{"id":"f73e88fa-aa40-4201-afb7-eb096c089b00","arxiv_id":"2411.13506","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"For layered planner-tracker control with Bezier reference trajectories, the set of feasible terminal conditions is a polytope, computable from a linear inequality on Bezier control points.","lead":"This paper introduces Bezier reachable polytopes, a certificate that describes all endpoints a robot's planning layer can request such that the low-level tracking controller still respects state and input limits. The authors show the reachable set can be written as a simple linear inequality, enabling long-horizon tasks to be planned by graph search instead of expensive trajectory optimization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reachable set in Section V is not exact for p>2γ−1: the D† substitution ignores the boundary-map nullspace, so F(x0) is an underapproximation, not the set of all terminal conditions.","rationale":"The reader's weakest assumption is the constructibility of the Lipschitz-derived constants. That concern is about reproducibility and completeness, but it does not directly falsify the existence claim in Theorem 1; missing constants could in principle be supplied. The concern raised here is different and more load-bearing: the Section V claim that F(x0) equals the set of all terminal conditions for which a feasible Bézier curve exists is internally inconsistent with the paper's own allowance of p>2γ−1. The pseudoinverse substitution is only valid when the boundary map is injective (p=2γ−1); otherwise it fixes a particular minimum-norm control-point vector and ignores other control points that also meet the boundary conditions. The scalar example is a clean demonstration: the same terminal condition is feasible through one control-point choice but rejected by the D†-based certificate. This does not invalidate the sufficient-condition direction of Theorem 1, which remains a valid conservative certificate, but it does mean the paper's advertised 'reachable set' is only an underapproximation. Because the overclaim is local and fixable by either restricting to minimal degree or projecting the polyhedron properly, conditional acceptance remains appropriate, but the exactness wording must be corrected. The reader's concerns about missing constants and deferred proof details are also valid, but the nullspace issue is a stronger, more falsifiable problem.","tokens_in":11326,"tokens_out":13727,"duration_ms":149397,"concrete_test":"Implement the scalar counterexample: a single-integrator planning model (γ=1, n=m=1) with Bézier order p=2, boundary x0=0, xT=2, and constraint b1 ≤ −1 on the middle control point. If the D†-based condition rejects xT=2 while the explicit control points (0,−2,2) satisfy both the boundary map and F⃗p ≤ G, the 'all terminal conditions' claim is false. For a broader check, compute the true projection of F⃗p ≤ G onto boundary values via Fourier–Motzkin elimination for a random instance with p>2γ−1 and verify that F D†v ≤ G does not coincide with the projection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V defines F(x0) = {xT | F D† [x0^T xT^T]^T ≤ G} and claims it represents all terminal conditions for which a feasible Bézier curve exists. This is not exact when the Bézier order p > 2γ−1, a case explicitly allowed in Property 3 and Remark 1. The boundary constraints ⃗D ⃗p = v are then underdetermined; the control points are ⃗p = D†v + Ny, where N spans the nullspace of ⃗D. Substituting only the minimum-norm solution D†v into F⃗p ≤ G gives a sufficient condition, but it ignores the y freedom. The true projection of {⃗p | ⃗D ⃗p = v, F⃗p ≤ G} onto v requires eliminating y, not replacing ⃗p with D†v. As a concrete scalar example with γ=1, m=1, p=2, take x0=0, xT=2, and let the inequality be b1 ≤ −1 on the middle control point. The minimum-norm solution is (b0,b1,b2)=(0,0,2), which violates b1 ≤ −1, so the D†-based set omits xT=2. Yet the control points (0,−2,2) satisfy the same boundary conditions and b1 ≤ −1, so a feasible Bézier curve exists. Thus F(x0) is an inner approximation, not the set of all feasible terminal conditions; the strongest reachability claim is false.