{"id":"df1e3bf4-9467-447f-aa9c-d67ef31d8c80","arxiv_id":"2507.10602","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Orbitally Stable Motion Primitives learn periodic robot motions with a stable latent oscillator, a learned invertible encoder, and task conditioning, claiming exponential orbital stability and smooth interpolation between behaviors.","lead":"The paper trains robot motion policies for periodic tasks by routing demonstrations through a learned invertible mapping into a latent space with a stable oscillator, and adds task conditioning so one policy can switch between behaviors. The authors claim formal exponential convergence to the learned cycle and demonstrate the approach on arms, a soft robot, and a swimming turtle robot.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The exponential contraction claim is unsupported: Proposition 1's variational vector δ=c(1,0,1,...,1) only satisfies the orthogonality condition δᵀM f=0 when y_{3:n}=0, so the contraction inequality is applied to a non-admissible perturbation; the gap propagates to Theorem 1.","rationale":"The reader's weakest_assumption matches my own: the proof of exponential stability hinges on exact orthogonality in the transverse contraction criterion, and uniform convergence of a residual along the flow is not a substitute for the pointwise algebraic condition. The paper does contain an independent Lyapunov proof for asymptotic orbital stability, so that weaker claim is supported; but the abstract and Theorem 1 promise exponential convergence, which is the claimed differentiator from prior work. The Lsci-based interpolation claim is empirically supported but not formal; I regard the contraction proof gap as more load-bearing. I recommend keeping the current conditional verdict: the method and experiments are promising, but the headline guarantee should be downgraded to asymptotic orbital stability unless Proposition 1 is repaired. Since my read agrees with the reader's weakest_assumption and does not change the verdict, no adjustment is needed.","tokens_in":49932,"tokens_out":5569,"duration_ms":64604,"concrete_test":"Independently re-derive Proposition 1 without the asserted choice of δ: for a general state (r, φ, y_{3:n}) with r>0 and y_{3:n}≠0, solve the orthogonality constraint δᵀ M_pol f_pol = 0 for δ_φ in terms of δ_{3:n}, substitute the resulting admissible family into the left-hand side of Eq. S61, and check its sign over Y_pol. If any admissible δ yields a positive value, the transverse contraction and exponential orbital stability claim fails; if the quantity is non-positive for all admissible δ, the proof can be repaired by making this family explicit. A numerical counterpart is to sample (r, φ, y_{3:n}) and admissible δ uniformly and evaluate the quadratic form.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1's Theorem 1 claims transverse contraction and hence exponential orbital stability. The load-bearing step is Proposition 1 in the Supplementary Text. Definition 1 requires the contraction inequality to hold for every variational vector δ with δᵀ M_pol f_pol = 0 at each point of the region. The proof instead fixes δ = c(1,0,1,...,1)ᵀ and computes δᵀ M_pol f_pol = −cβ y_{3:n} (Eq. S60), then asserts the condition is fulfilled because −cβ y_{3:n} converges uniformly to zero. That is a statement along the flow, not a pointwise algebraic condition: at any state with y_{3:n} ≠ 0, the chosen δ is not orthogonal to f_pol, so the inequality in Eq. S61 is evaluated for a vector outside the admissible set. The same gap is inherited by Proposition 2 through the coordinate change and by Theorem 1. The asymptotic orbital stability proof (Theorem 2) is a separate Lyapunov argument and is not affected; what is unsupported is the stronger exponential rate emphasized in the abstract and Section 2.1.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Orbitally Stable Motion Primitives (OSMPs), a dynamical-system motion policy that combines a task-conditioned bijective encoder (Euclideanizing flow) with latent supercritical Hopf-oscillator dynamics. The paper claims two formal results: asymptotic orbital stability (Theorem 2) and, when the velocity scaling is f_s=1, transverse contraction implying exponential orbital stability (Theorem 1), with proofs relegated to Supplementary Propositions 1 and 2. The method is evaluated against MLP, RNN, LSTM, NODE, Diffusion Policy, iFlow, and SPDT baselines on IROS letters, Drawing2D, image-contour, and turtle-swimming oracles, and is deployed on the UR5, KUKA, Helix soft robot, and Crush turtle robot. Additional