{"id":"4b756307-449e-41b2-b2b6-c60ae8401d6a","arxiv_id":"2411.10336","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"B-splines are reformulated as probabilistic movement primitives, giving trajectory distributions that can satisfy arbitrary boundary conditions while retaining B-spline local support.","lead":"This paper introduces B-spline Movement Primitives (BMPs), a way to write B-spline curves as probabilistic movement primitives so robots can satisfy start and end conditions while modeling a distribution of motions. It tests BMPs on digit-writing imitation learning and a simulated box-pushing reinforcement learning task, reporting that BMPs match or beat a recent baseline while adding explicit boundary control.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (11) and (12) omit the 1/T phase-to-time conversion that Eq. (9) itself requires, so exact velocity boundary conditions are off by the movement duration unless T=1.","rationale":"I read the paper in good faith. The core B-spline-to-MP rewrite is standard, and the clamped-endpoint interpolation is correctly described. The Gaussian extension in Eq. (13) is a legitimate modeling choice inherited from ProMP; it does not by itself threaten the central claim because the paper does not promise to model non-Gaussian trajectory distributions, and all MPs in this family share that linear-Gaussian structure. The most load-bearing issue is instead the velocity boundary condition. The paper's own Eq. (9) statement says phase velocity must be divided by duration T to get real velocity, yet Eq. (11) and Eq. (12) set derivative control points equal to the desired velocity without that factor. For non-unit T this changes the actual initial and end velocities by exactly T, so the advertised exact boundary-condition satisfaction fails in general. The zero-velocity constraints used in the box-pushing experiment are insensitive to this factor, which may explain why the error went unnoticed, but the paper's central claim is broader. The fix is elementary (insert 1/T), and the empirical conclusions for the tested zero-velocity case likely survive, so the appropriate verdict remains conditional rather than reject. The reader flagged velocity verification as an empirical fragility; I agree partially, but I locate a concrete algebraic error rather than a missing hardware experiment.","tokens_in":8727,"tokens_out":10022,"duration_ms":98052,"concrete_test":"Instantiate a degree-2 clamped B-spline with n=4 control points, knot spacing Delta=0.5, and movement duration T=2. Set desired real initial velocity v_d=1 and c_0=0. Using the paper's Eq. (11), c_1 = Delta/p * v_d + c_0 = 0.5/2 = 0.25. Evaluate the actual real initial velocity: dy/du at u=0 equals p/Delta * (c_1 - c_0) = 2/0.5 * 0.25 = 1, so dy/dt = 1/T = 0.5, not 1. Recompute with c_1 = c_0 + Delta/(p*T) * v_d = 0.125; then dy/du = 2/0.5 * 0.125 = 0.5 and dy/dt = 1. This single numerical check settles whether Eq. (11) needs the 1/T factor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (10) defines c^(1)_i as derivative control points with respect to phase u, and the text explicitly says these phase velocities must be divided by duration T to obtain real velocities. Equation (11) then imposes an initial velocity boundary condition as c^(1)_0 = p/Delta * (c_1 - c_0) = dot_y0 and solves c_1 = Delta/p * dot_y0 + c_0. If dot_y0 denotes a physical desired velocity, the chain rule dy/dt = (1/T) dy/du gives the actual initial real velocity as c^(1)_0 / T = dot_y0 / T, so the constraint is violated for any T != 1. The same missing factor appears in Eq. (12) for the end-velocity condition. If dot_y0 is instead intended as a phase velocity, then the paper never shows how to impose a real velocity boundary condition, and the robot experiments (e.g., Fig. 5b, which reports velocities at 0 s and 2 s) are not supported by the stated equations. Since exact boundary-velocity satisfaction is one of the two main pillars of the central claim, this is a load-bearing correctness issue, not merely a missing hardware experiment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces B-spline Movement Primitives (BMPs), a representation that casts clamped uniform B-splines as probabilistic movement primitives in the style of ProMP. Trajectories are written as linear combinations of B-spline basis functions with control-point weights, and a Gaussian distribution over the interior control points yields a Gaussian trajectory distribution (Eq. 13). Boundary conditions on position and velocity are enforced by fixing the first and last one or two control points via linear equations (Eqs. 11-12). The paper reports three sets of experiments: a regression comparison of B-spline versus ProMP/ProDMP on digit-writing trajectories, a toy planning task, an IL experiment with a conditional encoder-decoder for digit images, and an episodic RL experiment on a box-pushing task comparing BMP with ProDMP under BBRL and TCE.","tokens_in":8936,"tokens_out":10068,"duration_ms":85966,"significance":"The core idea is attractive and the mathematical development in Section 3 is mostly sound: the Cox-de Boor recursion, derivative relation (Eqs. 9-10), and clamped-endpoint construction are correctly stated, and the formulation as a linear-Gaussian trajectory distribution is a natural and potentially useful bridge between B-spline planning and probabilistic MP learning. The regression and RL results are suggestive of practical benefits. However, the central claim of exact boundary-velocity satisfaction is undermined by a missing phase-to-time conversion factor in the boundary-condition equations, and the IL experiment does not include any baseline MP. These issues need to be resolved before the paper's claims are fully supported.","major_comments":[{"comment":"The boundary-condition equations are off by a factor of 1/T. The text after Eq. (9) explicitly states that the phase velocity c^(1) must be divided by the duration T to obtain real velocity. Yet Eq. (11) sets c^(1)_0 = ˙y0 and Eq. (12) sets c^(1)_{n−2} = ˙ye, where ˙y0 and ˙ye are described as desired initial and end velocities. If these are physical velocities, the actual real velocity at the start is c^(1)_0 / T = ˙y0 / T, so the condition is satisfied only for T = 1; if they are phase velocities, the paper never provides a mechanism to impose a real velocity boundary condition. This directly affects the central claim that BMPs retain their ability to satisfy boundary conditions and the velocity demonstration in Fig. 5(b). The equations should use T ˙y0 and T ˙ye (or equivalently define c^(1) as time-derivative control points), and the authors should verify that the experiments used the corrected formula.","section":"Section 3, Eqs. (11)-(12)"},{"comment":"The digit-writing IL experiment reports only the BMP model and does not include ProMP or ProDMP baselines trained with the same encoder-decoder architecture and the same pairwise log-likelihood loss. As a result, this experiment cannot support the abstract's claim that BMPs offer greater expressiveness compared to existing MP variants in an IL setting. Adding at least a ProMP baseline (and ideally ProDMP) is necessary to substantiate the comparative claim, or the claim should be explicitly restricted to the regression and RL experiments.","section":"Section 4.2, Imitation Learning"}],"minor_comments":[{"comment":"The noise term is written as \"ϵy ∼ (0, σy)\" without the normal-distribution symbol; it should be \"ϵy ∼ N(0, σy)\".","section":"Section 2, Eq. (2)"},{"comment":"The definition of N_{i,0} is garbled: \"if ui ≤ u < ui+1, if ui ≤ u ≤ ui+1 for i = n − 1\" should be \"if ui ≤ u < ui+1, and if ui ≤ u ≤ ui+1 for i = m − 1\" (or similar).","section":"Section 3, Eq. (6)"},{"comment":"The notation \"u0, u1, ..., up = 0\" is imprecise; it should be \"u0 = u1 = ... = up = 0\", and similarly for the right end. Also, \"i ∈ [p, ..., m-p-1]\" is nonstandard notation for an index range.","section":"Section 3, after Eq. (8)"},{"comment":"The index set [2:n-3] is ambiguous; with n control points and four boundary constraints, the free control points are c_2, ..., c_{n-3}, and the notation should be defined explicitly as inclusive of both endpoints or as a set.","section":"Section 3, Eq. (13)"},{"comment":"The basis-function notation N_{1,0} and N_{2,0} used for higher-degree basis functions is inconsistent with the standard B-spline notation; the figure and text should use N_{0,1} and N_{0,2} (or clearly redefine the indexing).","section":"Section 3, Fig. 1(a) and text"},{"comment":"The acknowledgments section contains placeholder boilerplate (\"If a paper is accepted, the final camera-ready version will (and probably should) include acknowledgments...