{"id":"8932cebd-44cc-4e10-b500-e528a75b2d73","arxiv_id":"2505.07802","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Flow Planner combines a UNet with local inpainting conditioning, action-noise data augmentation, and train/inference trajectory splitting to enable flow models to stitch novel trajectories for robotic manipulation.","lead":"This paper presents Flow Planner, a trajectory-planning method that lets generative flow models compose pieces of previously seen robot motions into new plans. The method avoids obstacles up to four times larger than prior diffusion planners, shown on a Franka Panda arm in simulation and on real hardware.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Obstacle avoidance headline lacks statistical grounding; no trial counts or error bars for the claimed 'up to four times as large' improvement.","rationale":"The reader identified the local receptive field assumption as the weakest, but that is a concern about the mechanistic explanation, not about the empirical central claim. Even if the receptive field is not perfectly local, the stitching and obstacle avoidance results could still be valid for other reasons. The more load-bearing condition is that the headline quantitative claim—avoiding obstacles 'up to four times as large'—is reliably measured. This condition is insecure because the paper provides no trial counts, error bars, or baseline success-rate curves in the obstacle avoidance experiment. Without these, the claimed factor cannot be assessed or reproduced. The stitching metric is also ambiguous: if the boundary states are inpainted as exact conditioning, the reported MSE should be near zero, so the metric as described may not measure what it claims. These empirical gaps directly undermine the central claim, whereas the locality assumption is secondary. The reader's conclusion of conditional acceptance remains appropriate, so the verdict is unchanged.","tokens_in":79,"tokens_out":3012,"duration_ms":38850,"concrete_test":"Re-run the obstacle avoidance experiments with a fixed number of trials (e.g., 50) per obstacle size for each method, report mean and 95% confidence interval of success rate, and define the maximum obstacle as the largest radius with lower CI bound at or above 0.7. Also clarify the inpainting protocol in §5.1 by stating whether boundary states are hard-conditioned or soft-conditioned, and recompute the stitching error accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Flow Planner 'avoids obstacles up to four times as large' as baselines rests on the obstacle avoidance experiment in §5.2 and Fig. 5. Yet the paper reports only a single 'maximum obstacle radius that each method was able to reliably avoid' with no trial counts, no error bars, and no definition of 'reliably' in the main text. Appendix B later defines success as reaching the goal without collision, but Fig. 7 still omits the number of rollouts per obstacle size and does not include baseline methods. Without knowing the sample size, the reported maxima could be driven by a single lucky plan or by outlier noise. Moreover, the VAE baseline is described as a 'local planner' and MPD as unable to stitch, but no quantitative comparison of success rates across obstacle sizes is given for these baselines, so the claimed factor of four cannot be independently checked. The stitching benchmark itself is also under-specified: the text says 'the true first and last are inpainted' and then measures MSE against those same true states, which is trivially zero if conditioning is exact; the protocol for measuring 'stitching error' therefore needs clarification before the out-of-distribution boundary condition claim can be evaluated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Flow Planner, a flow-matching trajectory model for robot manipulation that aims to improve trajectory 'stitching'—composing novel complete plans from sub-trajectories—through three components: a UNet with inpainting conditioning to enforce local temporal receptive fields, Gaussian action-noise augmentation of the training dataset, and a trajectory-splitting procedure applied at both training and inference. The authors evaluate stitching on a Franka Panda in simulation using a benchmark where boundary-condition pairs are held out, and evaluate guided planning in an obstacle-avoidance task in simulation and on real hardware. They report that Flow Planner stitches out-of-distribution boundary conditions better than MPD, transformer, and VAE baselines and can avoid obstacles up to four times larger than the baselines. Code and videos are provided.","tokens_in":10118,"tokens_out":5107,"duration_ms":50861,"significance":"If the results hold, the paper provides a practical recipe for making generative planners compose rather than select trajectories, with concrete architectural and dataset-augmentation