contributions include phase synchronization of multiple primitives, online velocity-field shaping, and encoder conditioning with a smooth-interpolation loss (L_sci).","tokens_in":50227,"tokens_out":5985,"duration_ms":75692,"significance":"If the exponential-contraction guarantee were fully established, this would be a meaningful advance over prior orbitally stable DMPs: a single demonstration could yield a policy with a built-in limit cycle and guaranteed exponential convergence, and the empirical study is unusually broad, including real-hardware results, ablations, and an inference-time analysis. The asymptotic orbital stability proof (Theorem 2) is a useful formalization, and the phase-synchronization and online-shaping components are practically valuable. However, the central exponential claim rests on a gap in the transverse-contraction proof, and the zero-shot interpolation claim is overstated because L_sci trains directly on the tested conditioning values. Both issues are localized and potentially fixable, so the appropriate disposition is major revision rather than rejection.","major_comments":[{"comment":"The orthogonality condition required by Definition 1 is not established. Definition 1 demands that the contraction inequality hold for every variational vector δ with δ^T M f = 0 at each point of the region. The proof instead fixes δ = c(1, 0, 1_{n-2}^T)^T and computes δ^T M_pol f_pol = -c β y_{3:n}, then asserts that the orthogonality condition is fulfilled because -c β y_{3:n} converges uniformly to zero along the flow. That is a statement about the limit as t → ∞, not a pointwise algebraic condition. At any state with y_{3:n} ≠ 0, the chosen δ is not orthogonal to f_pol, so inequality (S61) is applied to a non-admissible perturbation. Consequently, transverse contraction of the polar latent dynamics is not proven, and Theorem 1's exponential orbital stability claim is unsupported as written.","section":"Supplementary Proposition 1, Step 2 (Eq. S60)"},{"comment":"Proposition 2 inherits the same orthogonality gap. The Cartesian variational vector δ_y = c(y_1/r, y_2/r, 1_{n-2})^T is obtained from the same non-orthogonal polar vector, and the computation δ_y^T M_y f_y = -c β y_{3:n} again only vanishes along the flow, not at every point of the region. Since Proposition 2 is the bridge from polar to Cartesian coordinates in the proof of Theorem 1, the exponential contraction conclusion is not established in either coordinate system.","section":"Supplementary Proposition 2 (Eqs. S66-S67)"},{"comment":"The theorem's region X = {x : sqrt(Ψ_1^2 + Ψ_2^2) > 0} corresponds to r > 0, but Proposition 1 states its contraction-rate lower bound only for r ≥ r_ε, with r_ε > 0. For 0 < r < r_ε, no uniform exponential rate is provided, so the claimed 'almost-global exponential orbital stability' does not follow from the supplied proof. The authors should either restrict the exponential claim to a region where a uniform rate is proven or provide a separate argument covering the neighborhood of the origin.","section":"Section 2.1, Theorem 1 and Supplementary Proposition 1 (region mismatch)"},{"comment":"The 'zero-shot generalization to unseen motion objectives' claim is circular in an important sense. The smooth conditioning interpolation loss L_sci samples conditioning values z̃ uniformly from the convex hull conv(Z) during training and supervises the encoder output at those values against a linear interpolation of the training endpoints. The values marked 'unseen' in Fig. 7D, such as z = -0.5 and z = 0.5, lie inside conv(Z) and are therefore directly optimized during L_sci training. If the authors intend 'within the training distribution' to include all hull points, then the experiment demonstrates interpolation to points that were seen by the loss, not zero-shot generalization; the terminology should be revised or the test values should be genuinely held out from L_sci.","section":"Section 2.6 and Supplementary Eq. (S12)-(S16)"}],"minor_comments":[{"comment":"The abstract's 'consistently outperforming state-of-the-art baselines such as diffusion policies' is stronger than the data support. In the IROS Letters row of Table 1, OSMP has higher trajectory RMSE (0.344) and velocity RMSE (0.895) than both MLP (0.257, 0.630) and Diffusion Policy (0.255, 0.648). The Discussion acknowledges these exceptions, but the abstract and Section 2.2 should be qualified accordingly.","section":"Abstract and Table 1"},{"comment":"The expression for the smallest eigenvalue condition in the positive-definiteness proof of M_pol is difficult to verify because of the nested