{"id":"729ad43f-3f0b-4ac0-9fa7-760feb10cc25","arxiv_id":"2507.09134","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new governor module lets path planners and nonlinear MPC work together with guaranteed recursive feasibility and an expanded region of attraction, validated in quadrotor simulation.","lead":"This paper introduces the Path Feasibility Governor, a module that selects intermediate reference points along a planned path so that an MPC controller always has a feasible plan, and it proves safety and stability guarantees. It matters because it could let robots use shorter, computationally cheaper prediction horizons in cluttered environments while keeping formal guarantees.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proofs cover a fixed-constraint MPC, but the implemented and simulated controller re-linearizes obstacle constraints at every step; the stability argument in Appendix B is not established for this time-varying OCP.","rationale":"The reader's weakest-assumption entry focuses on Assumption 3 and the strict path condition, which is a real practical precondition but is at least explicit in the assumptions. The more damaging gap, also noted in the reader's rationale, is that the implemented successive linearization is not the object of the proof. I checked whether the recursive-feasibility part of the shift argument could be repaired using Lemma 2B, and it likely can; however, the ISS/Lyapunov part used for convergence is substantially harder and is not addressed. This does not disprove the nominal theory, which is internally plausible, but it makes the paper's validation conditional. The reader's CONDITIONAL verdict remains appropriate, so I do not recommend moving the verdict. I disagree in emphasis only: I would put the implementation/proof mismatch ahead of Assumption 3 as the primary load-bearing issue.","tokens_in":21872,"tokens_out":24415,"duration_ms":346595,"concrete_test":"Instrument the existing MATLAB/Simulink loop to record the optimal cost J_k and the one-step difference J_{k+1}-J_k, and verify the ISS inequality (66) with a fixed K-class function along every logged trajectory in Section VI. Then re-run the same scenario with the exact nonlinear OCP (4) (using X = C ∩ ∩_j O_j^c, or a fixed linearization about the planned path instead of the previous trajectory) under the same PathFG update. If the successively linearized implementation violates the ISS estimate, or if the fixed-constraint implementation behaves materially differently, the simulation does not support the theorem as stated and the paper must either prove the time-varying case or restrict its claims.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Theorems 1 and 2 proofs, and particularly Lemma 4 in Appendix B, are written for the OCP (4) with one fixed constraint set X and a fixed value function J(x,r). Section V.A replaces X by the stage- and step-dependent linearizations Xi|k in (37)/(53), which are re-computed at every MPC step from the previous predicted trajectory. The feasible set and the optimal cost J_k therefore change at every time step. Recursive feasibility of a shifted candidate may be salvageable because the shifted states are exactly the next linearization points (Lemma 2B), but the ISS decrease inequality (66), the uniform continuity estimate (65), and the asymptotic-gain argument (67) all rely on a single time-invariant J. No corresponding ISS or Lyapunov argument is given for the successively linearized OCP. Consequently, the controller simulated in Section VI is not the controller covered by the theorems, so the numerical demonstration does not validate the claimed safety and asymptotic-stability guarantees. This is load-bearing because the central claim is a guarantee for PathFG+MPC, and the only evidence that the practical implementation inherits the guarantee is the simulation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Path Feasibility Governor (PathFG), an add-on module that integrates a path planner with a nonlinear MPC controller. Given a continuous path p(s) from the initial state to a target reference, PathFG selects, at each timestep, the furthest auxiliary reference s along the path for which the previous optimal terminal predicted state lies in the terminal set, thereby maintaining MPC feasibility while the path parameter progresses to 1. The paper proves recursive feasibility, constraint satisfaction, and asymptotic stability of the combined PathFG+MPC closed loop under Assumptions 1–3, with a region of attraction equal to the set of states connectable to the target by a feasible path. It then specializes the construction to linear systems with convexified obstacle constraints, gives a Lyapunov-based terminal set, and reports quadrotor navigation simulations with RRT* and potential-field planners.","tokens_in":22116,"tokens_out":7747,"duration_ms":94901,"significance":"If the theoretical claims hold, PathFG is a valuable modular contribution: it extends the feasibility-governor idea from convex to non-convex settings, significantly enlarges the MPC region of attraction without