{"id":"fb1a84be-248d-4b34-8212-011345672f4a","arxiv_id":"2504.18860","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A signed distance field based diffeomorphic transform lets contractive robot skills avoid obstacles while preserving contraction stability.","lead":"This paper shows how robot motions learned from demonstrations can avoid obstacles while keeping their mathematical stability guarantee. It uses signed distance fields to reshape the motion around objects and tests the method on simulated and real kitchen tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (27) defines SDC as the pullback of fc, which is trajectory-equivalent to the original dynamics and cannot provide obstacle avoidance; the obstacle-avoiding pushforward is never analyzed.","rationale":"The most load-bearing question is whether the system whose contraction is proven is the system that actually avoids obstacles. Eq. (27) is the pullback of fc. In differential geometry, the pullback of a vector field by a diffeomorphism is just the same vector field expressed in new coordinates; integral curves map to integral curves of the original field. Therefore y-trajectories of fSDC map via ψ to ordinary fc-trajectories in C, so no obstacle avoidance follows from the coordinate change alone. The standard diffeomorphic-obstacle-avoidance construction (e.g., Zhi et al. DT) uses the pushforward: the original stable dynamics is defined in the latent obstacle-free space and pushed forward to the workspace, ẋ = Jψ(ψ^{-1}(x)) f(ψ^{-1}(x)). The paper never defines such a system, and Algorithm 2's variable naming (send y_{t+1}) is ambiguous about which space is commanded. If the implemented controller commands y directly, the executed dynamics is the pullback, and no avoidance is expected; if it commands ψ(y), then the contraction proof in Appendix B4, which analyzes fSDC = J^{-1} fc(ψ(y)), is for the wrong system. This is more fundamental than the reader's concern: even granting a global nonsingular flow ψ, the pullback cannot produce the claimed avoidance. The SDDC is even less specified: fSDDC is never defined, and a differential coordinate change alone does not define a new vector field. For these reasons the central claim is not established by the paper's mathematics. A focused re-derivation and a code/simulation check of the commanded variable would settle the issue.","tokens_in":28600,"tokens_out":18514,"duration_ms":195149,"concrete_test":"Reconstruct the exact dynamics executed in Algorithm 2: determine whether the signal sent to the joint impedance controller is y_{t+1} (as written) or ψ(y_{t+1}). If it is y_{t+1}, simulate the LASA Sine example with a box obstacle and compare the executed trajectory with an unmodified NCDS rollout: the pullback dynamics predicts they coincide, so no avoidance should appear. If it is ψ(y_{t+1}), then the executed dynamics is the pushforward ẋ=Jψ(ψ^{-1}(x)) fc(ψ^{-1}(x)); re-run the Appendix B4 proof for this system and check whether the contraction condition ˙G+A^T G+GA+2αG≼0 holds with G=(Jψ^{-1})^T Jψ^{-1}. Also inspect the pseudocode variable conventions (y vs q) in the released code.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that the SDC/SDDC-transformed system is both obstacle-avoiding and contractive. For SDC, Eq. (27) defines fSDC(y)=Jψ(y)^{-1} fc(ψ(y)), which is exactly the pullback of fc by ψ (Eq. 8). A pullback is a coordinate representation of the same dynamics: if ẏ=fSDC(y), then q=ψ(y) satisfies q̇=fc(q). Hence the trajectory in configuration space is the unmodified learned trajectory; it cannot avoid an obstacle that fc would traverse. Obstacle avoidance requires the pushforward f_avoid(q)=Jψ(ψ^{-1}(q)) fc(ψ^{-1}(q)) (equivalently ẏ=fc(y) with command q=ψ(y)), but this system is never defined or analyzed. For SDDC, fSDDC is never defined; Eq. (26) is only a virtual-displacement relation, and using it as ẏ=Jψ^{-1} fc(q) again yields the pullback. Appendix B4 proves contraction of the pullback, not of any obstacle-avoiding system. Even granting global smoothness and nonsingularity of Jψ, the central claim does not follow from the presented equations; the paper proves stability of a system that does not perform the claimed avoidance.