parentheses and the notation m_φφ appearing on both sides of the inequality. The derivation should be rewritten more transparently or moved to a supporting calculation.","section":"Supplementary Eq. (S56)"},{"comment":"The phase-synchronization law multiplies the angular velocity by the factor 1 - k_ps Σ_j sin(δΦ*_ij + iφ - jφ). For sufficiently large k_ps or large phase errors, this factor can become nonpositive, which would violate the assumption ω > 0 used in the latent-dynamics analysis. A bound on k_ps or a projection that preserves positivity should be stated.","section":"Section 4.3, Eq. (6)"},{"comment":"The labels 'unseen during training' in Fig. 7D should be reconciled with the fact that L_sci samples uniformly over conv(Z); as written, the caption overstates the novelty of the interpolated conditioning values.","section":"Figure 7D"}],"recommendation":"major_revision","confidential_remarks":"The paper has a substantial empirical component and the asymptotic orbital stability argument is plausible, but the exponential contraction proof contains a real gap in the orthogonality step and the zero-shot claim is overstated. Both issues are localized and likely repairable: the authors need either a pointwise-valid orthogonality argument or a reformulated theorem, and they should either hold out the interpolation values from L_sci or change the 'zero-shot' terminology. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid, well-executed extension of the SPDT line of work, with one genuinely important theoretical claim that is not yet established. The empirical side is the strongest part: extensive benchmarks across drawing, image contours, and turtle swimming, plus real-robot demos on UR5, KUKA, Helix, and the turtle. The new loss terms—limit-cycle matching, time guidance, velocity regularization, and the conditioning-interpolation loss—are sensible and the ablations show they pull real weight. The phase synchronization and online shaping are useful additions. If I worked on rhythmic DMPs, I would read this carefully and very likely cite it for the empirical comparison alone. The authors also write an unusually honest limitations section, which I appreciate.\n\nNow the soft spots, in proportion. The main issue is the transverse contraction proof. The stress-test note is correct: Proposition 1 fixes a variational vector delta = c(1,0,1,...,1) and says the orthogonality condition is fulfilled because -c*beta*y_{3:n} converges uniformly to zero. But transverse contraction requires delta^T M f = 0 pointwise at every state in the region. At any state with y_{3:n} nonzero, that vector is not orthogonal to the flow, so the contraction inequality is being evaluated outside the admissible set. That gap propagates to Theorem 1. The asymptotic orbital stability result in Theorem 2 is a separate Lyapunov argument and is not affected; what is unsupported is the exponential rate that the abstract and Section 2.1 emphasize. This is fixable in principle—maybe a different choice of variational vector or a modified metric can close it—but as written the strongest claim does not go through.\n\nSecond, the zero-shot interpolation claim is oversold. The L_sci loss trains the encoder at conditioning values sampled continuously across the hull between the trained anchors, so evaluating at z=-0.5 or z=0.5 is not really zero-shot; the model has been trained there. The interpolation behavior is still a nice empirical result, but calling it zero-shot is misleading.\n\nReproducibility is currently weak: the paper promises open-source code and data but releases nothing yet. That is a moderate concern, not a fatal one.\n\nMy bottom line: this deserves a serious referee. The empirical contribution is strong enough to publish, and the theoretical gap is concrete and addressable. I would send it to review with a clear request: either repair the transverse contraction proof or downgrade the claim to asymptotic orbital stability and soften the abstract. I would also ask for the interpolation claim to be reworded as trained interpolation within the conditioning hull, not zero-shot.