{"id":"1b3d1bd8-33a7-438a-b8aa-4be80e5343ce","arxiv_id":"2505.06794","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Solving Poisson's equation on an occupancy map with a tunable negative forcing function yields a smooth safety function that can be used as a control barrier function for collision-free robot control.","lead":"This paper uses Poisson's equation to turn sensor data about obstacles into smooth 'safety functions' that define safe regions for a robot. These functions can drive control-barrier safety filters, and the authors demonstrate real-time collision avoidance on a quadruped and a humanoid robot.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dynamic-safety claim rests on an unproven time-varying extension: Proposition 1 and Theorem 6 are static, while the dynamic experiments recompute h(t,y) at 10 Hz and approximate its time derivative via eq. (41); the humanoid's reported negative h is direct evidence of the gap.","rationale":"The reader's weakest-assumption analysis correctly identifies the central soft spot: the paper's formal invariance results are static, while the dynamic safety claim is carried by experiments with time-varying h. My read of the theory confirms that Theorem 1, Proposition 1, and Theorem 6 are all stated for fixed h, and nothing in the manuscript proves forward invariance when h is recomputed online and its time derivative is approximated by finite differences. The humanoid experiment's reported negative h is an explicit manifestation of this gap, and the limitations section does not resolve it. The static contribution—constructing smooth safety functions from Poisson's equation and using them in CBF filters—is otherwise well supported by classical elliptic theory and standard backstepping arguments, so I do not see grounds to reject the paper. The appropriate outcome remains a conditional acceptance requiring either a time-varying invariance theorem or a downgrade of the dynamic hardware claims to heuristic demonstrations.","tokens_in":22021,"tokens_out":8673,"duration_ms":99282,"concrete_test":"Reproduce the dynamic quadruped experiment in simulation with exact re-solved h(t,y) updated at 10 Hz and a moving obstacle, then evaluate the time-varying CBF inequality partial h / partial t + Dh * u >= -gamma h along sampled trajectories. If a trajectory leaves the set {h >= 0} while the frozen-time static condition holds at every sample, the dynamic claim requires a new theorem; if no exit occurs over a broad sweep of obstacle speeds and update rates, the practical gap is mitigated but still not formally closed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical guarantee—forward invariance of C via Theorem 1 plus Proposition 1 (first order) and Theorem 6 (higher order)—is proven only for a fixed, time-independent safety function h(y). In the dynamic hardware experiments (Sections VI.B.2 and VI.B.3), h is recomputed from perception data at roughly 10 Hz and the temporal change is approximated by finite differences in eq. (41), but no theorem states that this time-varying h(t,y) satisfies the CBF inequality. For a time-dependent h the correct forward-invariance condition is dot h = partial h / partial t + Dh * dot y >= -gamma h (plus the Lg h u term for control-affine systems). Thus either the finite-difference term is used without proof that the resulting filter is safe, or the static condition is applied at each frozen instant, in which case a moving obstacle can drive h negative even though every frozen h is a valid safety function. The reported humanoid result—briefly negative h—is direct evidence of this gap. Because the abstract and title claim 'dynamic safety' and the dynamic experiments carry that claim, this unproven time-varying extension is the load-bearing assumption. The paper should either provide a time-varying CBF theorem with explicit assumptions on the regularity and rate of change of h(t,y), or reclassify the dynamic demonstrations as heuristic rather than guaranteed safety.