coincide exactly, the linear-$\\alpha$ condition is not actually necessary, while if they differ, the unification holds only inside the stated scope. A more direct check is to solve the two general QPs of Theorem 1 for a small matrix $A$ with nonzero rows, such as $A=\\begin{bmatrix}1&1\\\\1&-1\\end{bmatrix}$ and $b=[1,\\,1]^T$, to machine precision and compare the minimizers, which the theorem predicts are identical.","supporting_citations":[{"cited_title":"Oc3: A reactive velocity level motion planner with complementarity constraint-based obstacle avoid- ance for mobile robots,","cited_arxiv_id":null,"evidence_quote":"Provides the complementarity switching formulation (11) for velocity-level obstacle avoidance."},{"cited_title":"On the synthesis of reactive collision-free whole- body robot motions: A complementarity-based approach,","cited_arxiv_id":null,"evidence_quote":"Presents the complementarity-based whole-body velocity controller whose solution is being unified."},{"cited_title":"Control barrier functions in sampled-data systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the sampled-data control-barrier margin formulation and the discrete invariance condition used in constraint (23)."},{"cited_title":"Complementarity-based dy- namic simulation for kinodynamic motion planning,","cited_arxiv_id":null,"evidence_quote":"Gives the complementarity-based kinodynamic planning direction that the unification links to CBF planning."},{"cited_title":"The linear complimentarity problem, sufficient matrices, and the criss-cross method,","cited_arxiv_id":null,"evidence_quote":"Supports the convexity claim for the linear complementarity problem in Corollary 1."}],"review_version":1}