choices that go beyond the usual whole-trajectory conditioning used in diffusion planners. The benchmark idea (train on opposite-side condition pairs, test on same-side pairs) is a clean way to isolate stitching ability, and the paper includes several meaningful baselines as well as a real-hardware deployment, which is a strength. However, the quantitative evidence for the headline claims is currently under-powered and some definitions are ambiguous: the stitching-error metric as written is difficult to interpret, the 'local receptive field' mechanism is asserted rather than demonstrated for the UNet used, and the obstacle-avoidance results lack statistical grounding. With a tightened evaluation and clearer reporting, this could be a useful contribution to the generative planning literature.","major_comments":[{"comment":"The definition of the stitching error is ambiguous and the current wording undermines the central quantitative claim. The paper states: 'We then record the mean squared error between the first and last planned states as the stitching error – since the true first and last are inpainted – and the true initial and final states.' If the endpoints are inpainted with the true states, then the endpoint error is trivially zero; if the error is instead computed at an interior point or at a stitching seam, this must be stated explicitly and the tables re-labeled accordingly. Please specify exactly which states are compared and what the conditioning inputs are, and clarify whether 'batches of 64' are independent rollouts whose standard deviations are reported.","section":"Section 5.1, Tables 1-2"},{"comment":"The paper's central explanation for the stitching improvement is that the UNet with inpainting has a 'completely local receptive field,' but this appears inconsistent with the architecture as described. The UNet in Table 3 has downsampling layers with channel dims [32, 64, 128, 256], so the effective receptive field grows with network depth, and inpainting conditions on the global start and goal states at full resolution. The claimed locality is therefore not a property of the architecture as presented. Please either provide a direct measurement of the effective receptive field (for example, via input-gradient or ablation analysis) or revise the claim to 'approximately local' and discuss the role of the bottleneck, since this mechanism is load-bearing for the paper's narrative.","section":"Section 4.1 and Table 3"},{"comment":"The headline claim that the method 'avoid[s] obstacles up to four times as large' is not supported by the reported evidence. Fig. 5 reports only a single 'maximum obstacle radius that each method was able to reliably avoid' with no trial counts, no error bars, and no definition of 'reliably' in the main text. Appendix B defines success as reaching the goal without collision and shows success-rate curves for FP and FP+split, but it omits the MPD and VAE baselines, so the factor-of-four comparison against baselines cannot be independently checked. Please report success-rate curves with confidence intervals or error bars for all methods and state the number of rollouts per obstacle size.","section":"Section 5.2, Fig. 5, Appendix B"},{"comment":"The standard deviation of the Gaussian action noise is never reported. Since action-noise augmentation is claimed to be uniquely effective and the Limitations section explicitly describes a trade-off between trajectory smoothness and stitching robustness as a function of noise level, this is a key free parameter for reproducibility. Please state the value used in the experiments (in Table 3 or elsewhere) and include a sensitivity analysis showing how stitching performance and trajectory smoothness vary with the noise level.","section":"Section 4.2 and Section 7 (Limitations)"},{"comment":"The claim that train-time trajectory splitting 'completely eliminates' mode collapse is supported only by the qualitative plots in Fig. 4a. The paper does not provide a quantitative measure of mode collapse, nor does it define the training stage or guidance scale at which the phenomenon was assessed. Please add a quantitative metric (for example, the consistency error or the variance of generated trajectories across seeds) and describe the exact conditions under which the collapse was observed and eliminated.","section":"Section 4.3 and Fig. 4"}],"minor_comments":[{"comment":"There are several copyediting issues: 'plan viastitching' should be 'plan via stitching', and 'massively improvement flexibility' should be 'massively improved flexibility'.","section":"Section 1"},{"comment":"The caption contains an awkward comma: 'a local receptive, field' should be 'a local receptive field'.","section":"Figure 1 caption"},{"comment":"The guidance weighting bt = 1-t is described as found to work better 'in practice'; please state the range of values