{"id":"379c5d27-a18f-45d8-8fda-dbec8622c842","arxiv_id":"2505.11941","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A control barrier function for safe navigation is synthesized online by solving Laplace's equation on a local occupancy grid map with obstacle and safety-region boundary conditions.","lead":"This paper builds control barrier functions for robot safety by solving Laplace's heat equation on a local occupancy grid map, turning obstacle avoidance into a fast linear algebra problem. The method computes safety values in about nine milliseconds on a 200 by 200 grid and is demonstrated in simulation and on two TurtleBot robots.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The truncated local OGM breaks the formal safety claim: h=b on the map edge ignores obstacles just outside the sensor range, and no error bound links the local harmonic solution to the global one. This makes the 'safe in unknown environments' claim unsupported, though experiments may still work.","rationale":"I read the paper as claiming that a single harmonic CBF computed from a local OGM provides safe navigation with formal guarantees. The strongest supporting evidence is the clear PDE construction, the sparse linear solver (Table I shows ~10 ms synthesis), and successful simulation and real-world runs. The load-bearing weakness is the leap from the global Dirichlet problem in Section IV-A to the finite local window used in Algorithm 1. This is not a matter of disagreeing with a consensus; it is an internal gap: the propositions use ∂W, while the implementation uses the map edge. The numerical solution also discretizes the Laplacian, but the more serious issue is the artificial h=b boundary. Unless one assumes obstacles are sparse enough to enter the OGM before they can be hit, no safety guarantee follows. The proposed test settles this by checking whether the method collides with an obstacle just beyond the sensing range. The reader's weakest_assumption identified the same issue, and their CONDITIONAL verdict is the right one; I would not move it because the concern is addressable and the experiments are promising. I would, however, insist on either an assumption about sensing horizon and stopping distance or a bound on the local-global mismatch before claiming formal safety. A secondary unresolved point is the unicycle mapping in Eq. (19), which is asserted rather than derived, but the local-window mismatch is more fundamental.","tokens_in":12376,"tokens_out":7434,"duration_ms":84855,"concrete_test":"Run the Gazebo/TurtleBot pipeline with the goal directly behind a 0.2 m obstacle placed just outside the robot's 2 m×2 m sensing window, specifically at 1.05 m from the robot center along the nominal path, with max speed 0.15 m/s. If the robot collides before the obstacle enters the OGM or before the CBF constraint activates, the finite-window assumption is falsified. As an analytical check, compute the global harmonic solution on the full 3 m×3 m arena with that obstacle and compare its sup-norm difference from the local-grid solution over the region explored; if the difference is large near the map edge, the local h is not a valid proxy for the global CBF.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires the SSTF-CBF h to be a valid CBF over the actual workspace. Section IV-A formalizes this with Dirichlet data (10): h=-a on ∂O and h=b on ∂S∪∂W, where ∂W is the true workspace boundary. Algorithm 1, however, constructs the same problem on a 200×200 local OGM centered at the robot and, per lines 9-13, takes the finite map edge as the outer boundary, assigning it the 'safe' value b. Obstacles beyond the local map are simply absent. Consequently Propositions 1 and 3 prove safety for the exact global harmonic function, not for the truncated local-grid function used in the experiments. Because h rises to b at the artificial boundary, the CBF gradient can point outward from the sensed region and the QP (16) may actively direct the robot toward an obstacle that has not yet entered the OGM. The no-obstacle branch (Algorithm 1 lines 5-7) even applies unom unfiltered, so the robot can be driven straight into a just-outside-range obstacle. No theorem bounds the mismatch between the local and global harmonic solutions, and none of the experiments appears to place obstacles at the sensing boundary. Thus the paper's formal safety guarantee does not cover the actual online implementation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SSTF-CBF, a method for online synthesis of control barrier functions from local occupancy grid maps. The safety function h is obtained by solving Laplace's equation over the grid with Dirichlet boundary conditions: h=-a on obstacle boundaries and h=b on the