\") that should be removed or replaced with actual acknowledgments.","section":"Acknowledgments"}],"recommendation":"major_revision","confidential_remarks":"The missing 1/T factor in the boundary-velocity equations is a serious but easily correctable error; I would ask the authors to fix it and to confirm that their experiments (especially Fig. 5b) used the corrected formula. The IL experiment without baselines is also a gap that should be addressed either experimentally or by softening the claims. The paper is otherwise a solid contribution that could be suitable for publication after these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper does something natural and useful—writes uniform B-splines in ProMP form, then uses clamped control points to enforce start/end conditions. The Cox-de Boor material is correct, Eq (13) is a legitimate Gaussian trajectory distribution, and the idea of learning residuals from a constant line is pragmatic. I'd have been happy to see this in a movement-primitives venue.\n\nBut there is a load-bearing error in Section 3. Eq (9) gives velocities with respect to phase u, and the text correctly notes these must be divided by T to get real velocities. Then Eq (11) sets c^(1)_0 = dot_y0, and Eq (12) does the same at the end, treating the phase velocity as if it were the physical velocity. For T != 1, the imposed physical velocity is actually dot_y0/T. So the paper's advertised exact boundary-velocity satisfaction is off by a factor of T. If the authors meant dot_y0 to be a phase velocity, they never show how to prescribe a true physical velocity. This is not a typo; it breaks a central claim. The fix is easy (divide or multiply by T), but as written the equations are wrong.\n\nThe experimental section doesn't rescue it. The IL experiment has no ProMP or ProDMP baseline, so 'greater expressiveness' in generative modeling isn't demonstrated. The RL experiment uses one task, four seeds, no visible error bars, and no code or data. There's also an internal inconsistency: the text says TCE 'can only apply to BMP' and then discusses ProDMP's TCE results a few lines later. The conclusion overstates—'B-spline is generally a better trajectory representation' doesn't follow from these experiments.\n\nWhat the paper has going for it: a clean, checkable derivation that connects B-splines to a distribution; local support is a real advantage for episodic RL; and the boundary-condition construction would be valuable once the velocity conversion is fixed. The math is mostly transparent, and the mistake is identifiable and correctable.\n\nWho it's for: people building on ProMP and ProDMP, especially in IL/RL with reset conditions. I'd send it to a serious referee, but with a clear request to verify the boundary-velocity equations and to get a baseline in the IL experiment. Without those fixes, I wouldn't cite it.","headline":"A useful but flawed twist on ProMP: the boundary-velocity equations miss a factor of 1/T, so the paper's central guarantee doesn't hold as stated.","tokens_in":9507,"tokens_out":3749,"would_cite":false,"duration_ms":35518,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By rewriting B-splines as movement primitives with control-point weights, BMPs let B-splines model trajectory distributions for imitation and reinforcement learning while exactly satisfying start and end position and velocity conditions.","keywords":["B-spline Movement Primitives","Probabilistic Movement Primitives","Imitation Learning","Episodic Reinforcement Learning","Trajectory Distributions","Boundary Conditions","Robot Learning"],"falsifier":"Run BMP on a demonstration set with two clearly separated movement modes, such as reaching above and below an obstacle, and evaluate the likelihood of the demonstrations under the fitted Gaussian trajectory distribution; with distinct modes the fitted Gaussian will place most probability mass between them, yielding low likelihood. A second check is a physical robot experiment that measures the achieved end velocity and compares it to the prescribed value, because the paper derives velocities from a phase derivative divided by duration $T$ and provides no hardware verification that exact end-velocity conditions are met.","tokens_in":8492,"feed_emoji":"🤖","tokens_out":7315,"duration_ms":59911,"temperature":0.7,"pith_summary":"The paper introduces B-spline Movement Primitives (BMPs), a representation that combines the boundary-condition guarantees of B-splines with the probabilistic modeling of ProMPs. The core move is to treat B-spline control points as the weights in a linear-Gaussian trajectory model, so the distribution over trajectories is Gaussian with mean and