tried and whether this was selected on a validation set, since it is an ad hoc modification of the theoretically justified weighting.","section":"Section 3.2"},{"comment":"The hardware experiment is described with a single example and no quantitative success metric or number of trials. Please state how many hardware trials were performed and report success rates or at least a clear statement that this is a qualitative demonstration.","section":"Section 5.2"},{"comment":"The sentence 'The numbers in the right figure are the maximum object radius that the arm was able to reliably avoid' should refer to 'the right panel of Fig. 5' and should state the units explicitly in the caption.","section":"Section 5.2, Fig. 5"},{"comment":"There are two typos: 'A interesting direction' should be 'An interesting direction', and 'less destructive manor' should be 'less destructive manner'.","section":"Section 7"},{"comment":"Several references are incomplete or have line-broken URLs (for example, [1], [2], [3], [9]); please ensure all entries are formatted consistently with venue and publisher information.","section":"References"},{"comment":"The word 'significantly' is used in the abstract and Section 5 without statistical tests. Please either add significance tests (e.g., confidence intervals or paired tests across seeds) or replace 'significantly' with a descriptive term such as 'substantially'.","section":"Abstract and Section 5"}],"recommendation":"major_revision","confidential_remarks":"The core recipe is interesting and could be publishable in a robotics/AI journal after a substantial revision that strengthens the quantitative evaluation. The main gap is between the strong abstract-level claims and the thinness of the evidence in Section 5.2. The self-citations to the group's prior work ([24] and [40]) appear natural and appropriate; I do not see a citation-pattern concern. The paper's fit with the journal is reasonable, but the evaluation standards need to be raised before I would support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on generative planning for manipulation. The paper's real contribution is a clean empirical demonstration that replacing MPD's transformer-plus-direct-conditioning setup with a UNet inpainted on start/goal states, adding action noise to the dataset, and mixing half-length trajectories into training lets a flow-matching planner compose new trajectories from seen sub-trajectories. The stitching benchmark is well designed—train on a cross-shaped dataset with only opposite-side condition pairs, test on same-side pairs—and the ablations in Tables 1 and 2 are informative: each component matters, and the architecture and augmentation choices are properly compared. A model trained this way gets stitching error 0.16 versus 1.00 for MPD. That is a real, reproducible result, with code released. The limitations section is honest about the noise/smoothness trade-off and the hardware filtering it forced.\n\nThe local receptive field idea is a useful mechanistic frame, though 'completely local' is an overstatement: the UNet downsamples, so the effective receptive field grows at the bottleneck. That doesn't kill the empirical result, but the explanation should be 'more local' rather than 'local-only.' The paper partly owns this in the limitations with the failed transformer masking experiments.\n\nWhere the paper is soft: the 'obstacles up to four times as large' headline rests on a single 'maximum radius reliably avoided' per method, with no trial counts, no error bars, and 'reliably' only defined in the appendix. Appendix B gives success-rate curves only for the two FP variants, no baselines, and no number of rollouts per radius; the factor of four cannot be checked as reported. The stitching-error protocol is also ambiguous: if the true first and last states are hard-inpainted, the MSE measured against them would be zero; since the reported errors are non-zero, a step is missing from the description. That is a clarification, not a fatal flaw. The action-noise standard deviation, the main new hyperparameter, is not reported. Mode-collapse elimination is shown qualitatively, and the hardware result is a single demo.