safety-region and workspace boundaries. The resulting h is used to filter a nominal velocity controller by a quadratic program with the single-integrator CBF constraint. The authors claim that this construction yields a valid CBF (Proposition 1), that critical points of h are isolated (Proposition 2), and that safety is maintained even if the robot enters the zero-sublevel set through a critical boundary point (Proposition 3). Experiments in Gazebo and with two real TurtleBots show millisecond-level synthesis times on 200x200 local OGMs. The computational claim is supported by the sparse linear-system formulation and by Table I.","tokens_in":12631,"tokens_out":9899,"duration_ms":100407,"significance":"The core idea of using the steady-state heat equation to turn a local occupancy map into a smooth safety field is genuinely attractive, and the sparse-matrix formulation is a sensible alternative to SDF-based smoothing or per-obstacle ellipsoidal CBFs. The experiments are informative and the reported computation times support the real-time feasibility claim. I also credit the authors for a construction that is not fitted to make the experiments pass; the parameters a, b, delta, and alpha are user-chosen design parameters. However, the formal safety guarantees stated in Propositions 1 and 3 are proved for an idealized global harmonic solution, not for the truncated local-grid function actually used in Algorithm 1, and the transfer from the single-integrator model to the unicycle experiments is asserted rather than proved. These gaps are load-bearing because the title and abstract promise safe navigation in unknown environments. The underlying approach is promising and the gaps appear addressable in principle, but the manuscript in its current form does not establish its central formal claim.","major_comments":[{"comment":"The theoretical results in Propositions 1 and 3 concern the harmonic function defined over the full workspace W with Dirichlet data on the true obstacle boundary ∂O and the true outer boundary ∂W. Algorithm 1, however, constructs the same boundary-value problem on a 200x200 local occupancy grid and, per lines 9-13, assigns the artificial map edge the 'safe' value b. Obstacles beyond the current sensor range simply do not appear in the boundary data, and no theorem or error bound relates the local finite-difference solution to the global harmonic solution. Consequently, the formal guarantee established in Propositions 1 and 3 does not cover the online implementation. This is not merely a theoretical nicety: in the no-obstacle branch (Algorithm 1, lines 5-7) the nominal command is executed unfiltered, so a goal-directed controller can drive the robot toward an obstacle that has not yet entered the sensing range. None of the experiments appears to place obstacles at the sensing boundary, so this regime is untested.","section":"§IV-A and Algorithm 1, Eq. (10)"},{"comment":"The safety filter is designed for the single-integrator model (15), but the experiments use the unicycle model (18). The near-identity transformation (19) is applied to velocity commands, not to the state, and with p defined as the robot center the resulting position dynamics are \\dot p = [cos θ; sin θ](cos θ v_x + sin θ v_y), which is not \\dot p = [v_x; v_y]ᵀ. Therefore the single-integrator condition L_f h + L_g h u ≥ -α(h) does not automatically imply the corresponding condition for the actual unicycle center dynamics. If p is instead intended to be the offset point used in the Robotarium reference [37], then h is evaluated at that offset point and additional reasoning is needed to conclude safety of the robot center. The paper asserts that safety transfers to the unicycle model but provides no proof; this is a load-bearing gap because the experiments and the claimed experimental safety rely on the unicycle implementation.","section":"§V-A, Eq. (19)"},{"comment":"The proof of Proposition 1 reduces to the observation that u=0 is admissible in the single-integrator model. For any continuously differentiable function h with 0 in the input constraint set U, condition (6) is satisfied on the safe set C because 0 ≥ -α(h(p)) whenever h(p) ≥ 0. Thus Proposition 1 does not distinguish the proposed SSTF-CBF from any smooth function; the Laplace construction is not needed for the CBF condition itself. The meaningful safety-relevant ingredient is the direction and magnitude of ∇h in the QP constraint (16), and the paper should present the non-vacuous invariance argument for that constraint directly. As written, the theoretical contribution of Section IV-B is overstated.","section":"§IV-B, Proposition 1"},{"comment":"The proof of Proposition 3 is incomplete. The quantity vmax is defined as the maximum safety