long horizons, and keeps the PathFG update as a cheap one-dimensional search. The paper is clearly written, the architecture is appealing, and the simulation comparison between short-horizon governed MPC and long-horizon ungoverned MPC is informative. The main proof strategy—using ISS of the tracking error and finite-time convergence of the path parameter—is sound for the fixed-constraint OCP considered in Section IV. However, the implementation in Section V uses successively re-linearized obstacle constraints, and the paper does not bridge the gap between that time-varying OCP and the theorems. The strength of the central claim therefore currently exceeds what is established.","major_comments":[{"comment":"The theoretical guarantees are proven for the OCP (4) with a fixed constraint set X, while the controller implemented in Section V solves a sequence of OCPs in which the constraint sets Xi|k are re-linearized around the previous optimal trajectory and hence change at every timestep. Lemma 4's ISS argument uses a single time-invariant value function J(x,s) in (64)–(66), and Lemmas 5–8 inherit that structure; no ISS or Lyapunov argument is given for the time-varying feasible sets in (37)/(53). Moreover, the shifted candidate in the proof of Theorem 1 (Eq. (60)) only guarantees feasibility for a fixed X; under (37) the terminal step i=N refers to the undefined prediction ξ*_{N+1|k-1}, so the terminal constraint and the terminal set T from Section V.B require a separate invariance argument that is not supplied. The numerical example therefore does not validate the proved guarantees. The authors should either prove the corresponding safety/stability results for the successively linearized OCP, or explicitly restrict the formal claims to the fixed-constraint formulation and present the simulation as a heuristic demonstration.","section":"§V.A (Eqs. (37), (53)) vs. §IV (Appendix A, Lemmas 4–8)"},{"comment":"Assumption 3 requires the planner to return a continuous path p:[0,1]→Rε such that every point p(s) is strictly steady-state admissible and (x0,p(0))∈Γ. This is a strong condition on the planner output, and the paper gives no procedure for verifying it. RRT* returns a discrete collision-free path whose interpolated segments are not guaranteed to maintain a positive margin from obstacles, and potential-field paths are typically generated as finite waypoint sequences and are not automatically continuous or strictly inside Rε. The simulation section does not report how the planner outputs were checked against Assumption 3 (e.g., minimum clearance, continuous interpolation, or post-processing). Since Theorems 1–2 and the set Dx(r) are defined through existence of a path satisfying (12), this unverified condition is load-bearing. Please add a verification/post-processing step and document its output, or state and prove a relaxed condition that standard planners can satisfy.","section":"Assumption 3 and §VI.E"},{"comment":"The proof of Lemma 5 relies on uniform continuity of the optimal final predicted state ξ*_N via Assumption 2E, but Assumption 2E (Lipschitz continuity of the optimal solution map ζ*) is not established for the general nonlinear OCP (4). The citation to [45, Section 7.1] concerns time-distributed optimization and does not by itself prove this regularity for the exact solution map; for nonconvex constraints the argmin may be discontinuous or set-valued even when Assumptions 2A–2D hold. Lemma 1 establishes Assumption 2E only for the linear-quadratic case through [52]. The general statements of Theorems 1–2 are therefore conditional on an unproven regularity property. Please provide a proof or precise sufficient conditions for Assumption 2E for problem (4), or reformulate the main theorems for the linear case in which Lemma 1 applies.","section":"Assumption 2E and Lemma 5"}],"minor_comments":[{"comment":"The region of attraction is stated as (Dx(r)×[0,1])∩Γ, but in the path-parameter coordinates the relevant feasible set is Γ̃ defined in (17); the tilde is missing in the theorem statement.","section":"Theorem 2 statement"},{"comment":"The indexing uses ξ*_{i+1|k-1} for i=N, where ξ*_{N+1|k-1} is undefined; the index should be clamped or the notation should define ξ*_{N+1|k-1} explicitly.","section":"Eqs. (37) and (53)"},{"comment":"The legend in subfigure 6 labels 'PathFG+MPC(N=5)' and 'PathFG(N=5)', which makes it difficult to distinguish the combined computation time from the PathFG-only overhead; separate axes or clearer annotations would help.","section":"Figure 6"},{"comment":"The simulation section does not report planner parameters, number of runs, or obstacle-clearance statistics, so the claim of consistent performance across different planners is not quantitatively supported.","section":"§VI.E"},{"comment":"The numerical example validates the controller on a linearized quadrotor model; the abstract's reference to nonlinear MPC is supported by the general formulation but not by this particular demonstration.","section":"§VI.A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is closely related to the authors' prior feasibility-governor line of work and to [36], but the extension to non-convex path planning is a meaningful step. The main gap—between the fixed-constraint proofs and the successively linearized implementation—is substantial but fixable by either extending the proofs or narrowing the claims. The paper fits the scope of eess.SY well, and the related-work discussion is adequate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the PathFG idea is genuinely nice, but the proofs cover a controller that isn't the one the simulation runs. The gap is load-bearing, not cosmetic.