\n\nNone of these are load-bearing for the stitching result; the obstacle-avoidance claim is the one being oversold. Audience: people building generative planners for manipulation, and anyone in diffusion/flow policy work who cares about conditioning and out-of-distribution generalization. It deserves a serious referee: send it to review, and ask for statistics on the obstacle avoidance, a precise stitching-metric definition, and the noise level.","headline":"A genuinely useful empirical paper on making diffusion/flow planners stitch sub-trajectories, with a well-supported stitching result and an under-evidenced obstacle-avoidance headline.","tokens_in":10616,"tokens_out":4985,"would_cite":true,"duration_ms":45873,"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 argues that generative trajectory planners fail on novel start-goal pairs because they select, not stitch, and presents Flow Planner, which uses a time-local UNet, action-noise augmentation, and trajectory splitting to compose…","keywords":["flow matching","trajectory stitching","robot manipulation","guided planning","obstacle avoidance","diffusion models for planning","local receptive field","action noise augmentation"],"falsifier":"Move the inpainted goal far away while holding the local neighborhood of a given time step fixed, and measure how much that state's predicted value changes: if the output shifts substantially, the receptive field is not local, and the locality explanation is falsified; equivalently, inspect the trained UNet's effective receptive field with gradient-based sensitivity analysis.","tokens_in":9682,"feed_emoji":"🤖","tokens_out":6488,"duration_ms":61233,"temperature":0.7,"pith_summary":"This paper argues that generative trajectory planners fail on novel start-goal pairs because they select whole trajectories from the training set instead of stitching sub-trajectories. It proposes Flow Planner, built from three ingredients: a UNet with time-local (inpainting) conditioning, action-noise augmentation of the training data, and trajectory splitting during training and inference. On a Franka Panda arm, the method generates feasible plans for out-of-distribution boundary conditions and avoids spherical obstacles up to four times larger than the Motion Planning Diffusion baseline and other comparators, in simulation and on real hardware. The claim matters because it suggests the combinatorial space of motion clips needed for flexible behavior shrinks exponentially once a model can compose, rather than retrieve.","feed_headline":"A flow planner that stitches can avoid obstacles 4x larger","feed_subtitle":"Local temporal conditioning and trajectory splitting let a Franka arm compose new plans from existing motion clips.","key_machinery":"The load-bearing mechanism is the local receptive field: each time step in the trajectory is predicted only from its temporal neighborhood, implemented by using a UNet with inpainting conditioning, as opposed to a transformer or direct conditioning that lets global start or goal information bias every state. Around this sit two auxiliary mechanisms. Action-noise augmentation adds Gaussian noise independently to each joint's torques during dataset collection, breaking state-to-state correlations so the model can switch between nearby overlapping clips. Trajectory splitting mixes half-length trajectories into training batches with 50 percent probability to prevent mode collapse, and at inference re-noises an initial plan up to 50 percent, splits it in half, and denoises each half separately with the midpoint as a new boundary condition to reduce dynamic inconsistency; trajectory lengths must remain powers of two because of the UNet's structure.","core_discovery":"The central discovery is that stitching failure in generative trajectory planners is not inherent to flow matching but is driven by three identifiable deficiencies, each with a fix. A model whose conditioning is non-local in time, such as a transformer or direct conditioning, learns to reproduce whole dataset trajectories, so novel boundary conditions produce inconsistent plans; a UNet conditioned by inpainting, with a local receptive field, instead maximizes local consistency and can compose new trajectories. Dataset augmentation with action-specific Gaussian noise breaks correlations between joint states and enables the model to jump between overlapping clips, whereas randomizing start or goal positions and injecting random forces do not. Finally, mixing half-length trajectories into training batches prevents a guidance-induced mode collapse, and an inference-time split, re-noising to 50 percent and denoising each half separately with the midpoint as a new boundary condition, slows the growth of dynamic inconsistency under strong guidance. Together these techniques yield Flow Planner, which outperforms diffusion and VAE baselines in stitching error and obstacle-avoidance radius, and the result transfers to a real Franka Panda arm.","pith_inferences":["If the local-receptive-field explanation is right, then transformer-based planners should be stitchable by designing attention masks that are truly local; the paper's unsuccessful tridiagonal mask suggests this is a concrete open problem rather than a settled impossibility.","The paper's correlation-breaking hypothesis predicts a testable extension: augmentations that decorrelate successive joint states in a dimension-wise way, if they are less destructive than full torque noise, should preserve the smoothing-safety trade-off while enabling