{"id":"fa3580f1-7be0-4167-8894-bd210b4d3d78","arxiv_id":"2412.00597","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Spline-FRIDA models diverse human brushstroke trajectories with a variational autoencoder and a differentiable polyline renderer, and human raters preferred its drawings over the Bézier-based FRIDA baseline.","lead":"This paper adds a new type of brush stroke to the FRIDA robot painter: trajectories captured from human artists, encoded by an autoencoder, and rendered by a custom differentiable renderer. In a 100-person survey, its drawings were judged more human-like, more artistic, and closer to the reference than the older Bézier-stroke system.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section V-B's rotation/translation normalization erases orientation-dependent stroke styles; the paper acknowledges but never tests this assumption, so the general 'diverse human styles' claim is not yet established.","rationale":"The reader identified the Section V-B normalization as the weakest assumption, and I agree that it is the most load-bearing issue for the central style-capture claim. The paper explicitly acknowledges the tradeoff, which is good practice, but it does not test whether real stroke styles are invariant to position and rotation, nor does it restrict the abstract's 'diverse human stroke styles' claim accordingly. If the assumption fails, the VAE cannot encode a whole class of plausible human styles, and no amount of renderer improvement can recover the lost information. This concern does not invalidate the paper's more modest contributions, such as Traj2Stroke's lower Sim2Real gap for sharpie strokes or the demonstrated style matching for the five selected drawings, but it does mean the evidence does not yet support the general claim. Conditional acceptance remains appropriate: the paper should either add an ablation with orientation-dependent styles or explicitly scope the claims. The concrete test proposed—comparing orientation-conditioned stroke statistics between human and robot drawings—would settle whether the invariance assumption actually holds in relevant cases.","tokens_in":10200,"tokens_out":9295,"duration_ms":103092,"concrete_test":"Record a new human drawing with a clearly orientation-dependent style, for example hatching strokes that are long and straight when drawn horizontally but short and curved when drawn vertically. Fine-tune a TrajVAE on fewer than 20 examples using the Section V-B normalization, plan a target drawing, and compare the human and robot stroke distributions conditioned on absolute stroke angle (for instance, bin strokes by their pose offset and compare mean length and curvature per bin). If the robot's orientation-conditioned stroke statistics do not match the human's, the normalization assumption is violated and the style-capture claim must be scoped to rotation-invariant styles.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing assumption is in Section V-B: every captured trajectory is standardized by translating its start to (0,0) and rotating it so the endpoint lies on the x-axis, with the authors explicitly noting this 'assumes trajectory style is not affected by position/rotation on the canvas.' This assumption is never tested, and it is not a harmless preprocessing choice. It strips away absolute drawing angle and canvas location, so TrajVAE cannot learn styles in which stroke shape depends on orientation, such as hatching aligned to a form, contour shading, or slanted handwriting. During planning, each stroke has an independent pose offset, so the optimizer can orient a canonical shape, but the VAE latent prior is trained on the normalized marginal distribution of shapes and carries no information about how shape should vary with orientation. Thus, for orientation-dependent styles, the style transfer pipeline cannot reproduce the human style, regardless of how good Traj2Stroke's rendering is. The paper's style evaluations use only five hand-picked lab drawings, and the abstract claims general 'diverse human stroke styles,' so this untested invariance directly limits the central claim. The limitation is honestly stated, but no experiment bounds its scope or restricts the claim to rotation-invariant styles.