acceptodds
Under review as a conference paper at ICLR 2027

ReasonMotion: Consequence-Aware Motion Generation via Reward-Driven Counterfactual Optimization

Abstract

Counterfactual motion generation optimizes continuous interventions based on predicted downstream consequences rather than imitating reference trajectories. While essential for actionable sports coaching, existing diffusion models fail here because they require paired data of flawed and perfect executions. To address this, we propose ReasonMotion, a consequence-aware framework that formulates motion generation as reward-driven counterfactual optimization. Our two-stage pipeline couples a bounded stochastic remediation stage that refines an athlete's preparation without paired targets, with a prediction stage that generates improved continuations conditioned on this adjustment. To optimize this continuous multi-step process, we introduce visual Group Relative Policy Optimization (vis-GRPO) tailored for diffusion models to improve motion quality, alongside Sub-Trajectory Step Subsampling (STSS) to improve training efficiency while maintaining comparable motion quality. Empirical evaluations across multiple sports domains demonstrate that ReasonMotion improves both geometric fidelity and movement quality. Compared with the supervised backbone, it reduces average joint-position error from 152.53 to 142.67 mm and raises execution scores from 0.15 to 1.63 in figure skating, while increasing proximity-based hit rates from 60% to 80% and improving striking kinematics quality metrics in boxing. ReasonMotion enables a novel paradigm of consequence-aware AI coaching, bridging the gap between kinematic generation and actionable athletic guidance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
motion generation, reinforcement learning, diffusion models, sports coaching, credit assignment

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HZokOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ReasonMotion edits an initial motion segment and generates its continuation using reward-driven counterfactual optimization. It combines motion similarity, a learned execution-quality scorer and an acceleration penalty, with a diffusion optimization scheme designed to reduce training cost. Experiments cover motion prediction, figure skating and boxing.

Critical strengths
  1. Optimizing an initial motion through its predicted consequences is an interesting formulation, especially when paired examples of flawed and corrected motion are unavailable.
  2. The method offers a useful compute–quality trade-off. Held-out scoring networks and expert judgments also provide evidence beyond the training reward alone.
Critical weaknesses
  1. The performance advantage over strong matched baselines is not decisive. In Table 2, average pose error improves only from 143.59 mm with PPO to 142.67 mm, while PPO achieves higher execution scores and lower final-frame error, albeit at greater compute cost. The expert study does not include PPO or the supervised backbone, leaving the practical benefit over these alternatives insufficiently validated.
  2. Applicability beyond tasks with a clear execution-quality metric remains unclear. The skating and boxing experiments rely on domain-specific scorers and rewards; the boxing evidence concerns striking kinematics rather than broader interaction success. These results do not establish how counterfactual optimization would work for general motion generation, where a preferred outcome may be ambiguous. Higher predicted scores also do not establish that the proposed correction is physically achievable.
Questions for authors
  • How would the method define and validate preferences for motions without an established execution-quality score? Is the intended scope limited to tasks with domain-specific outcome scorers?
2 comments
AuthorsOpenReview ↗

Reply to HZok — Comment 1/1: W1, W2 and Q

We thank the reviewer for these valuable points. ReasonMotion prioritizes the compute–quality trade-off, outperforming PPO in both quality and accuracy at matched training compute. It also generalizes to score-free motion prediction by outperforming all baselines on Human3.6M, and adapts modularly across both aesthetic and purposive sports:

W1: Advantage over PPO. Our core claim is an improved compute–quality trade-off. At matched compute (~203 PFLOPs), ReasonMotion substantially outperforms PPO in both motion quality (Trans-R 1.631.63 vs. 1.361.36) and geometric accuracy (142.67142.67 vs. 145.58145.58 mm). While PPO eventually reaches higher terminal metrics (Trans-R 1.851.85 at 796 PFLOPs; Figure 3), it requires 3.9×3.9\times more training compute because its learned critic struggles with diffusion timesteps (Appendix G.3).

The expert study compared ReasonMotion with the reference motion and two published baselines; it did not include PPO or the supervised backbone, so it does not establish a perceptual advantage over them. Our claim relative to PPO is limited to the compute–quality trade-off above.