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Bézier Reachable Polytopes as certificates for layered planner-tracker architectures. The authors show that by enforcing affine inequalities on Bézier control points, one can guarantee that the closed-loop system satisfies state and input constraints over continuous time (Theorem 1). They combine this with boundary-value constraints to define forward and backward reachable polytopes in the planning state space, and demonstrate the approach on pendulum swing-up and on the ARCHER 3D hopping robot. The core mechanism is the convex hull property of Bézier curves, which reduces continuous-time constraint enforcement to checking control-point inequalities.","tokens_in":11650,"tokens_out":6558,"duration_ms":60821,"significance":"The central idea is appealing: if the construction is valid, it provides a computationally efficient, certificate-based interface between a decision-making layer, a planner, and a tracking controller. The convex-hull argument (Property 2) and the sufficient-condition framework are sound and correctly applied. The paper ships code, which is a strength. However, the exactness claim for the reachable polytope in Section V is not correct for the general case allowed in the paper, and the construction depends on Lipschitz constants and a closed-form solution that are not provided. The underlying sufficient-condition certificate remains valuable after revision.","major_comments":[{"comment":"The statement that F(x0) = {xT | F D† [x0^T xT^T]^T ≤ G} represents all terminal conditions for which a feasible Bézier polynomial exists is false when p > 2γ−1, a case explicitly permitted by Property 3 and Remark 1. In this regime the boundary constraint D⃗p = v is underdetermined: the control points are ⃗p = D†v + Ny with N spanning the nullspace of D. Substituting only the minimum-norm solution is sufficient but not necessary. For example, with γ=1, m=1, p=2, x0=0, xT=2 and the constraint b1 ≤ −1, the minimum-norm solution (b0,b1,b2)=(0,0,2) violates the constraint, yet the control points (0,−2,2) satisfy the same boundary conditions and the constraint. Thus F(x0) is an inner approximation, not the exact reachable set. This affects the graph-search edge condition in Section VI.A, which may miss feasible connections. Please either restrict the construction to p = 2γ−1, compute the projection eliminating the nullspace variable y, or explicitly reframe F(x0) as an under-approximation.","section":"Section V, definition of F(x0)"},{"comment":"The constructive claim of Theorem 1 depends on Lipschitz constants LΠ, LΨ, Le, Lk, LG, Lf from Assumption 3 and Lemma 3, and on a 'closed form' solution to the set containment problem whose details are deferred to reference [19], a code repository. The paper does not provide these constants for the pendulum or ARCHER experiments, nor a procedure for computing them, nor a derivation of the closed-form solution. Since the certificate is only useful if the matrices F and G can actually be built, this missing support is load-bearing. Please supply the derivation or a citable reference, and state how the constants are obtained in the experiments.","section":"Lemma 3 and Assumption 3"}],"minor_comments":[{"comment":"The phrase 'is is extremely common' should be corrected to 'is extremely common'.","section":"Introduction"},{"comment":"The notation cM for the projection of M onto the positive semidefinite cone is not defined; a clearer notation such as M+ would help.","section":"Lemma 3 proof"},{"comment":"The line 'Taking bG = bG' appears to be a typo, and the definition of the refinement matrices ⃗Qi should be made explicit so that the step from (16) to (17) is verifiable.","section":"Corollary 1 proof"},{"comment":"The definition of BQ(t) is ambiguous: it is not clear which subinterval [Ti−1, Ti] is used when writing BQ(t) ≜ B(Ti + t/T (Ti+1 − Ti)).","section":"Property 4"}],"recommendation":"major_revision","confidential_remarks":"The exactness issue in Section V is the main obstacle; the paper's contribution can be salvaged by reframing the reachable set as an inner approximation or by restricting to the minimal-order case p = 2γ−1. The missing Lipschitz constants and the reliance on a code repository for a mathematical proof are also important for a journal submission. I would encourage the authors to address these points carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the core construction is useful and mostly correct, but the headline reachability claim is too strong. The paper's contribution is a polytopic certificate: using the convex hull property of Bezier curves, continuous-time state and input constraint satisfaction for a planner-tracker system is reduced to affine inequalities on control points. That idea is sound, and the graph-search interface plus the hopper demo show it can be plugged into a decision layer. Lemma 1 and Lemma 2 are simple and check out; the dependence on [12] for H and Lemma 4 is explicit and reasonable.