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for synthesizing safety functions from occupancy-map perception data by solving Poisson's equation with Dirichlet boundary conditions and a strictly negative forcing function. The resulting solution h is shown to be a safety function (Theorem 1), to define a forward-invariant safe set for first-order integrator systems (Proposition 1), and, via control barrier function backstepping, to yield safety filters for higher-order integrator chains (Theorem 6). The authors demonstrate the approach in simulations and on quadruped and humanoid hardware, including dynamic environments where h is recomputed online at about 10 Hz and its temporal derivative is approximated by finite differences in Eq. (41).","tokens_in":22343,"tokens_out":4158,"duration_ms":41294,"significance":"The static theoretical core is a clean, constructive bridge from geometric perception data to smooth safety functions with classical regularity, avoiding the gradient discontinuities of signed distance functions. The variational characterization and the explicit use of Hopf's lemma and elliptic regularity are credible and useful. If the dynamic-environment claims were backed by a time-varying invariance theorem, the paper would be a strong practical contribution to CBF-based safety filtering. As it stands, the static theory is solid, but the title- and abstract-level claim of 'dynamic safety' rests on an unproven extension to time-varying h, and one hardware experiment shows h becoming negative. The paper should be revised to either provide the missing time-varying guarantee or soften the dynamic-safety claims to heuristic demonstrations.","major_comments":[{"comment":"Proposition 1 and Theorem 6 establish forward invariance only for a time-independent safety function h(y), but the dynamic hardware experiments recompute h from perception at roughly 10 Hz and use the finite-difference term in Eq. (41) to approximate ∂h/∂t. No theorem or prior reference is provided showing that the resulting time-varying filter still enforces the CBF inequality ˙h(t,y) = ∂h/∂t + Dh·u ≥ −γh, and simply applying the static condition at each frozen instant does not preclude a moving obstacle from driving h negative between updates. The paper should either state and prove a time-varying version of the safety-filter guarantee with explicit regularity and rate-of-change assumptions on h(t,y), or explicitly reclassify the dynamic experiments as heuristic demonstrations rather than guaranteed safe.","section":"§VI.B.2, Eq. (41)"},{"comment":"The paper's own data show the humanoid's h(t) becoming briefly negative, and the text attributes this to reduced-order-model mismatch. This contradicts the statement that the safety filter renders C forward invariant, since the filter's guarantee applies to the modeled single-integrator state, not the actual robot state; if the claimed dynamic-safety result requires robustness to tracking error, the manuscript needs a formal treatment of that error (e.g., an input-to-state safety argument with a quantified tracking bound) or should present the humanoid result as a limitation rather than as validation of the invariance guarantee.","section":"§VI.B.3, Fig. 6"}],"minor_comments":[{"comment":"There is a typo: 'aribtrary' should be 'arbitrary'; similarly, the Appendix A heading 'POISSION'S EQUATION' should be 'POISSON'S EQUATION'.","section":"§II.B"},{"comment":"The sentence 'Since hB(y) ≤ h(y) for all y ∈ C' should read 'for all (y, ˙y) ∈ C × R^3', because hB depends on the velocity error as well as the position.","section":"§V.B"},{"comment":"The claim that 'h remained positive' is based on evaluation at sampled robot states; the paper should explicitly state the sampling rate and avoid implying a continuous-time certificate from discrete evaluations.","section":"§VI.B.1, Fig. 5"},{"comment":"References [30] and [43] are the same conference paper; please disambiguate or unify the citation entries.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The static safety-function synthesis is sound and likely publishable after modest revision; the main risk is the unproven time-varying extension underlying the 'dynamic safety' claim. I would not reject outright, but the authors must either supply a time-varying invariance theorem or significantly temper the claims in the abstract and hardware sections. The humanoid negative-h result should also be handled honestly, either as a robustness gap or as a demonstration limitation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the static part of this paper is a legitimate contribution, and the dynamic part is where the claim outruns the proof. The construction—solving Poisson with negative softplus forcing derived from a harmonic guidance field—is clean, and the regularity chain from Hölder f to C^{2+k} h to CBF backstepping is standard and correctly used. The static forward invariance (Prop 1, Thm 6) is fine as far as I can tell. The double-integrator simulation and static quadruped test are nice sanity checks, and the non-uniform boundary flux idea is a genuinely useful knob for steering gradients away from deadlocks.