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Spline-FRIDA, a robotic painting system that models human brush stroke trajectories using a variational autoencoder (TrajVAE) and a new differentiable renderer (Traj2Stroke). The authors collect motion-capture data of human drawings, fine-tune per-style TrajVAEs from fewer than 20 examples, and plan paintings via gradient descent through Traj2Stroke. Human evaluations on Mechanical Turk suggest that Spline-FRIDA captures stroke styles and produces more human-like, higher-quality, and more artistic drawings than the Bézier-based FRIDA baseline. The renderer is evaluated on out-of-distribution trajectories against CNN baselines, with Traj2Stroke achieving the lowest sharpie loss and Traj2Stroke with U-Net the lowest brush loss.","tokens_in":10489,"tokens_out":5194,"duration_ms":50357,"significance":"If the results hold, Spline-FRIDA offers a practical path to transferring a user's stroke style to a physical robot from a small number of demonstrations, and Traj2Stroke provides a sample-efficient differentiable renderer with a small Sim2Real gap. The paper's strengths include the Real2Sim2Real data collection, the modular VAE-decoder latent space, the out-of-distribution renderer experiments, and the use of physical robot executions in the human evaluation. The authors also honestly acknowledge computational limitations and the normalization tradeoff. However, the central diversity and human-likeness claims are broader than the current evidence, mainly because one load-bearing modeling assumption is untested and the survey results lack inferential statistics.","major_comments":[{"comment":"The trajectory normalization described in Section V-B translates every captured stroke so its start is at (0,0) and rotates it so its endpoint lies on the x-axis. The authors explicitly note this 'assumes trajectory style is not affected by position/rotation on the canvas,' but they do not test this assumption or limit the paper's claims to rotation-invariant styles. As a result, the TrajVAE cannot encode orientation-dependent stroke styles (e.g., hatching aligned with a form, slanted handwriting), so the claimed ability to model 'diverse human brush stroke trajectories' (abstract) and the style-transfer results in Section VI rest on an untested invariance. I would like to see either an experiment that checks VAE reconstruction on rotated/translated versions of held-out strokes and planning with different orientations, or a revised claim that explicitly restricts the method to styles that are position/rotation-invariant.","section":"V-B, VI"},{"comment":"The human evaluation in Section VI-A reports raw participant counts for five binary questions but provides no significance tests, confidence intervals, or inter-rater agreement measures. For example, the abstract's claim that Spline-FRIDA strokes are 'more human-like' is based on a 73/27 split in Table I; a binomial test would presumably support this, but no test is shown, and for the more balanced 'abstract' question (40/60) no conclusion can be drawn from the reported data alone. The claim that Spline-FRIDA 'improve[s] semantic planning' is supported only by the subjective 'better matches the reference image' question, not by a direct measure of planning objective loss. Please add appropriate statistical analyses and either directly measure planning quality (e.g., final loss or image similarity) or rephrase the claim.","section":"VI-A, Table I"},{"comment":"The conclusion that Traj2Stroke has a 'significantly smaller Sim2Real gap' (Section I contribution list) is based on average L1 losses in Table II, but the table reports no variance, error bars, or test of significance across repeated runs or trajectories. While the sharpie margin (0.00055 vs 0.00095) looks meaningful, the manuscript does not establish reliability, and on brush strokes Traj2Stroke with U-Net is the best method, which qualifies the headline result. Please report standard deviations or confidence intervals over the evaluation set and clarify that the choice of base Traj2Stroke for the full system is a design decision driven by generalization and simplicity rather than by absolute loss on the brush medium.","section":"VI-C, Table II"}],"minor_comments":[{"comment":"The training objective of TrajVAE is described as minimizing mean squared error; as a variational autoencoder it should also include the KL divergence term. Please state the full loss or clarify that only the reconstruction term is shown.","section":"V-C"},{"comment":"The five human drawings used in Fig. 6 are hand-picked lab members' drawings; while acceptable for a proof-of-concept, the paper would benefit from stating that the styles were selected and from reporting the number of trajectories and artists per style.","section":"VI"},{"comment":"There is a typo in 'inneficiencies' (should be 'inefficiencies') and a few other small grammatical errors that should be corrected in a final revision.","section":"VII"},{"comment":"If a trajectory has a zero-length segment (v = u), Eqs. (1)-(3) involve division by zero; please state the assumed handling for degenerate segments or add a small epsilon.","section":"V-D, Eq. (1)"},{"comment":"The qualitative claim that 'the vanilla CNN with and without CoordConv fails to generalize' is supported by Fig. 9, but the quantitative Table II does not break down the loss per style; reporting per-style losses would make the generalization claim more precise.","section":"VI-C"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid systems contribution with a clear evaluation structure, but the claims are broader than the evidence in three places: the untested orientation invariance, the lack of inferential statistics in the human survey, and the renderer comparison without variance. These are fixable within the scope of a revision. I recommend major revision rather than rejection because the core pipeline is plausible and the out-of-distribution renderer experiment is a good step. One mild concern for the editor: the paper's contribution list says 'significantly smaller Sim2Real gap' without a statistical test; the authors should be asked to justify or soften this wording."