covariance given by the spline basis applied to a learned distribution over control points. Boundary conditions are enforced deterministically by fixing the first and last control points through linear equations, so every sample from the distribution respects the specified start and end positions and velocities. If correct, BMP gives imitation and episodic reinforcement learning a trajectory representation that is both expressive and constraint-satisfying, which the paper demonstrates on digit-writing, goal-reaching, and robot box-pushing tasks.","feed_headline":"B-splines gain probabilistic learning without losing exact endpoints","feed_subtitle":"New movement primitive models trajectory uncertainty while satisfying boundary conditions for robot learning.","key_machinery":"The central object is the clamped uniform B-spline, defined by the Cox–de Boor recursion over a knot vector with repeated end knots so that the curve starts at the first control point and ends at the last. Its basis functions have local support, its derivative is another B-spline whose control points are differences of the original ones via $c^{(1)}_i = \\frac{p}{\\Delta}(c_{i+1} - c_i)$, and its convex hull property bounds the curve between its control points. Treating the interior control points as the weights of a ProMP-style linear-Gaussian model yields a trajectory distribution whose mean and covariance are computed by propagating a Gaussian over those control points through the spline basis, with boundary conditions folded in as a deterministic offset.","core_discovery":"The central claim is that B-splines, reformulated as movement primitives through their control points, inherit the probabilistic machinery of ProMPs while preserving exact boundary-condition satisfaction. Concretely, for given initial and final position and velocity conditions, the single-DoF trajectory distribution is Gaussian with mean $\\Phi^T_{[2:n-3]} \\mu_c + d$ and covariance $\\Phi^T_{[2:n-3]} \\Sigma \\Phi_{[2:n-3]} + I\\sigma_y^2$, where the selected control-point submatrix excludes the boundary-fixed endpoints and $d$ is the deterministic part fixed by the boundary conditions. This means the learned representation only needs to model the residual freedom of the interior control points, while the first two and last two control points are pinned down by equations such as $c_0 = y_0$ and $c_1 = \\frac{\\Delta}{p}\\dot{y}_0 + c_0$, with the symmetric equations at the end. The paper shows that B-splines alone regress digit-writing trajectories with lower mean squared error than ProDMP and ProMP, can represent constant segments without wiggles, and, when embedded in the BBRL and TCE algorithms, reach or exceed the success rates of ProDMP while also respecting prescribed end velocities.","pith_inferences":["The derivative recursion means BMP could impose acceleration or jerk boundary conditions as well, by fixing additional control points, which would matter for kinodynamic planning and physical feasibility.","The convex hull property of B-splines, which the authors mention as future work, could be used to certify velocity or torque bounds without sampling, giving hard safety guarantees in RL.","The linear-Gaussian assumption limits BMP to unimodal trajectory distributions; multi-modal demonstrations would need mixtures or latent variables, and the boundary-condition construction itself would survive but the probabilistic inference would not.","B-spline basis locality means changing one control point affects only a local segment of the trajectory, which may make BMP's optimization landscape smoother for black-box RL than the global-basis ProDMP."],"forward_implications":["BMP can be plugged into episodic reinforcement learning algorithms such as BBRL and TCE, and its trajectory distribution enables segment-level advantage-weighted policy updates that step-based algorithms cannot perform.","In the box-pushing task, BMP combined with TCE improves convergence speed and final success rate compared to BMP with BBRL, while enforcing zero end-velocity conditions that ProDMP fails to meet.","B-splines represent constant trajectory segments without oscillation, which ProDMP cannot do, so BMP is better suited for tasks with holding phases.","Imposing initial position and velocity and end position and/or velocity requires only fixing two control points at each end, and the same derivative recursion extends the construction to acceleration-level boundary