\n\nThe soft spot is exactly where the stress test points. Everything proven is for a time-independent h(y). In the dynamic experiments h is re-solved at 10 Hz and the time derivative is approximated by finite differences (eq. 41). There is no theorem stating that this time-varying h satisfies the CBF inequality, and no assumptions on how fast h may change. The humanoid trace going negative is not a small blemish—it is the predicted failure of the unproven extension. The authors do admit the brief violations and attribute them to ROM tracking lag, which is plausible, but the gap remains: they present dynamic safety as a guarantee in the abstract, and the experiments are the evidence. The fix is either a time-varying CBF theorem with explicit regularity and rate conditions, or reframing the dynamic demos as heuristic. A third option—reporting the actual safety violation metrics—would at least make the empirics honest.\n\nMinor points: no code or data release, so the 0.2–0.3 ms solve time and the full pipeline are not independently reproducible. The citation pattern is fair; elliptic potentials and PDE distance approximations are credited. The novelty is not in the PDE theory itself but in the specific forcing construction and its integration with CBF filtering. That is enough.\n\nVerdict: engage with it. The static contribution is solid and the dynamic gap is identifiable and fixable; a serious referee can ask for a time-varying theorem or an honest reframing.","headline":"Static Poisson safety-function synthesis is a real contribution, but the paper's dynamic-safety claim rests on an unproven time-varying extension.","tokens_in":22843,"tokens_out":2113,"would_cite":true,"duration_ms":22627,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35J05","93C85"],"pacs":[],"model":"deepseek-v4-flash","headline":"Solving Poisson's equation on an occupancy map produces a safety function that can certify safe robot motion and drive a CBF-based safety filter.","keywords":["control barrier functions","Poisson's equation","safety filters","safe set synthesis","occupancy maps","CBF backstepping","quadruped robot","humanoid robot"],"falsifier":"Run the hardware pipeline with an obstacle moving through the field faster than the roughly 10 Hz update rate; the claim fails if $h$ becomes negative before the next recomputation, or if one can construct two consecutive updates satisfying (41) and the filter constraint yet with $h$ crossing zero in between. A cheaper falsifier is a simulation of the double integrator with a periodically switching occupancy map that checks the continuous trajectory between safety-function updates.","tokens_in":21840,"feed_emoji":"🤖","tokens_out":10708,"duration_ms":101054,"temperature":0.7,"pith_summary":"This paper attempts to establish a constructive pipeline from raw perception to verified safety: take a local occupancy map, solve Poisson's equation with zero boundary values on obstacle surfaces and a strictly negative forcing function, and use the unique solution $h$ as the safety constraint in a Control Barrier Function (CBF) safety filter. If successful, the method replaces hand-designed or gradient-discontinuous constraints, such as signed distance functions, with one smooth function that is automatically synthesized from arbitrary obstacle geometry and comes with forward-invariance proofs. The paper proves that $h$ is a safety function, that it renders the safe set forward invariant for first-order models, and that CBF backstepping extends the guarantee to higher-order integrator chains. Hardware demonstrations on a quadruped and a humanoid in static and moving obstacle fields show the computation is fast enough for real-time use, which is why a general reader should care: it is a concrete step toward robots that can be handed a camera feed and a guarantee about not colliding.","feed_headline":"Poisson's equation turns occupancy data into certified safe sets","feed_subtitle":"A single smooth field built from obstacle boundaries can filter unsafe commands on quadruped and humanoid robots.","key_machinery":"The