\n\nWhat's new: using the previous MPC terminal state to select the next reference along a path via terminal-set membership is a clean trick. It extends the feasibility governor line from convex reference sets to non-convex paths, and the conceptual claim—that the region of attraction becomes the entire path-connected set—is attractive. For the fixed-constraint OCP (4), the proofs in Section IV and the appendices look plausible. The simulation also does real work: with PathFG, N=5 and N=15 run well below the 0.1s sampling period, while ungoverned MPC needs N=50 and can't run in real time. That is a useful practical demonstration.\n\nThe problem is in the implementation section. The OCP actually solved online, (31) plus (37)/(53), replaces the fixed state constraint X with linearizations X_{i|k} computed from the previous predicted trajectory. These sets change at every time step. Theorems 1 and 2, and especially the ISS argument in Lemma 4, assume a fixed constraint set and a fixed value function J. The decrease inequality (66) and the uniform continuity bound (65) have no analog for the time-varying OCP. So the simulated controller is not the controller covered by the theorems. The paper doesn't acknowledge this gap. A shift argument for recursive feasibility might be repairable because the shifted states are exactly the next linearization points (Lemma 2B), but the stability proof, including the finite-time convergence of s, relies on the invariant value function. No ISS or Lyapunov argument is given for the successively linearized OCP. So the numerical results, while promising, don't validate the claimed safety and asymptotic-stability guarantees.\n\nSecond concern: Assumption 3 asks for a path entirely in R_ε with (x0,p(0)) ∈ Γ. RRT* and potential fields don't come with that guarantee, and the paper doesn't report how the paths in the simulation were checked. This is a real gap between assumptions and practice, though secondary to the first.\n\nThe paper is not circular; it derives results from assumptions, and the simulation is a demonstration. Related-work coverage is fine, and the writing is clear. The central defect is the theory/practice mismatch.\n\nWho should read this: researchers working on MPC and motion-planning integration. It deserves a serious referee, but the revision needs to be substantial. Either prove recursive feasibility and stability for the linearized OCP, or present the linearized version explicitly as a practical heuristic and support that claim. Also, address how Assumption 3 is met in practice.\n\nRecommendation: send to peer review, major revisions.","headline":"Nice idea, but the proofs cover the fixed-constraint MPC while the simulated implementation uses trajectory-dependent linearized constraints, so the central guarantee is unproven for the actual controller.","tokens_in":22611,"tokens_out":7478,"would_cite":true,"duration_ms":81627,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C55","93D20","93D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"A small add-on unit, the Path Feasibility Governor, filters a planned path into feasible intermediate references for MPC and is proven to keep the closed loop safe, feasible, and asymptotically stable from any state connected to the…","keywords":["path feasibility governor","model predictive control","recursive feasibility","region of attraction","path planning","constraint satisfaction","asymptotic stability","quadrotor navigation"],"falsifier":"Feed a deliberately borderline path into the quadrotor setup, one whose waypoints graze the inflated obstacle boundary so that some equilibrium $\\bar{x}_{p(s)}$ lies on the boundary of $X$ or $U$ rather than strictly inside $\\mathcal{R}_\\varepsilon$, and run the closed loop; if the MPC problem becomes infeasible at any step, or if the observed increments of $s_k$ are not bounded below by a positive constant across runs, the premises of Theorems 1 and 2 are not satisfied. A simpler check: sample a typical RRT* or potential-field path in the paper's environment and test the strict margin $p(s)\\in\\mathcal{R}_\\varepsilon$ pointwise.","tokens_in":21674,"feed_emoji":"🚁","tokens_out":10796,"duration_ms":104827,"temperature":0.7,"pith_summary":"The paper aims to close the gap between geometric path planning and constrained predictive control: planners produce collision-free paths that ignore dynamics, while a model predictive controller (MPC) can only track references it can reach within its prediction horizon. The