{"id":"ab39a08f-e15c-447c-bd15-ee583d96be7a","arxiv_id":"2506.18460","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A distributed bearing-only estimator plus a pointing controller make all agents' headings converge to a stationary target using only two non-collinear sensing agents.","lead":"This paper builds a distributed estimator and controller that lets a group of fixed agents aim their headings at a stationary target when only two agents can measure the target's bearing. The result drops the collinearity and persistent-excitation assumptions of earlier pointing-control work and gives convergence proofs plus a UE4 simulation video.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2 rests on a citation that is not shown to cover a time-varying estimate, and as stated it admits a trivial anti-aligned fixed point; the pointing stage needs a self-contained, qualified proof.","rationale":"The reader's weakest assumption, that Theorem 2 is asserted by citation without checking whether Trinh et al.'s theorem covers a time-varying estimate, is the same core concern I would identify. The localization half of the paper is internally consistent: the LTI error dynamics, the Hurwitz argument for the SA layer, and the ISS argument for the NSA layer support Theorem 1, modulo minor imprecision such as calling a nonsymmetric matrix positive-definite. The pointing half is different. Theorem 2 is not proved in the paper, and a theorem about pointing to a fixed target does not automatically carry over to a reference that moves because the observer is still converging. Whether the cascade works is a concrete technical question, not a matter of taste; it depends on the growth of the perturbation d_i(t) and on excluding the unstable anti-aligned equilibrium. The counterexample with all estimates initialized at q0 shows that the universal statement is false as written, so the authors need at least an almost-everywhere qualifier. For these reasons the reader's CONDITIONAL verdict appears correct: the central idea is likely sound, but the paper should be revised to include a self-contained proof of the pointing stage, a qualifier on initial headings, and a cleanup of the textual inconsistency around Eq. (7b).","tokens_in":6100,"tokens_out":15683,"duration_ms":167804,"concrete_test":"Obtain Theorem 3.4 of Trinh et al. (2020), write out its hypotheses, and check each against the closed loop (7a)-(9), specifically whether it permits the reference position in (9) to be the time-varying estimate \\hat q_i(t) instead of a fixed q0. Independently, derive the heading error dynamics \\dot\\psi_i = -\\|q_0-p_i\\|\\sin\\psi_i + d_i(t) with d_i(t) bounded by \\|\\hat q_i(t)-q_0\\|, and verify a standard cascade lemma (e.g., Panteley-Loria or Khalil Lemma 10.3) using the exponential bound \\|\\hat q_i(t)-q_0\\| \\le Ce^{-\\lambda t}. If the cascade condition holds, Theorem 2 is true once restated with the anti-aligned exception removed; if it fails, the pointing stage is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two halves. The localization half is well supported: Lemma 1 gives exponential convergence of the two sensing-agent estimates under Assumption 1, and Theorem 1 extends this to non-sensing agents under Assumption 2 via an ISS argument with exponentially decaying inputs. The pointing half, however, is not established. Theorem 2 is proved only by the sentence 'See Theorem 3.4 in Trinh et al. (2020)'. That theorem is not restated, and the paper does not verify its hypotheses for the closed loop (7a)-(9). In particular, the controller (9) is driven by \\hat q_i(t), a time-varying estimate that converges only asymptotically, whereas a pointing theorem for a fixed target position may not extend to a moving reference without an additional cascade or robustness argument. This gap is load-bearing because it is the only support for the pointing half of the result. Moreover, the statement as written is literally false without a generic-initial-condition qualifier: if all estimates start at q0 and some h_i(0) is exactly opposite to q0-p_i, then \\dot h_i = M_{h_i}(q0-p_i)=0 for that agent, and its heading never points toward the target. Thus Theorem 2 needs both a self-contained proof and an 'almost all initial headings' or non-anti-aligned qualification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a network of n stationary agents in the plane, each with a unit heading vector. Two sensing agents measure the bearing to a stationary target; the remaining agents communicate over a directed graph. The authors propose a distributed bearing-only estimator (7) for the target position and a pointing controller (9) that steers each heading toward the agent's current estimate. Lemma 1 and Theorem 1 claim exponential convergence of all estimation errors under Assumptions 1–2; Theorem 2 claims that all headings asymptotically point at the target. The localization half is supported by a linear analysis and a virtual-fusion-node graph argument; the pointing half is delegated to a citation.","tokens_in":6252,"tokens_out":10422,"duration_ms":114125,"significance":"The localization result is a