central object is the Poisson safety function: the solution $h$ of the Dirichlet problem $\\Delta h = f$ on the occupancy domain with $h=0$ on obstacle boundaries and $f$ strictly negative. The forcing function is the design knob. The paper's main construction routes through a smooth guidance vector field $\\vec{v}$: each component solves Laplace's equation with boundary data $b(y)\\,\\hat{n}(y)$ encoding the desired flux, and the forcing term is a smooth softplus-style function of $\\nabla \\cdot \\vec{v}$ that stays strictly negative, so the solution is automatically smooth and superharmonic. This also makes $h$ the unique minimizer of the variational problem $\\min_h \\int_\\Omega \\left(\\tfrac{1}{2}|Dh|^2 - h\\, \\beta^{-1}\\ln(1+e^{-\\beta \\nabla \\cdot \\vec{v}})\\right) dV$. The companion machinery is CBF backstepping: auxiliary safe controllers $k_i$ are designed recursively, and the function $h_B(\\vec{y}) = h(y) - \\sum_i \\tfrac{1}{2\\mu_i}\\| y^{(i)} - k_i\\|^2$ is shown to define a shrunken invariant safe set, with Lemma 1 supplying the Lipschitz regularity of $D^r h$ that makes the controller locally Lipschitz and well-defined.","core_discovery":"The paper's central claim is that a single scalar function encoding an entire obstacle field can be manufactured by solving an elliptic boundary value problem. Let $\\Omega$ be the unoccupied region bounded by obstacle surfaces $\\partial\\Omega$, and let $f \\in C^{k,\\alpha}(\\Omega;\\mathbb{R}_{<0})$ be H\\\"older continuous and strictly negative. The unique classical solution $h$ of $\\Delta h = f$ in $\\Omega$, $h=0$ on $\\partial\\Omega$, is a safety function of order $2+k$: its zero superlevel set is a compact safe set whose boundary is $\\partial\\Omega$, and Hopf's lemma guarantees $Dh \\cdot \\hat{n} < 0$ on the boundary, so the gradient is nonzero exactly where it is needed. Because $\\Delta h < 0$, the weak minimum principle keeps $h$ nonnegative inside $\\Omega$, so the safe set is the whole free space, and each obstacle interior can be made unsafe by solving with a positive forcing function there. The paper then proves this safety function is a CBF: Proposition 1 gives a Lipschitz controller for $\\dot{y} = w$ satisfying $\\dot{h} \\geq -\\gamma h$, and Theorem 6 extends the argument to integrator chains of relative degree $r \\geq 2$ using backstepping functions $h_B$ that render a shrunken subset $C_B \\subset C \\times \\mathbb{R}^{3(r-1)}$ forward invariant.","pith_inferences":["A rigorous dynamic-safety theorem would need to treat $h(t,y)$ as time-varying; one route is to bound $\\|\\partial h/\\partial t\\|$ and add it to the CBF condition, turning the current finite-difference approximation (41) into a provable margin.","The per-obstacle boundary flux $b$ in (26) is a tuning handle for shaping gradient magnitudes; this suggests a data-driven or optimization-based flux assignment to push the system away from deadlock equilibria, though no such guarantee is given.","The same variational formulation could be applied to higher-dimensional state spaces or to directly synthesizing a CLF-CBF pair, since the guidance field construction is independent of the robot's dynamics."],"forward_implications":["Occupancy maps can be converted directly into safety constraints: the same pipeline yields a smooth CBF for arbitrary obstacle geometry, avoiding the gradient discontinuities of signed distance functions that cause chattering in safety filters.","Safety is decoupled from the goal: the synthesized safety function is geometry-only, so a single $h$ serves any nominal controller in a safety-filter quadratic program.","The guarantees cover higher-order dynamics, not just kinematic models: for integrator chains of relative degree $r \\geq 2$, the backstepped CBF makes a shrunken safe set forward invariant.","The regularity of the safety function is tunable through the forcing function: H\\\"older-continuous forcing already suffices for low relative degree, while smooth forcing gives smooth $h$ for higher relative degree.","On hardware, the pipeline runs at about 10 Hz with PDE solve times of 0.2--0.3 ms, which the experiments indicate is fast enough to keep a quadruped safe while obstacles move."],"supporting_citations":[{"why":"Defines CBFs and the quadratic-program safety filter that the Poisson safety function is plugged into.","marker":"[8]"},{"why":"Supplies the classical elliptic