\n\nRecommendation: accept for peer review, conditional on the proof gap being fixed or the claim being scaled back.","headline":"OSMP extends SPDT with real empirical gains, but the headline exponential orbital stability guarantee rests on a gap in the transverse contraction proof and the zero-shot interpolation claim is diluted by the training loss.","tokens_in":50798,"tokens_out":1275,"would_cite":true,"duration_ms":19559,"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":"Orbitally Stable Motion Primitives: learned periodic robot motions with exponential convergence and zero-shot task interpolation.","keywords":["orbital stability","limit cycles","imitation learning","dynamic movement primitives","transverse contraction","Hopf bifurcation","task conditioning","zero-shot generalization"],"falsifier":"Evaluate the transverse-contraction inequality of Proposition 1 at a fixed polar state with $y_{3:n} \\neq 0$ and $r \\neq R$: if the left-hand side is positive there, the exponential contraction claim is not established. In simulation, initialize two nearby trajectories with the same phase but different transverse coordinates; if their transverse separation does not decay at the predicted exponential rate $e^{-\\zeta t}$, the central guarantee fails.","tokens_in":49716,"feed_emoji":"🤖","tokens_out":9839,"duration_ms":107321,"temperature":0.7,"pith_summary":"Orbitally Stable Motion Primitives (OSMPs) aim to turn a single periodic demonstration into a robot controller with a formal guarantee: every trajectory converges exponentially to the learned limit cycle, not merely asymptotically. The construction maps robot configurations through a bijective, task-conditioned encoder into a latent oscillator built from a supercritical Hopf bifurcation, then pulls the latent velocity back to the robot state space. The paper further claims that one policy, conditioned on a scalar task variable $z$, can reproduce several motions and interpolate smoothly between them, giving consistent zero-shot behavior on unseen tasks inside the training distribution. If these claims hold, rhythmic tasks such as wiping, brooming, and swimming become learnable from very little data while staying robust to disturbances and predictable under human interaction.","feed_headline":"Robots learn periodic motions with exponential stability from one demo","feed_subtitle":"Motion primitives promise fast disturbance recovery, smooth task switching, and synchronized multi-limb locomotion.","key_machinery":"The load-bearing object is the latent-space oscillator in polar coordinates, $$\\dot{y}_{\\mathrm{pol}} = (\\$\\alpha$(1-$r^{2}$/$R^{2}$)r,\\; f_\\omega(\\varphi),\\;-\\$\\beta$ y_{3:n}),$$ the polar form of a supercritical Hopf bifurcation: radius $r$ is driven toward $R$ by gain $\\alpha$, the transverse coordinates $y_{3:n}$ decay with gain $\\beta$, and motion around the cycle is governed by a learned positive angular velocity $f_\\omega(\\varphi)$. A bijective Euclideanizing-flow encoder $\\Psi(x;z)$ -- a composition of invertible coupling layers -- maps the robot configuration $x$ into this latent space, and the policy velocity is the pullback $\\dot{x} = f_s(x)J_\\Psi^{-1} f_y(\\Psi(x;z))$. The convergence proof runs through a transverse contraction metric in polar coordinates, lifted through the polar-to-Cartesian map and then through the diffeomorphism; Theorem 1 packages that chain into almost-global exponential orbital stability. The mechanism for multi-task behavior is conditioning: the encoder depends on a scalar task parameter $z$, and the smooth-conditioning-interpolation loss $\\mathcal{L}_{\\mathrm{sci}}$ enforces that interpolated latent-space cycles map to linearly interpolated oracle-space cycles.","core_discovery":"On the paper's own terms, the central discovery is that a diffeomorphic encoder combined with a supercritical Hopf oscillator in latent space provides both the expressive freedom to reproduce complex periodic trajectories and the structure to prove strong convergence. Theorem 1 states that, whenever the velocity scaling factor is set to $f_s(x)=1$, the OSMP dynamics are transverse contracting almost everywhere, and transverse contraction implies exponential orbital stability: trajectories converge to the periodic orbit at an exponential rate from any initial state outside the center of the cycle. This strengthens the previously asymptotic orbital-stability observation for such systems into an exponential one, under mild architectural assumptions. A second claimed result is that conditioning the encoder on a scalar task variable, with a dedicated interpolation loss, makes the learned limit cycle vary smoothly with the conditioning, so a single policy can transition between behaviors and generalize zero-shot within the training distribution. The paper supports these claims with benchmark rollouts, ablations, and hardware experiments on several robot platforms.","pith_inferences":["Editorially, if the exponential guarantee is genuine, OSMPs should remain stable under piecewise-constant switching of the conditioning $z(t)$; the paper names this