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the Spline-FRIDA paper. The headline: it's a solid, well-scoped robotics+SBR paper, and the main novelty—Traj2Stroke, a 7-parameter differentiable polyline renderer—genuinely works better than the CNN alternative for sharpie strokes, and it transfers to out-of-distribution trajectories. The trajectory VAE trained on mocap data is also a sensible way to capture stroke style, and the sample efficiency (fine-tune on <20 strokes) is a real selling point.\n\nThe human survey is convincing at a coarse level: 73/27, 84/16 etc. all favor Spline-FRIDA on human-likeness, quality, match, and artistic value. But there are no significance tests or confidence intervals, and with 100 MTurk participants it's cheap to compute. That's a minor fix, not a flaw in the method.\n\nThe soft spots are mostly about scope. The paper demonstrates five hand-picked styles from the lab, all of which are plausible rotation-invariant doodles. The stress-test about the normalization is correct: the preprocessing rotates every trajectory so the endpoint lies on the x-axis, and the authors note it assumes style is position/rotation invariant. For styles like calligraphy or hatching where the shape of a stroke depends on its absolute direction, the VAE cannot encode that, and the claimed 'diverse' styles are not demonstrated. This is an honest limitation in the text, but it should be scoped in the abstract and, ideally, tested—e.g., by comparing a VAE trained on normalized vs. orientation-preserving versions of the same strokes.\n\nAnother practical issue: the preprint claims 'open-source' but I didn't see a link to code or data. If that's in the camera-ready, fine; if not, the claim needs to be made good.\n\nWho'd get value from this: people working on stroke-based rendering, human-robot creative systems, and differentiable simulation. It's not a paradigm shift, but it's a useful combination with a real robot. I'd bring it to a reading group and I'd cite it if I were working on SBR. The central claims hold up to the extent demonstrated; the paper deserves a serious referee. Recommendation: send it out, but ask for significance tests, code/data release, and a scoping of the style claims.","headline":"A solid, scoped contribution to robot painting: the new differentiable polyline renderer is the real win, and the style-diversity claim needs explicit scope.","tokens_in":10970,"tokens_out":3663,"would_cite":true,"duration_ms":50730,"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 representing robot brush strokes as VAE-decoded trajectories with a new differentiable renderer lets a physical robot capture a human's stroke style from fewer than 20 examples, and that human raters judge the…","keywords":["robot painting","brush stroke modeling","variational autoencoder","differentiable rendering","stroke-based rendering","trajectory style transfer","motion capture","human-robot co-creation"],"falsifier":"Record the same artist drawing the same nominal stroke shape in several different canvas locations and orientations, normalize the trajectories exactly as the paper does, and inspect or train the TrajVAE: if strokes that are identical after normalization produce visibly different real marks, or if the VAE's reconstruction error rises systematically for strokes drawn at certain angles, the normalization premise fails. A more direct test is to define a style by spiral strokes centered at a particular corner of the canvas; since the normalization rotates every trajectory to a horizontal position, such a location-dependent style could not be encoded.","tokens_in":1843,"feed_emoji":"🎨","tokens_out":2436,"duration_ms":115206,"temperature":0.7,"pith_summary":"The paper argues that the limiting factor in robot painting is not the planning algorithm but the stroke primitive itself. It introduces Spline-FRIDA, which represents each brush stroke as a full trajectory sampled from a variational autoencoder trained on motion-captured human drawings, and renders that trajectory with a new differentiable stroke dynamics model, Traj2Stroke. The central claim is that this trajectory representation