clean analytic contribution: with two non-collinear bearing sensors and a rooted communication graph, the linear estimator (7) avoids persistent-excitation conditions and requires no data fitting. The virtual fusion node is a reasonable way to convert the two-leader problem into a standard consensus-tracking form, and the paper provides a UE4 demonstration. However, the central two-step claim is incomplete: Theorem 2 is not proved in the manuscript and is false as stated for anti-aligned initial headings. The contribution is potentially useful but currently conditional on a substantiated pointing-convergence proof.","major_comments":[{"comment":"The proof of Theorem 2 consists of the single sentence 'See Theorem 3.4 in Trinh et al. (2020).' The hypotheses of that theorem are not restated and are not verified for the closed loop (7a)–(9). In particular, the controller (9) is driven by \\hat q_i(t), a time-varying signal that converges to q_0 only asymptotically; a pointing-convergence theorem for a fixed target position does not automatically extend to this cascade without an explicit robustness or two-time-scale argument. Because Theorem 2 is the sole support for the pointing half of the central claim, this gap is load-bearing. Please give a self-contained proof, or state the external theorem and verify all its conditions, including the time-varying reference.","section":"Section 3.2, Theorem 2"},{"comment":"As stated, Theorem 2 is not correct without a generic-initial-condition qualifier. If all estimates are initialized at q_0 and one agent has h_i(0)=-(q_0-p_i)/||q_0-p_i||, then \\dot h_i = M_{h_i}(q_0-p_i)=0 for that agent; because the estimator has zero initial error in this configuration, \\hat q_i remains at q_0 and the anti-aligned heading never converges to the target direction. The theorem should be qualified as holding for almost all initial headings, or the anti-aligned equilibrium should be explicitly excluded and its instability analyzed.","section":"Section 3.2, Theorem 2 statement"},{"comment":"The proof of Theorem 1 states that (L+\\bar B_f)\\otimes I_2 is positive definite by Theorem 3.6 of Ren and Beard (2005). For the directed NSA graph, L+\\bar B_f is generally not symmetric, and the cited theorem is a consensus result rather than the matrix fact used here. The needed statement is that L+\\bar B_f is a nonsingular M-matrix, hence positive stable, under Assumption 2, which then makes the linear system (20) exponentially stable. Please correct the matrix-theoretic claim and the citation.","section":"Section 3.2, Theorem 1 proof"}],"minor_comments":[{"comment":"The proof says the Hurwitz condition is 'easy to verify' but does not show the verification. Please include the Routh array or the explicit inequalities that, together with sin^2(\\theta_1-\\theta_2) \\neq 0, imply all roots have negative real parts.","section":"Lemma 1, Eq. (15)"},{"comment":"The symbol B_f is used both as a vector and, through \\bar B_f = diag(B_f), as a diagonal matrix; please make the notation consistent and state the dimensions of the Kronecker products.","section":"Eq. (20)"},{"comment":"The step from ISS plus an exponentially decaying input to exponential convergence of \\tilde q^* is asserted rather than shown; for the linear system at hand the implication is true, but it should be stated explicitly.","section":"Theorem 1 proof"},{"comment":"The text contains 'Fig. Fig. 6'; the duplicate word should be removed.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is compact and the localization half is sound in substance. The blocking issue is Theorem 2, which is currently a citation and is not qualified against the anti-aligned equilibrium. I see no circularity or data-fitting concern; the gap is fixable but requires a real proof rather than a local edit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: read this for the two-agent bearing-only estimator and the virtual fusion node argument; do not rely on Theorem 2 as written.\n\nWhat is actually new: the paper gives a distributed bearing-only localization scheme that needs only two non-collinear sensing agents and removes the persistent-excitation and collinearity assumptions used in earlier work. Lemma 1 is sound; I checked the error dynamics and the Hurwitz condition. The characteristic polynomial is correct and the non-collinearity condition is the necessary and sufficient condition for exponential convergence. The virtual fusion node is a genuinely nice device: it turns the non-sensing-agent estimation problem into a consensus-tracking problem with an exponentially decaying input. Theorem 1 is plausible under Assumption 2, given an ISS argument that is sketched but not fully detailed. The contribution relative to Deghat et al., Chen et al., and Wu et al. is real.