\n\nThe soft spot is Section V. The set F(x0) is defined as {xT | F D† [x0; xT] ≤ G} and described as all terminal conditions for which a feasible Bezier curve exists. That is only true when p = 2γ−1, so D is square and invertible. Property 3 and Remark 1 explicitly allow p > 2γ−1, where D has a nullspace and control points take the form p = D†v + Ny. Substituting only the minimum-norm solution gives a sufficient condition, not the exact projection. The stress-test example is correct: with γ=1, p=2, x0=0, xT=2, and a constraint on the middle control point, the minimum-norm solution violates the inequality but a feasible curve exists. The paper should either restrict to p=2γ−1 or eliminate the nullspace variables and call the set an inner approximation. As written, the central reachability claim is false in a case the paper itself allows.\n\nSmaller issues: Corollary 1's proof is essentially missing, Lemma 3 defers the closed-form derivation to the repository, and the Lipschitz constants LΠ, LΨ, Le, Lk, LG, Lf needed to build F and G are not reported for the experiments. There is also a sign slip in Lemma 3: the text says cM ⪯ M, but the implication as used requires cM ⪰ M. That one looks like a typo, but it should be fixed.\n\nBottom line: the paper is worth a serious referee, and a good referee will ask for the D† issue and the missing constants. I would not cite it in its current form, but with the reachable-set claim corrected and the constants documented it would be a useful tool for layered motion planning.","headline":"Good certificate idea, but the reachable-set claim in Section V is an inner approximation for p>2γ−1; the paper overstates it.","tokens_in":12221,"tokens_out":5143,"would_cite":false,"duration_ms":46953,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B03","93C85","68T40"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single affine inequality on Bézier control points certifies every trajectory a layered planner-tracker system can safely track, turning long-horizon planning into polytope intersections.","keywords":["Bézier curves","reachable polytopes","layered control architectures","motion planning","tracking certificates","constraint satisfaction","graph search","hopping robot"],"falsifier":"Simulate a Bézier trajectory that passes the certificate on the hopping robot's full hybrid dynamics; if any closed-loop sample violates the state or input bounds, the claimed equivalence between $F\\vec{p}\\leq G$ and constraint satisfaction is false for those constants.","tokens_in":11053,"feed_emoji":"🤖","tokens_out":5501,"duration_ms":57693,"temperature":0.7,"pith_summary":"This paper claims that for a robot controlled by a layered architecture—a planner that emits Bézier polynomial trajectories and a low-level feedback tracker—the set of trajectories the low level can track while respecting state and input limits is certified by one affine inequality on the Bézier control points. If the inequality holds, tracking keeps the closed-loop system inside the safe state set and the actuator commands inside their allowed box for every instant. This turns long-horizon planning into checking linear constraints and intersecting polytopes rather than solving expensive boundary-value problems. The authors demonstrate the idea on a simulated pendulum swing-up and on a physical 3D hopping robot.","feed_headline":"One Bézier check certifies every safe planned trajectory","feed_subtitle":"Passing one inequality guarantees the tracker stays inside state and input limits for the entire plan.","key_machinery":"The central object is the Bézier Reachable Polytope: a polytope in the space of terminal or initial conditions cut out by an affine inequality $F\\vec{p} \\leq G$ on Bézier control points. The construction combines three ingredients: the convex-hull property of Bézier curves, so linear inequalities on control points bound the entire continuous curve; a tracking certificate $E$ that bounds closed-loop error as a function of the planning input; and Lipschitz bounds on $\\Pi$, $\\Psi$, $e$, $k$, $g_d^{-1}$, and $f_d$. Lemma 3 is the technical core: it rewrites the nonlinear norm terms that appear in the state and input constraints as a finite system of linear inequalities, using projection onto the positive semidefinite cone and a closed-form set-containment step.","core_discovery":"The central claim is Theorem 1: given a planning model with a tracking certificate, there exist matrices $F$ and $G$ such that any Bézier curve $B:I\\to X_d$ whose control points satisfy $F\\vec{p} \\leq G$, when tracked, yields $\\Pi(x_{cl}(t))\\in C_X$ and $k(x_{cl}(t),x_d(t))\\in C_U$ for all $t\\in I$. Because boundary conditions enter linearly through the matrix $D$, the set $F(x_0)=\\{x_T \\mid F D^\\dagger [x_0^\\top\\ x_T^\\top]^\\top \\leq G\\}$ of terminal conditions for which a feasible curve exists is a polytope. This gives a constructive certificate of reachability for the combined planner-tracker system.","pith_inferences":["The construction is likely