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework, Signed Distance Field Diffeomorphic Transform (SDT), for adding obstacle avoidance to contractive dynamical systems while preserving contraction. It constructs an infinitesimal generator from the gradient of an SDF weighted by inverse or swept barrier functions, integrates this generator into a flow psi (Eq. 24), and uses psi to modulate a neural contractive dynamical system either through a differential coordinate change (SDDC, Eq. 26) or a coordinate change (SDC, Eq. 27). The authors claim that Theorem 1 entails contraction preservation and therefore safe obstacle avoidance. They introduce two metrics (RFC, VM), compare against MM, DT, ARPF on LASA synthetic tasks, and demonstrate two real-world kitchen tasks with learned RDF/CDF representations.","tokens_in":28852,"tokens_out":11134,"duration_ms":111153,"significance":"If the central claim were correct, the paper would fill a real gap: combining implicit robot/scene representations with contraction-preserving whole-body obstacle avoidance. The experimental apparatus is substantial: learned RDF/CDF representations, swept barrier features, comparisons over SDF architectures, real-robot demonstrations, and inference-time decompositions. The paper also makes a good-faith attempt to quantify trajectory curvature and vector-field alignment. However, the core mathematical object analyzed for stability is not the object that performs avoidance; the defined SDC is a pullback of the original system, and SDDC is not defined as a dynamical system. Because the main contribution is precisely the claim that these transforms yield both avoidance and contraction, the paper's central conclusion is unsupported.","major_comments":[{"comment":"Equation (27) defines fSDC(y)=Jpsi(y)^{-1} fc(psi(y)), which is the pullback of fc by psi, not an obstacle-avoiding pushforward. With q=psi(y), one immediately obtains q_dot=Jpsi(y) y_dot=fc(psi(y))=fc(q), so the configuration-space trajectory is exactly the unmodulated NCDS trajectory. Since psi is a diffeomorphism, a trajectory avoids the obstacle set O if and only if its preimage avoids psi^{-1}(O); a coordinate change cannot turn a colliding trajectory into a non-colliding one. Therefore the claim in Section IV-B3 that 'the robot can successfully avoid obstacles while preserving the stability of the underlying skill' does not follow from Eq. (27), and the avoidance visible in Figures 6 and 9 must be produced by some other, undisclosed controller. The authors need to define and analyze the pushforward system q_dot=Jpsi(psi^{-1}(q)) fc(psi^{-1}(q)) (or an equivalent) and prove obstacle avoidance for that system.","section":"IV-B4, Eq. (27)"},{"comment":"Equation (26) is a virtual-displacement relation delta y=Jpsi^{-1} delta q; no dynamical system fSDDC is ever defined. The text concludes 'the transformed system fSDDC is contractive', but there is no ODE for y or q to which Theorem 1 could be applied. If the intended definition is y_dot=Jpsi^{-1} fc(psi(y)), it is again a pullback and inherits the avoidance problem of the previous comment; if something else is intended, it needs to be written down and analyzed.","section":"IV-B3, Eq. (26)"},{"comment":"The contraction proof assumes that psi is a global smooth diffeomorphism with pointwise nonsingular Jpsi and a uniformly positive-definite induced metric Gpsi=Jpsi^T Jpsi. The generator (22) contains the inverse barrier 1/(Gamma_SDF - tsave), which is singular on the obstacle boundary; global existence of the flow (24) is not established, Jpsi may become unbounded or singular in a neighborhood of the boundary, and uniform positive-definiteness of Gpsi is not checked. Consequently the claimed global contraction guarantee is not proven even for the pullback system considered in Appendix B4.","section":"Appendix B4, Eq. (24)"},{"comment":"The statement 'According to Theorem 1, contraction is preserved under affine transformations, consequently the friction (28) does not compromise the contraction guarantees' is unsupported: the factor eta_f ||fc(y)||/||fm(y)|| is a state-dependent scalar scaling of the vector field, not an affine feedback transformation as in Theorem 1(1), and positive scalar scaling does not in general preserve contraction. Since the friction variant is presented as part of the method (Figure 6c), the contraction guarantee for that variant needs a separate proof.","section":"IV-B5, Eq. (28)"}],"minor_comments":[{"comment":"The noise-injected loss in Eq. (14) is defined with a leading minus sign, so minimizing Lnoise maximizes the squared error; this is likely a sign error and should be corrected.","section":"Eq. (14)"},{"comment":"The symbol for the safety threshold appears as tsave in Eq. (19) and tsafe in Eq. (37); please unify.","section":"IV-B2, Eqs. (19) and (37)"},{"comment":"The notation psi(x, qt) is ambiguous because Eq. (24) takes (x, y) and returns a configuration q; clarify which variables are inputs and which are outputs in