proposed Path Feasibility Governor (PathFG) is a small add-on unit that slides an auxiliary reference $s\\in[0,1]$ along the planned path, always selecting the furthest point the MPC can feasibly reach in the remaining steps. The central claim is that this one-dimensional filter guarantees recursive feasibility, constraint satisfaction, and asymptotic stability from every initial state that can be connected to the target by a continuous, strictly admissible path, expanding the region of attraction from the $N$-step backward reachable set to the full path-connected set. A practical consequence is that prediction horizons can be kept short and cheap because the planner, not the horizon, handles the non-convex obstacles.","feed_headline":"Governor widens MPC's safe zone to every path-connected goal","feed_subtitle":"A cheap scalar filter feeds MPC reachable waypoints, so short horizons stay feasible and stability is proven.","key_machinery":"The central object is the auxiliary-reference selection map $g(\\xi^*_N)=\\max\\{s\\in[0,1]\\mid(\\xi^*_N,s)\\in\\tilde{T}\\}$, where $\\tilde{T}=\\{(x,s)\\mid(x,p(s))\\in T\\}$ is the terminal set sliced along the path and $\\xi^*_N$ is the optimal final predicted state of the previous MPC solution. Because that state was reachable in $N$ steps from the previous state, it is reachable in $N-1$ steps from the current state, so pulling the reference as far forward as the terminal set allows makes recursive feasibility hold by construction. The companion pieces are the Lyapunov-based terminal set $T=\\{(x,r)\\mid\\Delta(x,r)\\leq 0\\}$ built from the explicit-reference-governor threshold construction, and the input-to-state stability of the MPC closed loop with respect to reference changes (Lemma 4), which together imply that whenever the predicted terminal state gets close to the path equilibrium the reference jumps at least a fixed margin $\\alpha$ forward (Lemma 6), so $s_k$ reaches $1$ in finite time (Lemma 8).","core_discovery":"The paper establishes that a scalar auxiliary reference moving along a planner-supplied path can carry all the non-convexity of the motion planning problem, leaving the MPC with only the convex tracking task. Given a target $r\\in\\mathcal{R}_\\varepsilon$ and a continuous path $p:[0,1]\\to\\mathcal{R}_\\varepsilon$ with $p(1)=r$ whose start is MPC-feasible, the PathFG update $s_k=g(\\xi^*_{N|k-1})=\\max\\{s\\in[0,1]\\mid(\\xi^*_{N|k-1},s)\\in\\tilde{T}\\}$ selects the next reference, and Theorem 1 proves the closed loop satisfies $(x_k,s_k)\\in\\tilde{\\Gamma}$, so state and input constraints hold for all $k$. Theorem 2 then proves that the equilibrium $(\\bar{x}_r,1)$ is asymptotically stable with region of attraction $(D_x(r)\\times[0,1])\\cap\\tilde{\\Gamma}$, where $D_x(r)$ is the set of all initial states from which a feasible path to the target exists; the proof combines input-to-state stability of the MPC policy with respect to reference increments with a minimal-jump lemma showing $s_k$ advances by at least a fixed amount whenever the predicted terminal state is close to the path, yielding finite-time convergence of $s_k$ to $1$. The numerical study shows a quadrotor navigating a cluttered environment with prediction horizon $N=5$ where ungoverned MPC requires $N\\geq 20$, at computation times near $0.01$ s, well below the $0.1$ s sampling period.","pith_inferences":["A testable extension the paper leaves implicit: add a post-planner filter that projects the planned path inward by the margin $\\varepsilon$, or certifies pointwise membership in $\\mathcal{R}_\\varepsilon$, before PathFG engages, so that Assumption 3 becomes checkable rather than assumed.","The proof structure only needs a uniform minimum jump $\\alpha$ along the path and continuity of $p$, so the same machinery should extend to piecewise-smooth paths, to replanning with moving targets, and to robust MPC paired with a robust invariant terminal set, since the ISS and terminal-set arguments carry over.","The simulation data show a performance trade-off the authors frame mildly: ungoverned long-horizon MPC converges faster but exceeds the real-time budget, so a practical reading is that PathFG is the enabler of short-horizon operation rather than a faster tracker."],"forward_implications":["The region of attraction of the closed loop grows from the $N$-step backward reachable set of the terminal set to $D_x(r)$, the set of all initial states from which any feasible path to the target exists.","Prediction horizons can be drastically shortened: in the quadrotor study, PathFG+MPC with $N=5$ navigates an obstacle-dense environment where ungoverned MPC needs $N\\geq 20$, keeping total computation around $0.01$ s per step, below the $0.1$ s sampling period.","Feasibility is recursive by construction: if $(x_0,p(0))\\in\\Gamma$, then $(x_k,s_k)\\in\\tilde{\\Gamma}$ for all $k$, so constraint satisfaction holds at every time step without re-checking.","The auxiliary reference reaches $s=1$ in finite time, after which the loop behaves as standard MPC tracking the target reference.","Because the planner