captures a human artist's intra-stroke style from fewer than 20 demonstration examples, and that paintings planned with it look more human-made, better match the reference image, and are judged more artistic than paintings made with the quadratic Bezier strokes used by FRIDA. A 100-participant survey supports this: participants matched robot drawings to the correct human style source, and preferred Spline-FRIDA over FRIDA on human-likeness, overall quality, reference match, and artistic value. The paper also reports that Traj2Stroke roughly halves the L1 prediction error on out-of-distribution Sharpie strokes compared to a CNN renderer, visibly narrowing the simulation-to-real gap.","feed_headline":"Human-style robot strokes beat Bezier curves in 100-person survey","feed_subtitle":"Learns a stroke style from under 20 examples; raters preferred its drawings as more human-like, accurate, and artistic.","key_machinery":"The two load-bearing components are TrajVAE and Traj2Stroke. TrajVAE is a variational autoencoder that compresses a 32-point polyline trajectory, with coordinates $(x,y,h)$, into a 64-dimensional latent code and decodes it back; pretrained on aggregated sessions and fine-tuned on a single drawing, it is the module that encodes stroke style. Traj2Stroke is the differentiable brush stroke dynamics model: it reorients the decoded trajectory to canvas coordinates, then for each segment computes a distance map, a height map by linear interpolation of brush elevation, an affine thickness map $\\alpha h + \\beta$, and a darkness profile $(1 - d/t)^c$, taking the max over segments. With only seven learnable parameters $m_x, m_y, b_x, b_y, \\alpha, \\beta, c$, the whole renderer is differentiable end-to-end, so gradient descent can optimize latent vectors $z$, pose offsets $\\Delta$, and colors $\\rho$ against a CLIP-based feature loss.","core_discovery":"The central discovery is that representing a stroke as a polyline trajectory in a learned latent space, rather than as a three-parameter Bezier curve, is what lets a robot adopt a human drawing style. The authors record artists with motion capture, normalize each stroke to a fixed pose, train a variational autoencoder (TrajVAE) on aggregated recordings, and fine-tune it on a single session using fewer than 20 trajectories. During planning the decoder generates stroke trajectories, and a new differentiable renderer, Traj2Stroke, turns each trajectory into a grayscale stroke image with variable thickness using only seven learnable parameters. The paper claims this pipeline preserves distinct human styles, such as long zig-zags, tiny curls, and small circles, and that survey participants reliably identified which human drawing sourced each robot style. In direct comparison, participants judged Spline-FRIDA drawings as more human-like than FRIDA's (73 percent versus 27), better overall (84 versus 16), closer to the reference image (84 versus 16), and more artistic (82 versus 18).","pith_inferences":["Because every trajectory is normalized before being encoded, the VAE cannot represent how stroke style depends on canvas position or drawing angle; testing whether artists actually vary stroke shape with pose would directly extend the method.","Traj2Stroke renders each segment independently and composites by max darkness, so it cannot model how overlapping wet strokes physically mix; a sequential renderer with ink transfer is a natural next step.","The survey compares the new trajectory representation against one Bezier-based system, so how much of the gain comes from the renderer versus the representation itself is untested and could be probed by giving the Bezier planner the same renderer.","The paper limits itself to intra-stroke style, leaving inter-stroke elements such as spacing, stroke order, and overlap unmodeled; a hierarchical stroke-pattern model could capture those without changing Traj2Stroke."],"forward_implications":["A user can hand a robot a drawing with a distinctive stroke style, and the robot can reproduce that style in new images after fine-tuning the style autoencoder on fewer than 20 recorded strokes.","Because Traj2Stroke is trained once per drawing medium and generalizes to out-of-distribution trajectories, new styles can be added by swapping in a differently fine-tuned TrajVAE without recollecting renderer data.","Planning remains fully differentiable: gradient descent can optimize latent stroke codes, canvas poses, and colors directly against a CLIP feature loss, so the method inherits FRIDA's text-, image-, and audio-prompt workflow.","The sim-to-real gap for Sharpie strokes is visibly reduced, meaning planned images can be trusted as previews of the physically painted result.","This style flexibility comes at a computational cost: a 400-stroke painting takes about an hour to plan on a 4090 GPU, with only 80 strokes