Algorithm 2.","section":"Algorithm 2, line 5"},{"comment":"The definition of RFC takes maxima over points x in tau_base and tau_m, but curvature is a function of time; the reparametrization and alignment of the two trajectories should be specified.","section":"V-A, Eq. (31)"},{"comment":"In the dynamic-obstacle barrier formula, the denominator |qSDT| should presumably be |q_dot_SDT|; the expression is dimensionally inconsistent as written.","section":"VII-B, Eq. (39)"},{"comment":"Several entries contain misplaced commas (e.g., '18 , 63' and '5 , 13') that should be decimal points.","section":"Table II"}],"recommendation":"reject","confidential_remarks":"The problem is structural: the system that is proved contractive is the pullback of the original dynamics, and therefore cannot produce the claimed obstacle avoidance; SDDC is not even defined as a dynamical system. Correcting this would require replacing the central construction with a pushforward, redoing the contraction analysis, and re-evaluating the experimental claims, which goes beyond a normal major revision. I therefore recommend rejection rather than major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the paper on contraction-preserving obstacle avoidance for NCDS. The short version: the core idea is attractive and the experiments are extensive, but the central theoretical claim as written is not supported by the equations. Eq. (27) defines fSDC(y) = J_psi(y)^{-1} fc(psi(y)), which is exactly the pullback of fc under the diffeomorphism psi. A pullback is just a coordinate change: if you integrate ẏ = fSDC(y) and set q = psi(y), then q satisfies q̇ = fc(q). So the physical trajectory in configuration space is the original unmodified trajectory. That cannot avoid an obstacle that fc would traverse. The obstacle-avoiding system you would want is the pushforward q̇ = J_psi(y) fc(y) with q = psi(y), but it is never defined or analyzed. For SDDC, Eq. (26) is only a virtual-displacement relation, not a dynamical system, and using it as ẏ = J_psi^{-1} fc(q) again gives a pullback. Appendix B4 proves contraction of the pullback, which is correct but irrelevant: it is stability of the wrong system. The figures do show avoidance, which makes me suspect the implementation actually uses something like the pushforward, but the paper does not say that and no code is provided to check.\n\nTo give credit where it is due: the combination of SDF/RDF/CDF with contraction-preserving diffeomorphic transforms is well-motivated and timely. The two new metrics, RFC and VM, are genuinely useful for comparing reactive avoidance methods. The experiments are thorough, including real-robot tasks, and the discussion of limitations (concave obstacles, discretization, solver trade-offs) is honest and useful.\n\nThe soft spots beyond the main flaw are minor by comparison: no error bars on the quantitative tables, no code or data release, and the global contraction claim assumes the flow psi is a global diffeomorphism with a uniformly positive-definite metric, which is not established. All of those are fixable. The pullback direction is not.\n\nI would send this to peer review because the combination of ideas deserves a serious look and this particular error is subtle enough that a careful referee might catch it. But my own recommendation would be major revision or reject in current form: the central guarantee does not follow. I would not cite it until the authors either replace the pullback with the pushforward and re-prove contraction, or clarify with equations and code exactly which system the robot actually executes.","headline":"The central avoidance claim is not supported by the equations: SDC/SDDC are defined as pullbacks, so they are trajectory-equivalent to the original dynamics and cannot avoid obstacles.","tokens_in":29378,"tokens_out":9344,"would_cite":false,"duration_ms":89942,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a contractive dynamical system can be reshaped around obstacles by a flow generated from a signed distance field, preserving the contraction guarantee through diffeomorphic coordinate changes.","keywords":["contraction theory","signed distance fields","diffeomorphic transforms","obstacle avoidance","learning from demonstration","neural contractive dynamical systems","robot manipulation","stability preservation"],"falsifier":"Take two nearby initial conditions of the modulated system near a concave obstacle, integrate them with a high-accuracy solver, and measure their separation over time; if it does not decay exponentially at the predicted contraction rate (or grows) before the