is decoupled from the controller, any planner satisfying the path condition can be swapped in, and replanning in dynamic environments only requires refreshing the path."],"supporting_citations":[{"why":"The feasibility-governor scheme for linear MPC that PathFG extends from convex to non-convex search spaces.","marker":"[41]"},{"why":"Supplies the explicit-reference-governor Lyapunov threshold computation used to build the terminal set (38)-(41).","marker":"[35]"},{"why":"Provides the Lyapunov-based terminal set construction adopted in Section V-B and the Crazyflie dynamic model used in the simulations.","marker":"[36]"},{"why":"Establishes the standard fact that the MPC cost is a Lyapunov function for the closed loop, which Lemma 4's ISS proof relies on.","marker":"[49]"},{"why":"Its Theorem 4.4.2 is the base result ensuring MPC safety and stabilization under Assumptions 2A-2D.","marker":"[10]"},{"why":"The RRT* planner used in the simulations to demonstrate compatibility with an optimal sampling-based planner.","marker":"[47]"},{"why":"The potential-field planner used in the simulations to demonstrate compatibility with a suboptimal planner.","marker":"[48]"}],"fun_headline_variants":["A single scalar keeps MPC safe with short horizons","Scalar governor gives MPC full path safety, tiny horizon","Path Feasibility Governor: short-horizon MPC, proven safe","One filter step: MPC stability expanded to all paths","Governor makes MPC safe without long prediction horizons"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire guarantee rests on the planner returning a continuous path whose every point is a strictly admissible reference (equilibrium strictly inside the state and input constraints) and whose start is already feasible for the MPC, a strict-margin condition that standard planners do not guarantee and that the paper gives no procedure for verifying.","fun_headline_variants_meta":{"raw":{"variants":["A single scalar keeps MPC safe with short horizons","Scalar governor gives MPC full path safety, tiny horizon","Path Feasibility Governor: short-horizon MPC, proven safe","One filter step: MPC stability expanded to all paths","Governor makes MPC safe without long prediction horizons"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1739,"prompt_tokens":1034,"completion_tokens":705,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":626}},"tokens_in":650,"tokens_out":705,"duration_ms":8779,"temperature":1.0,"reasoning_tokens":626,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:04:36.025130+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed a deliberately borderline path into the quadrotor setup, one whose waypoints graze the inflated obstacle boundary so that some equilibrium $\\bar{x}_{p(s)}$ lies on the boundary of $X$ or $U$ rather than strictly inside $\\mathcal{R}_\\varepsilon$, and run the closed loop; if the MPC problem becomes infeasible at any step, or if the observed increments of $s_k$ are not bounded below by a positive constant across runs, the premises of Theorems 1 and 2 are not satisfied. A simpler check: sample a typical RRT* or potential-field path in the paper's environment and test the strict margin $p(s)\\in\\mathcal{R}_\\varepsilon$ pointwise.","supporting_citations":[{"cited_title":"Governor-parameterized barrier function for safe output tracking with locally sensed constraints,","cited_arxiv_id":null,"evidence_quote":"The feasibility-governor scheme for linear MPC that PathFG extends from convex to non-convex search spaces."},{"cited_title":"The explicit reference governor: A general framework for the closed-form control of constrained nonlinear systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the explicit-reference-governor Lyapunov threshold computation used to build the terminal set (38)-(41)."},{"cited_title":"Automated driving: Safe motion planning using positively invariant sets,","cited_arxiv_id":null,"evidence_quote":"Provides the Lyapunov-based terminal set construction adopted in Section V-B and the Crazyflie dynamic model used in the simulations."},{"cited_title":"Time- distributed optimization for real-time model predictive control: Stability, robustness, and constraint satisfaction,","cited_arxiv_id":null,"evidence_quote":"Establishes the standard fact that the MPC cost is a Lyapunov function for the closed loop, which Lemma 4's ISS proof relies on."},{"cited_title":"Motion planning for mobile manipula- tors—a systematic review,","cited_arxiv_id":null,"evidence_quote":"Its Theorem 4.4.2 is the base result ensuring MPC safety and stabilization under Assumptions 2A-2D."},{"cited_title":"Model predictive control for tracking using artificial references: Fundamentals, recent results and practical implementation,","cited_arxiv_id":null,"evidence_quote":"The RRT* planner used in the simulations to demonstrate compatibility with an optimal sampling-based planner."},{"cited_title":"Kouvaritakis and M","cited_arxiv_id":null,"evidence_quote":"The potential-field planner used in the simulations to demonstrate compatibility with a suboptimal planner."}],"review_version":1}