conditions."],"supporting_citations":[{"why":"Supplies the ProMP linear-Gaussian trajectory model that BMP reformulates using B-spline basis functions and control-point weights.","marker":"[2]"},{"why":"Defines ProDMP, the closest probabilistic movement primitive baseline, which BMP compares against in regression, imitation learning, and reinforcement learning.","marker":"[4]"},{"why":"Provides the definition of B-spline basis functions, the Cox–de Boor recursion, and the derivative and convex hull properties that the formulation relies on.","marker":"[14]"},{"why":"Introduces the Temporal Correlated Exploration (TCE) algorithm, whose segment-level updates require a trajectory distribution and are applied to BMP in the paper.","marker":"[9]"},{"why":"Introduces the Deep Black-Box Reinforcement Learning (BBRL) algorithm used to demonstrate BMP in episodic reinforcement learning.","marker":"[10]"},{"why":"Supplies the synthetic-MNIST digit-writing dataset and the DMP training setup used in the imitation learning experiment.","marker":"[16]"},{"why":"Defines dynamic movement primitives, the earliest MP formulation that motivates the boundary-condition and attractor properties BMP seeks to combine.","marker":"[1]"}],"fun_headline_variants":["B-spline movement primitives blend exact endpoints with learned uncertainty","Probabilistic B-splines: exact boundary conditions meet trajectory learning","Robot learning gets B-spline movement primitives with exact end velocities","B-splines become movement primitives for probabilistic robot motion","Exact boundary constraints preserved in B-spline movement primitives"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the assumption that a Gaussian distribution over the learned control points adequately captures the trajectory variability in real imitation and reinforcement learning tasks; if the true variability is strongly non-Gaussian or multi-modal, BMP still enforces boundary conditions but its probabilistic predictions will be misleading.","fun_headline_variants_meta":{"raw":{"variants":["B-spline movement primitives blend exact endpoints with learned uncertainty","Probabilistic B-splines: exact boundary conditions meet trajectory learning","Robot learning gets B-spline movement primitives with exact end velocities","B-splines become movement primitives for probabilistic robot motion","Exact boundary constraints preserved in B-spline movement primitives"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000716,"raw_usage":{"total_tokens":3261,"prompt_tokens":1029,"completion_tokens":2232,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":2143}},"tokens_in":645,"tokens_out":2232,"duration_ms":15624,"temperature":1.0,"reasoning_tokens":2143,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:44:18.244902+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BMP on a demonstration set with two clearly separated movement modes, such as reaching above and below an obstacle, and evaluate the likelihood of the demonstrations under the fitted Gaussian trajectory distribution; with distinct modes the fitted Gaussian will place most probability mass between them, yielding low likelihood. A second check is a physical robot experiment that measures the achieved end velocity and compares it to the prescribed value, because the paper derives velocities from a phase derivative divided by duration $T$ and provides no hardware verification that exact end-velocity conditions are met.","supporting_citations":[{"cited_title":"Paraschos, C","cited_arxiv_id":null,"evidence_quote":"Supplies the ProMP linear-Gaussian trajectory model that BMP reformulates using B-spline basis functions and control-point weights."},{"cited_title":"Prautzsch, W","cited_arxiv_id":null,"evidence_quote":"Provides the definition of B-spline basis functions, the Cox–de Boor recursion, and the derivative and convex hull properties that the formulation relies on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Temporal Correlated Exploration (TCE) algorithm, whose segment-level updates require a trajectory distribution and are applied to BMP in the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Deep Black-Box Reinforcement Learning (BBRL) algorithm used to demonstrate BMP in episodic reinforcement learning."},{"cited_title":"Pahi ˇc, B","cited_arxiv_id":null,"evidence_quote":"Supplies the synthetic-MNIST digit-writing dataset and the DMP training setup used in the imitation learning experiment."}],"review_version":1}