obstacle is cleared, the preservation claim fails. Equivalently, evaluate $\\det J_\\psi$ along the flow on a grid approaching the obstacle surface: if it approaches zero or changes sign anywhere in the working domain, the flow is not a global diffeomorphism and the contraction-invariance theorem cannot be applied.","tokens_in":28383,"feed_emoji":"🤖","tokens_out":8793,"duration_ms":80386,"temperature":0.7,"pith_summary":"This paper claims that a contractive dynamical system, a robot skill whose trajectories converge exponentially to a learned motion, can be reshaped around obstacles without losing that convergence guarantee. The device is a flow generated by a barrier-weighted gradient of a signed distance field: the flow defines a diffeomorphism, and because contraction is invariant under diffeomorphic coordinate changes, the transformed system stays contractive. The paper introduces two versions, a differential coordinate change (SDDC) and a pullback coordinate change (SDC), and evaluates them on synthetic 2D skills and real kitchen manipulation tasks with whole-body obstacle avoidance. If the claim holds, learned skills can be given reactive safety without redesigning their stability certificates.","feed_headline":"Robot skills can dodge obstacles without losing contraction stability","feed_subtitle":"A signed-distance-field flow reshapes whole-body motion around obstacles while preserving exponential convergence.","key_machinery":"The load-bearing object is the flow-based diffeomorphism $\\psi$: the time-$t$ map of the vector field $V(x,q)=-b(x,q)\\nabla_q\\Gamma_{\\mathrm{SDF}}(x,q)$, where $\\Gamma_{\\mathrm{SDF}}$ is a signed distance field, a robot distance field (RDF), or a configuration-space distance field (CDF), and $b$ is an inverse barrier that diverges at the obstacle surface. The flow turns scalar distance information into a coordinate transformation, and its Jacobian $J_\\psi$ supplies the contraction metric $G_\\psi=J_\\psi^\\top J_\\psi$; contraction invariance under such coordinate changes is what lets the modulated system inherit the original skill's exponential convergence. A swept barrier term optionally masks the avoidance field where the motion moves away from the obstacle, and a friction term rescales the modulated velocity to better match the original skill.","core_discovery":"The central claim is that obstacle avoidance for contractive robot skills can be achieved by composing the skill dynamics with a diffeomorphism built from the environment's implicit signed distance field. Specifically, the gradient $\\nabla_q\\Gamma_{\\mathrm{SDF}}$ scaled by an inverse barrier $b_{\\mathrm{inv}}=s_{\\mathrm{grad}}/(\\Gamma_{\\mathrm{SDF}}-t_{\\mathrm{save}})$ defines an infinitesimal generator whose flow $\\psi(y)=y-\\int_0^t b\\,\\nabla_q\\Gamma_{\\mathrm{SDF}}\\,du$ is a candidate diffeomorphism. Transforming the contractive dynamics $f_c$ either as a differential coordinate change $\\delta y=J_\\psi^{-1}\\delta q$ (SDDC) or as a pullback $\\dot y=J_\\psi^{-1}f_c(\\psi(y))$ with Riemannian metric $G_\\psi=J_\\psi^\\top J_\\psi$ (SDC) preserves contraction by Theorem 1, so the robot avoids obstacles while retaining exponential convergence to the intended motion. The paper also claims that this reshaping stays close to the learned vector field, as measured by its new curvature and misalignment metrics, and demonstrates the approach on learned whole-body kitchen skills.","pith_inferences":["The construction implies that the contraction-preservation guarantee is inherited only on the region where the flow is a true diffeomorphism; a practical implementation could monitor the condition of the flow's derivative online and treat a near-singular Jacobian as a signal to stop modulating.","Because the method works with any signed-distance representation, a natural extension is to learn the scene as a neural signed distance field and use the same flow for mobile manipulation; the paper mentions this direction but does not test it.","The relative-motion barrier sketched in the future-work section suggests a concrete extension to moving obstacles: the same flow construction with an obstacle-velocity-dependent barrier would give reactive avoidance in dynamic scenes, with the contraction guarantee still riding on the flow being a diffeomorphism."],"forward_implications":["A contractive skill learned in joint space can be given whole-body obstacle avoidance without retraining the skill or changing its learned vector field.","The approach applies to any contractive dynamical system, not only the neural contractive systems used in the experiments, because