an under-approximation of the true feasible set: the Lipschitz bounding and the single reference point add conservatism, so some feasible Bézier curves may be rejected by $F\\vec{p} \\leq G$; practitioners should treat the polytope as a safety certificate rather than the exact reachable set.","The same affine inequality could be embedded directly as constraints in trajectory optimization or model-predictive control, replacing heuristic penalty terms with a closed-form feasibility oracle.","The derivation relies mainly on the convex-hull property, so a similar certificate may extend to other convex basis parameterizations or B-splines, though the exact linear boundary-condition structure would need to be re-derived.","A natural quantitative test is to measure the actual tracking error on the hopping robot and compare it with the assumed tracking certificate; if the real error exceeds the certificate, the polytope must be recomputed with the measured error bound."],"forward_implications":["Long-horizon plans can be certified without closed-loop simulation: a plan is feasible if its Bézier control points pass the linear check $F\\vec{p} \\leq G$.","The forward and backward reachable sets $F(x_0)$ and $B(x_T)$ are polytopes, so graph-search planners can connect sampled states by checking polytope intersection instead of solving two-point boundary-value problems.","Refining the time interval with piecewise-constant reference points reduces conservatism at the cost of more polytope faces, giving an explicit tightness-versus-computation tradeoff.","The same certificate can serve as a model-predictive layer, since the refined certificate is written as linear constraints around a sequence of reference points.","The input bound $u_{max}$ and horizon $T$ reshape the reachable polytope predictably, so task difficulty is reflected directly in the number of graph-search steps required."],"supporting_citations":[{"why":"Supplies the Bernstein-basis derivative matrix $H$ and the constraint-tightening lemma used to build the state-constraint inequality.","marker":"[12]"},{"why":"Provides the convex-hull and splitting properties of Bézier curves that let control-point inequalities bound the whole continuous curve.","marker":"[18]"},{"why":"Describes the ARCHER 3D hopping robot used for the hardware demonstration of the certificate.","marker":"[20]"}],"fun_headline_variants":["One Bezier polytope checks every safe trajectory","Layered control gets one-shot safety proofs","Bezier reachable sets certify whole-plan safety","Single inequality verifies entire motion plan","Polytope certificates for robust hierarchical plans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The certificate is only as strong as the tracking error bound supplied by the low-level controller and the six smoothness constants the paper assumes without showing how to compute them—if those constants are wrong or missing, the polytope stops certifying anything.","fun_headline_variants_meta":{"raw":{"variants":["One Bezier polytope checks every safe trajectory","Layered control gets one-shot safety proofs","Bezier reachable sets certify whole-plan safety","Single inequality verifies entire motion plan","Polytope certificates for robust hierarchical plans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000785,"raw_usage":{"total_tokens":3408,"prompt_tokens":834,"completion_tokens":2574,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":2506}},"tokens_in":450,"tokens_out":2574,"duration_ms":23420,"temperature":1.0,"reasoning_tokens":2506,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:20:27.535494+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a Bézier trajectory that passes the certificate on the hopping robot's full hybrid dynamics; if any closed-loop sample violates the state or input bounds, the claimed equivalence between $F\\vec{p}\\leq G$ and constraint satisfaction is false for those constants.","supporting_citations":[{"cited_title":"Multi-Rate Planning and Control of Uncertain Nonlinear Systems: Model Predictive Control and Control Lyapunov Functions","cited_arxiv_id":"2204.00152","evidence_quote":"Supplies the Bernstein-basis derivative matrix $H$ and the constraint-tightening lemma used to build the state-constraint inequality."},{"cited_title":"A primer on bézier curves,","cited_arxiv_id":null,"evidence_quote":"Provides the convex-hull and splitting properties of Bézier curves that let control-point inequalities bound the whole continuous curve."},{"cited_title":"Creating ARCHER: A 3D Hopping Robot with Fly- wheels for Attitude Control,","cited_arxiv_id":null,"evidence_quote":"Describes the ARCHER 3D hopping robot used for the hardware demonstration of the certificate."}],"review_version":1}