regularity result that H\\\"older-continuous forcing yields a $C^{2+k,\\alpha}$ solution, the basis of Theorem 1.","marker":"[14]"},{"why":"Provides the weak maximum principle, Hopf's lemma, Laplace equation regularity, and Dirichlet's principle used throughout.","marker":"[38]"},{"why":"States the maximum/minimum principles and Hopf's lemma that give the boundary gradient sign and safety-function property.","marker":"[51]"},{"why":"Gives the safe backstepping theorem used to render the shrunken set $C_B$ forward invariant for high-order chains.","marker":"[43]"},{"why":"Frames constructive backstepping for partially feedback linearizable systems, the template for Theorem 6.","marker":"[24]"},{"why":"Introduces the guidance vector field approach from Poisson image editing that the paper adapts to construct smooth forcing functions.","marker":"[26]"},{"why":"Supplies the SOR finite-difference solver used for the real-time PDE solves.","marker":"[45]"}],"fun_headline_variants":["Poisson's equation encodes obstacles into safe control sets","One PDE turns perception data into safety filters for robots","Certified safe sets from occupancy maps via a single Poisson solve","Safe robot control from a single solution to Poisson's equation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that forward invariance for the static safety function $h$ transfers to the time-varying $h(t,y)$ recomputed online and differentiated by finite differences in (41), an extension the paper assumes but does not prove or cite.","fun_headline_variants_meta":{"raw":{"variants":["Poisson's equation encodes obstacles into safe control sets","One PDE turns perception data into safety filters for robots","Certified safe sets from occupancy maps via a single Poisson solve","Safe robot control from a single solution to Poisson's equation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000953,"raw_usage":{"total_tokens":4094,"prompt_tokens":1003,"completion_tokens":3091,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":3025}},"tokens_in":619,"tokens_out":3091,"duration_ms":20099,"temperature":1.0,"reasoning_tokens":3025,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:33:17.883799+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the hardware pipeline with an obstacle moving through the field faster than the roughly 10 Hz update rate; the claim fails if $h$ becomes negative before the next recomputation, or if one can construct two consecutive updates satisfying (41) and the filter constraint yet with $h$ crossing zero in between. A cheaper falsifier is a simulation of the double integrator with a periodically switching occupancy map that checks the continuous trajectory between safety-function updates.","supporting_citations":[{"cited_title":"Control barrier function based quadratic programs for safety critical systems,","cited_arxiv_id":null,"evidence_quote":"Defines CBFs and the quadratic-program safety filter that the Poisson safety function is plugged into."},{"cited_title":"Gilbarg and N","cited_arxiv_id":null,"evidence_quote":"Supplies the classical elliptic regularity result that H\\\"older-continuous forcing yields a $C^{2+k,\\alpha}$ solution, the basis of Theorem 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the weak maximum principle, Hopf's lemma, Laplace equation regularity, and Dirichlet's principle used throughout."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"States the maximum/minimum principles and Hopf's lemma that give the boundary gradient sign and safety-function property."},{"cited_title":"Safe backstepping with control barrier functions,","cited_arxiv_id":null,"evidence_quote":"Gives the safe backstepping theorem used to render the shrunken set $C_B$ forward invariant for high-order chains."},{"cited_title":"Construc- tive safety-critical control: Synthesizing control barrier functions for partially feedback linearizable systems,","cited_arxiv_id":null,"evidence_quote":"Frames constructive backstepping for partially feedback linearizable systems, the template for Theorem 6."},{"cited_title":"Poisson image editing,","cited_arxiv_id":null,"evidence_quote":"Introduces the guidance vector field approach from Poisson image editing that the paper adapts to construct smooth forcing functions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the SOR finite-difference solver used for the real-time PDE solves."}],"review_version":1}