\n\nThe soft spots are concentrated in the pointing stage. Theorem 2 is proved by the sentence \"See Theorem 3.4 in Trinh et al. (2020)\", with no restatement of that theorem's hypotheses and no demonstration that it covers a time-varying, asymptotically converging reference. The controller (9) is driven by the estimate, not by the true q0, and a cascade argument is needed. Without it, the pointing convergence is asserted, not derived. Worse, the statement as written is false: if an agent's initial heading is exactly opposite to q0 - p_i, then M_{h_i}(q0 - p_i) = 0, so that heading never changes and the agent never points at the target. Theorem 2 needs a generic-initial-condition qualifier and a self-contained proof. This is not fatal to the localization idea, but it means the headline claim of \"pointing consensus\" is not established as stated.\n\nThere are a few smaller issues. The abstract omits the graph-connectivity assumption, so \"only two agents not collinear\" sounds stronger than it is. The description of Eq. (7b) mentions an orthogonal projection term that is not present in that equation. The simulation is a single six-agent scenario with no gains and no baseline; it illustrates the behavior but does not validate the theorem. The citation pattern is otherwise fine; the problem is the weight placed on one external theorem.\n\nBottom line: I largely agree with the reader. The paper deserves a serious referee because the localization idea is clean and likely correct, but the authors should either supply a real proof of Theorem 2 with the anti-aligned qualifier or soften the claim to a conjecture supported by simulation. I would not desk-reject this; I would send it to review and ask for a major revision.","headline":"The localization half of this paper is solid and new, but the pointing half is a citation with an unqualified statement that is literally false in an anti-aligned initial condition.","tokens_in":6882,"tokens_out":2305,"would_cite":true,"duration_ms":24905,"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 two non-collinear bearing sensors let a distributed network localize a stationary target and asymptotically point all agent headings at it.","keywords":["target-pointing consensus","bearing-only localization","distributed estimation","cooperative localization","virtual fusion node","multi-agent systems","projection matrix","networked control"],"falsifier":"Simulate the closed loop (7a)-(9) with two non-collinear sensing agents and a non-sensing layer satisfying Assumption 2, using very slow estimator gains; if the headings fail to approach the true target direction even after the estimation error has essentially vanished, the pointing claim is false as stated. Conversely, to refute the localization claim, run (7a)-(7b) with arbitrarily small initial errors and check whether $\\|\\tilde q_i(t)\\|$ decays exponentially; any persistent residual at arbitrarily small initial error would contradict Theorem 1.","tokens_in":5799,"feed_emoji":"🎯","tokens_out":10289,"duration_ms":97892,"temperature":0.7,"pith_summary":"This paper takes on the problem of making every agent in a distributed network point its heading at a common stationary target when only two agents can measure the target's direction (its bearing). It proposes a two-step strategy: a distributed bearing-only estimator through which every agent reconstructs the target position, followed by a pointing controller that turns each heading toward the estimate. The central claim is that two sensing agents whose lines of sight to the target are not collinear are enough for localizability, replacing the stronger persistent-excitation and collinearity assumptions of earlier pointing and localization schemes. If the claim holds, networks of platforms such as optical transceivers or observation instruments could acquire and track a target cooperatively without a prescribed formation.","feed_headline":"Two bearing-only sensors can aim a whole network at a target","feed_subtitle":"A distributed estimator plus pointing control removes earlier collinearity and persistent-excitation assumptions.","key_machinery":"The central object is the orthogonal projection matrix $M_{z_i}=I_2-z_i z_i^T$ formed from a sensing agent's unit bearing vector $z_i$; it cancels the line-of-sight component of the estimation error and drives the estimate along the perpendicular direction. Coupled through $k_{12}$ and $k_{21}$, the two sensing agents' errors become an LTI system whose stability is decided by the Hurwitz criterion, and the constant term of its characteristic polynomial ties stability exactly to the non-collinearity condition $\\sin(\\theta_1-\\theta_2)\\neq 0$. The second key device is the virtual fusion node: in the non-sensing layer the two sensing estimates are merged into a single leader node, converting the estimation law (7b) into a consensus-tracking system that is input-to-state stable when the fused graph has a spanning tree. The pointing control $M_{h_i}(\\hat q_i-p_i)$ applies the same projection idea to rotate each heading toward the estimated target.","core_discovery":"On the paper's own terms, the discovery is a bearing-only distributed localizability result: with two stationary sensing agents whose lines of sight to a stationary target are not collinear, the target position can be recovered exponentially by every agent in the network. The coupled error dynamics of the two sensing agents form a linear time-invariant system whose characteristic