extension as future work, and it would turn the interpolation result into a discrete mode-switching controller.","Beyond the paper, the scalar conditioning channel is a minimal case of a broader idea: replacing $z$ with an observation or language embedding would let the same orbital-stability machinery select behaviors from perception, a direction the paper flags as future work.","Since contraction is preserved under interconnection, separate OSMPs synchronized by the phase-feedback law should scale toward high-degree-of-freedom locomotion, with each limb or module kept on its own cycle while only the phase is coupled."],"forward_implications":["A single demonstration suffices to train a working rhythmic policy, with exponential return to the cycle after disturbances.","Because the policy is time-invariant, phase shifts such as a half-period offset do not derail it, unlike classical time-indexed trajectory tracking.","Multiple OSMPs can be phase-locked by adjusting only their polar angular velocities, preserving each primitive's spatial motion and convergence properties.","A single task-conditioned policy can morph between learned motions and execute unseen motions inside the training distribution without retraining.","Exponential transverse contraction makes OSMPs composable: interconnecting them with other contracting systems should preserve stability, supporting modular and hierarchical control."],"supporting_citations":[{"why":"Supplies the transverse contraction criterion and the definition of exponential orbital stability that Theorem 1 relies on.","marker":"[35]"},{"why":"Supplies the Euclideanizing-flow bijective encoder and the pullback velocity construction.","marker":"[13]"},{"why":"Closest prior rhythmic-diffeomorphism architecture whose training losses the paper replaces with velocity imitation and limit-cycle matching.","marker":"[22]"},{"why":"Provides the Imitation Flow baseline with stochastic latent limit-cycle dynamics that OSMPs are benchmarked against.","marker":"[14]"},{"why":"Proves transverse contraction for the Andronov-Hopf oscillator, the argument Proposition 1 adapts for the polar latent dynamics.","marker":"[23]"},{"why":"Foundational contraction analysis that supplies the rate and composition background for the exponential claim.","marker":"[34]"},{"why":"Source for the supercritical Hopf bifurcation dynamics used to shape the latent space.","marker":"[30]"},{"why":"Supplies the Real NVP coupling-layer invertibility that underlies the bijective encoder.","marker":"[31]"},{"why":"Provides the Diffusion Policy baseline and the Gaussian Fourier feature embedding used for conditioning.","marker":"[3]"}],"fun_headline_variants":["Exponential orbital stability for learned rhythmic robot motions","Stable periodic robot motions with zero-shot task switching","One policy for many rhythmic tasks with orbital stability","Orbital stability guarantees for learned periodic motion primitives","Task-conditioned Hopf oscillators for stable robot locomotion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The exponential-convergence guarantee rests on treating a chosen variation between two nearby trajectories as exactly perpendicular to the cycle's motion, but at any state with nonzero transverse coordinates that perpendicularity is only approximate and becomes exact only as the coordinates shrink to zero.","fun_headline_variants_meta":{"raw":{"variants":["Exponential orbital stability for learned rhythmic robot motions","Stable periodic robot motions with zero-shot task switching","One policy for many rhythmic tasks with orbital stability","Orbital stability guarantees for learned periodic motion primitives","Task-conditioned Hopf oscillators for stable robot locomotion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1308,"prompt_tokens":964,"completion_tokens":344,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":268}},"tokens_in":580,"tokens_out":344,"duration_ms":4230,"temperature":1.0,"reasoning_tokens":268,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:59:35.737974+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the transverse-contraction inequality of Proposition 1 at a fixed polar state with $y_{3:n} \\neq 0$ and $r \\neq R$: if the left-hand side is positive there, the exponential contraction claim is not established. In simulation, initialize two nearby trajectories with the same phase but different transverse coordinates; if their transverse separation does not decay at the predicted exponential rate $e^{-\\zeta t}$, the central guarantee fails.","supporting_citations":[],"review_version":1}