optimized at a time due to memory limits."],"supporting_citations":[{"why":"Provides the FRIDA robotic painting platform, the Bezier stroke baseline, and the optimization-based planning loop that Spline-FRIDA extends.","marker":"[10]"},{"why":"Defines the Bezier-curve stroke primitive and differentiable-rendering approach that motivates the paper's more flexible trajectory representation.","marker":"[9]"},{"why":"The DiffVG differentiable rasterizer considered as an alternative; it lacks thickness differentiability and boundary softness, which motivates Traj2Stroke.","marker":"[17]"},{"why":"Supplies the variational autoencoder formulation that TrajVAE is built on for encoding and decoding stroke trajectories.","marker":"[23]"},{"why":"CLIP features serve as the perceptual representation used in the planning loss to compare generated paintings to target images.","marker":"[24]"},{"why":"Introduces the cosine-similarity loss over CLIP features that the planner backpropagates through Traj2Stroke.","marker":"[25]"},{"why":"CoordConv is used as a renderer-baseline variant tested against Traj2Stroke for coordinate-to-pixel mapping.","marker":"[27]"},{"why":"U-Net architecture forms the refinement network in the Traj2Stroke with U-Net variant, which performs best for brush strokes.","marker":"[28]"}],"fun_headline_variants":["Robot learns human brush style from under 20 strokes","Survey: humans prefer robot's stroke style over Bezier","Sample-efficient stroke model gives robot a human touch","Spline-FRIDA: humanlike robot painting from few examples","Robot paintings more humanlike with learned stroke paths"],"cache_read_input_tokens":13184,"weakest_assumption_plain":"The load-bearing assumption is that the style of a stroke does not depend on where it is drawn on the canvas or at what angle, because each captured trajectory is normalized by moving its start to $(0,0)$ and rotating its endpoint onto the x-axis before the VAE ever sees it.","fun_headline_variants_meta":{"raw":{"variants":["Robot learns human brush style from under 20 strokes","Survey: humans prefer robot's stroke style over Bezier","Sample-efficient stroke model gives robot a human touch","Spline-FRIDA: humanlike robot painting from few examples","Robot paintings more humanlike with learned stroke paths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1550,"prompt_tokens":938,"completion_tokens":612,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":534}},"tokens_in":554,"tokens_out":612,"duration_ms":118488,"temperature":1.0,"reasoning_tokens":534,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:11:12.943026+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the same artist drawing the same nominal stroke shape in several different canvas locations and orientations, normalize the trajectories exactly as the paper does, and inspect or train the TrajVAE: if strokes that are identical after normalization produce visibly different real marks, or if the VAE's reconstruction error rises systematically for strokes drawn at certain angles, the normalization premise fails. A more direct test is to define a style by spiral strokes centered at a particular corner of the canvas; since the normalization rotates every trajectory to a horizontal position, such a location-dependent style could not be encoded.","supporting_citations":[{"cited_title":"Frida: A col- laborative robot painter with a differentiable, real2sim2real planning environment,","cited_arxiv_id":null,"evidence_quote":"Provides the FRIDA robotic painting platform, the Bezier stroke baseline, and the optimization-based planning loop that Spline-FRIDA extends."},{"cited_title":"Learning to paint with model-based deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Defines the Bezier-curve stroke primitive and differentiable-rendering approach that motivates the paper's more flexible trajectory representation."},{"cited_title":"Differentiable vector graphics rasterization for editing and learning,","cited_arxiv_id":null,"evidence_quote":"The DiffVG differentiable rasterizer considered as an alternative; it lacks thickness differentiability and boundary softness, which motivates Traj2Stroke."},{"cited_title":"Auto-encoding variational bayes,","cited_arxiv_id":null,"evidence_quote":"Supplies the variational autoencoder formulation that TrajVAE is built on for encoding and decoding stroke trajectories."},{"cited_title":"An intriguing failing of convolutional neural networks and the coordconv solution,","cited_arxiv_id":null,"evidence_quote":"CoordConv is used as a renderer-baseline variant tested against Traj2Stroke for coordinate-to-pixel mapping."},{"cited_title":"U-net: Convo- lutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"U-Net architecture forms the refinement network in the Traj2Stroke with U-Net variant, which performs best for brush strokes."}],"review_version":1}