value among critical points in Int(I), but no argument is given that such critical points exist; if there are none, the proof cannot proceed. More seriously, the QP (16) is infeasible at any state with h(p)<0 and ∂h/∂p=0, because the constraint becomes 0 ≥ -α(h(p)) with -α(h(p))>0. Proposition 3 does not rule out reaching such a state, and the statement that the robot encounters no critical points between p0 and p(t2) is asserted without proof. The claim that the robot 'will not enter the obstacle region' before exiting Uc is therefore not established by the given argument.","section":"§IV-B, Proposition 3"}],"minor_comments":[{"comment":"The notation for grid values is inconsistent: the text uses both h_i,j and hij for the same quantity; please unify.","section":"§IV-A, Eq. (12)"},{"comment":"Several typos should be corrected, including 'yeilds' in §IV-A, 'simlified' in §III, and 'the CBFs needs' in the Introduction.","section":"Abstract / Introduction"},{"comment":"The text says a target is considered reached when the robot is within 0.5 cm of the destination; given the 3 m x 3 m arena this is likely a typo for 0.5 m. Please clarify.","section":"§V-B"},{"comment":"The caption states 'region O and ∂O are identical,' which is confusing because O is a set of obstacle cells and ∂O is its boundary; please rephrase or explain the discretized convention.","section":"§IV-A, Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript promises formal safety guarantees that the current proofs do not actually deliver for the implemented local-grid, unicycle-based system. The underlying PDE construction is elegant and the computational experiments are valuable, so I do not recommend rejection. The main gaps—local truncation, unicycle transfer, and the incompleteness of Proposition 3—are identifiable and potentially fixable with additional assumptions and proofs, but they require substantial revision rather than local edits."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Know this paper for two things: a clean way to turn a local occupancy grid into a single smooth safety function by solving Laplace's equation, and a formal guarantee that stops short of the actual algorithm. The experiments are real and the timings are credible.\n\nWhat's genuinely new is the combination: solve the heat equation on a 200x200 local OGM with a sparse iterative solver, then use that as the CBF for the whole workspace. That gives you one smooth safety function for arbitrary obstacle shapes, which is a real advantage over ellipse-based or SDF-based constructions. The authors know the related literature and the writing is clear. The experiments include both Gazebo and real TurtleBots, and showing ~10ms synthesis online is a legitimate contribution.\n\nThe soft spots are concentrated in the gap between the math and the implementation. The main theorem proves safety for the harmonic function on the full workspace with true obstacle boundaries. Algorithm 1, however, solves the same PDE on a truncated local map whose outer edge is artificially assigned the safe value b. Obstacles just outside the sensor range are simply absent from that boundary-value problem, and there is no error bound linking the local solution to the global one. In the worst case, the artificial boundary condition can actually create a gradient that steers the robot toward an obstacle that hasn't entered the map yet. So the 'safe in unknown environments' claim is wider than what the mathematics supports. That's the kind of thing a serious revision can fix—e.g., by adding a pessimistic buffer at the map edge or by analyzing the harmonic extension—but the current paper does not address it.\n\nSecond, the unicycle mapping in Eq. (19) is not a near-identity diffeomorphism in the usual sense: the commanded [vx,vy] gets projected onto the heading direction, so the single-integrator CBF guarantee does not automatically transfer. The experiments work, but that is empirical evidence, not a proof. Proposition 3 is also written in a way that is hard to follow; the vmax machinery seems unnecessary and it quietly assumes the path encounters no critical points. I think the conclusion can be recovered more directly—inside the unsafe set, the barrier constraint forces h nondecreasing—so this is a presentation problem more than a fatal flaw.\n\nThird, the experimental section has no baselines and no error bars. The timing numbers are nice, but without comparison to SDF or ellipse methods the practical claims stay suggestive. No code is released either.