the preservation argument is purely geometric.","Obstacle avoidance can be quantified beyond minimum distance: the proposed flow-curvature and vector-field-misalignment metrics allow direct comparison of how much a modulation distorts the original skill.","Real-time execution is feasible for cluttered scenes when the implicit distance is queried at the nearest point of the point cloud, as shown in the kitchen experiments.","The guarantees and reliable behavior are demonstrated for convex obstacles; concave obstacle shapes can trap the flow and stall the skill."],"supporting_citations":[{"why":"Supplies the neural contractive dynamical system used as the base contractive skill being transformed.","marker":"[4]"},{"why":"Shows that modulation-matrix obstacle avoidance can preserve contraction, the prior result this work extends to implicit signed distance representations.","marker":"[18]"},{"why":"Provides the modulation-matrix baseline and the friction-term idea used to preserve the underlying velocity profile.","marker":"[19]"},{"why":"Defines contraction stability and the exponential-convergence criterion that the proposed transforms aim to preserve.","marker":"[38]"},{"why":"Establishes that contraction is invariant under differential coordinate changes, the theorem directly used to certify the transformed systems.","marker":"[41]"},{"why":"Provides the contraction conditions used to verify that the transformed system with metric $G_\\psi$ remains contractive.","marker":"[67]"},{"why":"Introduces flow-based diffeomorphic transforms for stable obstacle avoidance, the approach the paper adapts and extends to implicit signed distance fields.","marker":"[71]"},{"why":"Provides the Eikonal and implicit geometric regularization losses used to learn smooth signed distance fields.","marker":"[15]"},{"why":"Introduces configuration-space distance fields, one of the implicit robot representations used in the real robot experiments.","marker":"[35]"},{"why":"Introduces robot signed distance fields via smooth polynomial link models, used for whole-body distance queries in the experiments.","marker":"[36]"}],"fun_headline_variants":["SDF warps robot dynamics to dodge obstacles safely","Obstacle dodging that keeps robot convergence intact","Flow-based avoidance preserves contraction in robot skills","Contractive skills orbit obstacles via SDF flows","Implicit surfaces reshape motion without losing stability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on assuming the obstacle-avoidance flow is a smooth, invertible coordinate change over the entire working area, with a well-conditioned derivative everywhere and a uniformly well-behaved induced metric; the paper asserts these properties rather than proving them, even though the barrier that generates the flow becomes singular exactly at the obstacle surface.","fun_headline_variants_meta":{"raw":{"variants":["SDF warps robot dynamics to dodge obstacles safely","Obstacle dodging that keeps robot convergence intact","Flow-based avoidance preserves contraction in robot skills","Contractive skills orbit obstacles via SDF flows","Implicit surfaces reshape motion without losing stability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000492,"raw_usage":{"total_tokens":2436,"prompt_tokens":983,"completion_tokens":1453,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":1381}},"tokens_in":599,"tokens_out":1453,"duration_ms":8749,"temperature":1.0,"reasoning_tokens":1381,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:08:51.367482+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two nearby initial conditions of the modulated system near a concave obstacle, integrate them with a high-accuracy solver, and measure their separation over time; if it does not decay exponentially at the predicted contraction rate (or grows) before the obstacle is cleared, the preservation claim fails. Equivalently, evaluate $\\det J_\\psi$ along the flow on a grid approaching the obstacle surface: if it approaches zero or changes sign anywhere in the working domain, the flow is not a global diffeomorphism and the contraction-invariance theorem cannot be applied.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the contraction conditions used to verify that the transformed system with metric $G_\\psi$ remains contractive."},{"cited_title":"Diffeomorphic transforms for generalised imitation learn- ing","cited_arxiv_id":null,"evidence_quote":"Introduces flow-based diffeomorphic transforms for stable obstacle avoidance, the approach the paper adapts and extends to implicit signed distance fields."}],"review_version":1}