polynomial has constant term $k_{12}k_{21}\\sin^2(\\theta_1-\\theta_2)$, so exponential convergence is equivalent to $\\sin(\\theta_1-\\theta_2)\\neq 0$. Non-sensing agents then track the sensing agents' estimates through a consensus law, and the virtual fusion node merges the two sensors into one leader so that the non-sensing error system is input-to-state stable under Assumption 2. The paper further claims that the pointing control law (9) makes every agent's heading asymptotically point at the true target.","pith_inferences":["The virtual fusion node construction is not obviously limited to two sensing agents; fusing any set of sensing agents into one effective leader would likely give the same leader-following analysis, provided no two lines of sight are collinear.","For a slowly moving or drifting target, the LTI error argument no longer applies, but a time-scale separation between fast estimation and slow target motion would be a natural route to ultimately bounded or practical convergence.","The pointing claim (Theorem 2) is inherited from an external theorem cited in one sentence; a self-contained direct proof for the full coupled estimator-controller system would settle whether the moving-estimate cascade preserves the theorem's hypotheses, or reveal a condition under which only practical pointing holds."],"forward_implications":["With only two non-collinear sensing agents, no persistent-excitation condition or special formation is required for the network to localize a stationary target.","Agents without any bearing measurement can still obtain exponentially converging target estimates, as long as the communication graph lets them be reached from the sensing layer.","The two-step design decouples localization from pointing: once the estimate converges, the heading control (9) can steer all agents toward the same target.","The only localizability obstruction in the two-sensor case is exact collinearity, $\\theta_2=\\theta_1$ or $\\theta_2=\\theta_1+\\pi$; any other pair of lines of sight works."],"supporting_citations":[{"why":"Provided a geometric bearing-only estimator whose localizability required persistent excitation, the baseline the paper's estimator removes.","marker":"Deghat et al. (2014)"},{"why":"Provided a cooperative bearing-only localization scheme also relying on persistent excitation, used as a comparison for the relaxed assumptions.","marker":"Chen et al. (2023)"},{"why":"Gave a distributed pointing controller under a collinearity assumption, which the paper's strategy avoids.","marker":"Wu et al. (2023)"},{"why":"Removed the collinearity assumption but required substantial prior target information, contrasting with the minimal sensing here.","marker":"Trinh et al. (2018)"},{"why":"Its Theorem 3.4 is cited as the proof that headings converge to the target once estimates are available.","marker":"Trinh et al. (2020)"},{"why":"Its Theorem 3.6 supplies the positive-definiteness of the fused Laplacian that the paper uses to prove exponential convergence of the non-sensing estimates.","marker":"Ren and Beard (2005)"}],"fun_headline_variants":["Two bearing sensors exponentially localize and aim a network","Bearing-only: two non-collinear sensors give exponential lock","Distributed pointing: two sensors suffice, no persistent excitation","Networked aiming from two bearing measurements alone","Two non-collinear bearings: exponential convergence for all agents"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the external theorem cited for heading convergence still applies when each agent's target estimate is time-varying and only asymptotically converging; the paper does not verify the theorem's hypotheses, so if the theorem assumes a fixed target point the pointing claim is not established.","fun_headline_variants_meta":{"raw":{"variants":["Two bearing sensors exponentially localize and aim a network","Bearing-only: two non-collinear sensors give exponential lock","Distributed pointing: two sensors suffice, no persistent excitation","Networked aiming from two bearing measurements alone","Two non-collinear bearings: exponential convergence for all agents"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000691,"raw_usage":{"total_tokens":3076,"prompt_tokens":837,"completion_tokens":2239,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":2161}},"tokens_in":453,"tokens_out":2239,"duration_ms":14759,"temperature":1.0,"reasoning_tokens":2161,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:49:39.048379+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the closed loop (7a)-(9) with two non-collinear sensing agents and a non-sensing layer satisfying Assumption 2, using very slow estimator gains; if the headings fail to approach the true target direction even after the estimation error has essentially vanished, the pointing claim is false as stated. Conversely, to refute the localization claim, run (7a)-(7b) with arbitrarily small initial errors and check whether $\\|\\tilde q_i(t)\\|$ decays exponentially; any persistent residual at arbitrarily small initial error would contradict Theorem 1.","supporting_citations":[],"review_version":2}