\n\nThe paper deserves a serious referee. It is an honest engineering effort with a promising idea and a gap that can be closed. I would bring it to a reading group and would cite it if I were working on sensor-based CBF synthesis, but I would cite it with a caveat about the truncation issue.","headline":"A useful engineering paper for real-time CBF synthesis from local grids, but the proof covers an exact global harmonic function rather than the truncated local map actually used, and the unicycle transfer is asserted, not proved.","tokens_in":13184,"tokens_out":5253,"would_cite":true,"duration_ms":56124,"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 claims that solving Laplace's equation over a local occupancy grid map produces a control barrier function that keeps a robot safe in unknown environments and can be recomputed in about 9 milliseconds.","keywords":["control barrier function","occupancy grid map","Laplace's equation","safe navigation","online synthesis","harmonic function","sparse linear solver","mobile robot safety"],"falsifier":"Run the controller in an arena with a thin pole placed just outside the robot's 200 by 200 local map in the direction of travel, so it is not yet in the occupancy grid; if the robot does not brake or deviate before the map expands to reveal the pole, the online safety claim is disproved.","tokens_in":12170,"feed_emoji":"🤖","tokens_out":5523,"duration_ms":52970,"temperature":0.7,"pith_summary":"This paper claims that a safe navigation command can be synthesized directly from a local occupancy grid map by treating the map as a boundary-value problem for steady-state heat conduction. The safety function h(p) is the solution of Laplace's equation, with obstacles held at temperature -a and safe boundaries at temperature b, and the paper argues that this h(p) is a valid control barrier function for single-integrator dynamics. Because the resulting linear system is sparse, the field can be recomputed online in about 9 ms for a 200 by 200 grid, fast enough to track a moving robot's changing sensor view. A single scalar constraint then keeps the robot inside the safe set regardless of how many obstacles or what shapes appear in the map. The payoff is a perception-driven safety filter that needs no pretraining and no per-obstacle modeling.","feed_headline":"A 9 ms heat-field solve keeps robots safe in unknown spaces","feed_subtitle":"One control-barrier constraint, recomputed from a local grid map, handles any number of obstacles without pretraining.","key_machinery":"The load-bearing object is the discrete Laplace equation with a five-point stencil: each transition cell's safety value equals the average of its four neighbors, assembled into the sparse linear system Ah = b, where A has diagonal 4 and off-diagonal entries -1 or 0. Boundary cells are fixed by the Dirichlet condition (10), and interior cells in the safety and obstacle regions are fixed at b and -a. The sparsity lets the system be solved iteratively by GMRES or BiCGSTAB, which is what makes online recomputation at millisecond scale possible. The harmonic nature of the solution supplies the smoothness, maximum principle, and isolated-critical-point properties on which the safety proof rests.","core_discovery":"The central claim is that the harmonic function obtained from Laplace's equation, with Dirichlet conditions h = -a on obstacle boundaries and h = b on the safety-region and workspace boundaries, is a control barrier function for the single-integrator model, and that its discrete version over a local occupancy grid map can be solved in real time. The paper defines this as the steady-state thermal field-inspired CBF (SSTF-CBF). Propositions 1 and 3 argue that h is continuously differentiable, that its critical points are isolated, and that even if a robot crosses the h = 0 boundary at a point where the gradient vanishes, it cannot reach an obstacle before leaving the unsafe sublevel set. Experiments with a TurtleBot in simulation and in the lab show the field synthesized in average times around 9 ms on 200 by 200 maps, with safety values remaining positive during navigation.","pith_inferences":["A natural next step is to derive a quantitative bound on the error between the local-window harmonic solution and the global harmonic solution; without such a bound the safety certificate formally covers the closed workspace, not a moving sensor window.","Because the map-to-CBF mapping is linear, the fixed-stencil computation could be shifted almost entirely to a GPU or embedded accelerator, potentially reducing the 9 ms figure further.","The same harmonic construction could serve as a navigation function, with gradient descent on h used for planning as well as for safety filtering, uniting the two layers in one scalar field.","For multi-robot operation, the experiments show mutual avoidance, but the proof treats other agents as static obstacles in each local map; a formal treatment of concurrent