stitching.","Inference-time splitting is a generic guidance-stabilization trick, so it could be transplanted to other generator-matching planners, including diffusion planners, wherever strong guidance causes endpoint inconsistency.","Because the paper identifies architectural locality rather than flow matching itself as the cause, related stitching results in offline reinforcement learning and model-based planning might be re-examined through the same lens."],"forward_implications":["Planners built this way no longer need dense coverage of every start-goal pair; the paper states the combinatorial space of required clips reduces exponentially once stitching works.","Larger guidance scales become usable with inference-time splitting, so obstacle-avoidance and other cost-guided tasks can demand bigger deviations from the unguided plan without the trajectory becoming dynamically inconsistent.","The same architecture and conditioning choices transfer from simulation to real hardware, since the obstacle-avoidance task succeeds on a real Franka Panda with a box inside the maximum sphere.","Stitching performance is quantifiable: on a cross-shaped Franka dataset, Flow Planner reports a stitching error of 0.16 ± 0.07, versus 1.00 ± 0.97 for the Motion Planning Diffusion baseline.","Action-noise augmentation is the dataset-side key: same-noise, random-start/goal, and random-force schemes all leave stitching error higher than the 0.16 of Flow Planner."],"supporting_citations":[{"why":"Supplies the cross-shaped 2D particle environment used to illustrate stitching failure and the consistency/cost trade-off that guided planning is interpreted as.","marker":"[4]"},{"why":"The Motion Planning Diffusion baseline; contributes the signed-distance guidance scheme and obstacle-avoidance setup that Flow Planner is compared against and outperforms.","marker":"[16]"},{"why":"Defines the flow-matching framework, including probability paths, velocity fields, and the flow ODE, that Flow Planner builds upon.","marker":"[33]"},{"why":"Establishes the conditional flow-matching objective used to train the velocity network.","marker":"[34]"},{"why":"Provides the transformation between velocity fields and score functions that underlies classifier guidance in the paper.","marker":"[38]"},{"why":"Represents the common generator-matching architecture and conditioning schemes, transformers and UNets with direct or inpainting conditioning, that the paper categorizes by receptive field.","marker":"[6]"},{"why":"Explores composing shorter clips into longer ones, the closest related use of splitting, which the paper contrasts with using splitting to aid guidance.","marker":"[29]"}],"fun_headline_variants":["Flow Planner stitches to dodge 4x larger obstacles","Stitching flow models yields 4x obstacle avoidance","Flow model stitching avoids obstacles 4x larger","Local conditioning enables flow stitching for 4x dodging","Flow stitching dodges 4x bigger obstacles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the claim that the UNet's receptive field is completely local in time, so inpainting conditioning never leaks the global start or goal into a state's prediction; if downsampling gives the UNet a non-local receptive field, the stated mechanism behind the stitching improvement is not established.","fun_headline_variants_meta":{"raw":{"variants":["Flow Planner stitches to dodge 4x larger obstacles","Stitching flow models yields 4x obstacle avoidance","Flow model stitching avoids obstacles 4x larger","Local conditioning enables flow stitching for 4x dodging","Flow stitching dodges 4x bigger obstacles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00101,"raw_usage":{"total_tokens":4240,"prompt_tokens":888,"completion_tokens":3352,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":3276}},"tokens_in":504,"tokens_out":3352,"duration_ms":23269,"temperature":1.0,"reasoning_tokens":3276,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:07:59.254751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Move the inpainted goal far away while holding the local neighborhood of a given time step fixed, and measure how much that state's predicted value changes: if the output shifts substantially, the receptive field is not local, and the locality explanation is falsified; equivalently, inspect the trained UNet's effective receptive field with gradient-based sensitivity analysis.","supporting_citations":[{"cited_title":"Janner, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the cross-shaped 2D particle environment used to illustrate stitching failure and the consistency/cost trade-off that guided planning is interpreted as."},{"cited_title":"Zheng, M","cited_arxiv_id":null,"evidence_quote":"Provides the transformation between velocity fields and score functions that underlies classifier guidance in the paper."}],"review_version":1}