updates would be needed before claiming pairwise guarantees between moving robots."],"forward_implications":["Only one CBF constraint is needed in the safety QP, independent of obstacle count or shape, so the optimization stays small even in cluttered maps.","The per-cycle synthesis time of about 9 ms on a 200 by 200 grid makes it practical to refresh the safety field as fast as sensor maps are produced, including with dynamic obstacles.","Because the field is harmonic in the transition region, there is no need for the interpolation or smoothing that signed-distance-function-based CBFs require.","A robot that momentarily crosses into the h < 0 region is still protected from collisions by Proposition 3, so the filter tolerates temporary violations of the nominal safe set.","The same construction works with local maps that treat other robots as obstacles, as demonstrated by the two-robot real-world experiment."],"supporting_citations":[{"why":"Defines control barrier functions and the forward-invariance condition that the proposed h(p) must satisfy.","marker":"[6]"},{"why":"Supplies the steady-state heat-conduction model and the smoothness of its solutions, grounding the Laplace-equation construction.","marker":"[32]"},{"why":"Provides the maximum principle and harmonic-function properties used in Proposition 3's safety argument.","marker":"[33]"},{"why":"GMRES is one of the sparse iterative solvers used to compute the discrete Laplace solution quickly.","marker":"[34]"},{"why":"BiCGSTAB is the alternative sparse iterative solver used for the same online computation.","marker":"[35]"},{"why":"Gives the result that critical points of a non-constant harmonic function are isolated, used to handle vanishing gradient on the safe-set boundary.","marker":"[36]"},{"why":"Provides the near-identity diffeomorphism used to map single-integrator safe commands to the unicycle model in experiments.","marker":"[37]"}],"fun_headline_variants":["Harmonic heat fields synthesize CBFs in 9 ms for safe navigation","Millisecond safety barrier synthesis from local occupancy grids","Laplace's equation yields real-time control barrier functions online","9 ms safety field from grid maps keeps robots collision-free","Online CBF synthesis via heat conduction solves in 9 ms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The safety guarantee rests on the assumption that the local map's outer boundary really is a safe boundary, so anything outside the current sensor range is treated as safe; an obstacle lurking just beyond the map edge is invisible to the computed field.","fun_headline_variants_meta":{"raw":{"variants":["Harmonic heat fields synthesize CBFs in 9 ms for safe navigation","Millisecond safety barrier synthesis from local occupancy grids","Laplace's equation yields real-time control barrier functions online","9 ms safety field from grid maps keeps robots collision-free","Online CBF synthesis via heat conduction solves in 9 ms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1458,"prompt_tokens":921,"completion_tokens":537,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":453}},"tokens_in":537,"tokens_out":537,"duration_ms":5765,"temperature":1.0,"reasoning_tokens":453,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:44:19.433035+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the controller in an arena with a thin pole placed just outside the robot's 200 by 200 local map in the direction of travel, so it is not yet in the occupancy grid; if the robot does not brake or deviate before the map expands to reveal the pole, the online safety claim is disproved.","supporting_citations":[{"cited_title":"Control barrier functions: Theory and applications,","cited_arxiv_id":null,"evidence_quote":"Defines control barrier functions and the forward-invariance condition that the proposed h(p) must satisfy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the steady-state heat-conduction model and the smoothness of its solutions, grounding the Laplace-equation construction."},{"cited_title":"Basic properties of harmonic functions,","cited_arxiv_id":null,"evidence_quote":"Provides the maximum principle and harmonic-function properties used in Proposition 3's safety argument."},{"cited_title":"Bi-CGSTAB: A fast and smoothly converging variant of Bi-CG for the solution of nonsymmetric linear systems,","cited_arxiv_id":null,"evidence_quote":"BiCGSTAB is the alternative sparse iterative solver used for the same online computation."},{"cited_title":"p-harmonic functions in the plane,","cited_arxiv_id":null,"evidence_quote":"Gives the result that critical points of a non-constant harmonic function are isolated, used to handle vanishing gradient on the safe-set boundary."}],"review_version":1}