W2 and Q: Tasks without an execution score and generalization to boxing. When no domain-specific scoring rubric is defined, the framework remains functional through geometric and kinematic smoothness objectives alone. On the standard Human3.6M benchmark (Table 1), we disable the scoring reward entirely (λscore=0\lambda_{\text{score}} = 0) and still outperform all evaluated baselines (ADE 0.3250.325, FDE 0.4230.423).

For counterfactual coaching, the optimization naturally requires a target objective. The philosophy of sport distinguishes aesthetic (judged) sports, where the movement form itself is scored, from purposive sports, where an outcome decides success regardless of kinematic style (Best, 1978; Hurka, 2015). Figure skating is aesthetic, so consequence lies in the subsequent jump trajectory (modeled via remediation followed by prediction). Boxing is purposive: what matters is whether the strike contacts the target. The two stages of ReasonMotion are modular: boxing remediates the full window directly to optimize that extrinsic physical consequence (punch reach and 30 cm contact distance; Table 5).

Applying the full counterfactual pipeline to a new coaching domain thus requires specifying an appropriate outcome metric, which we clarify in the limitations. Furthermore, while boxing tracking failures under physics-based simulation are close to real motion (see detailed metrics in reply to SkEo, Q1), we do not claim verified physical achievability.

References

  • Best, D. (1978). Philosophy and human movement. Routledge.
  • Hurka, T. (2015). On judged sports. Journal of the Philosophy of Sport, 42(3), 317–325.
AuthorsOpenReview ↗

Follow-up on W2: Additional Experiment with Radar-Measured Pitch Velocity

We thank the reviewer again for the thoughtful feedback regarding tasks governed by objective physical outcomes (W2).

Alongside our score-free Human3.6M results (λscore=0\lambda_{\text{score}} = 0) and boxing experiments, we would like to share an additional baseball pitching experiment optimizing radar-measured pitch velocity on the public OpenBiomechanics Project dataset (full experimental details and biomechanical analyses are posted in our new General Response (Update) above). On held-out test pitchers, ReasonMotion increases estimated pitch velocity by +4.0 mph (85.4 to 89.4 mph) and model-free peak hand speed from 19.4 to 21.3 m/s while remaining within real-world biomechanical bounds, further showing that the framework extends naturally to tasks with measurable physical outcomes.

Reviewer_tE8QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ReasonMotion studies motion correction by editing an observed initiation and generating a continuation conditioned on that edit, without requiring paired flawed and corrected executions. The framework combines a text-conditioned diffusion backbone, partial re-denoising for remediation, and group-relative reward optimization using motion similarity, execution quality, and smoothness rewards. The remediator is trained using a frozen supervised consequence predictor; a separately reward-tuned predictor generates idealized continuations at inference. Sub-Trajectory Step Subsampling (STSS) retains the full denoising rollout while updating selected timesteps and sharing injected noise at inactive steps. Validation includes figure skating, whole-window boxing remediation, and Human3.6M prediction, with comparisons against supervised and reinforcement-learning baselines, stage ablations, multiple execution-quality judges, computational-cost measurements, and an evaluation by ten skating experts.

Critical strengths
  1. The separation between remediation and consequence prediction is useful and supported by meaningful stage comparisons. Training the remediator through a frozen supervised predictor provides a concrete way to evaluate proposed initiation changes using their predicted downstream consequences. The distinction between this predictor and the reward-tuned predictor used for inference is clearly specified in Sections 3.6–3.7. Table 3 provides useful controlled evidence: at t0 = 15, optimizing remediation while retaining the supervised predictor increases Trans-R from 0.146 to 0.866 and reduces initiation A-MPJPE from 91.27 to 89.02 mm. This supports the value of the remediation stage beyond simply improving the final predictor, while remaining appropriately understood as improvement under a learned consequence model.
  2. The evaluation provides several complementary sources of evidence for practical utility. The paper compares optimization methods using a shared backbone and reward formulation, and evaluates generated motions beyond the training scorer. In Table 2, adding STSS to vis-GRPO reduces cumulative training compute at the selected checkpoints from 343.4 to 202.1 PFLOPs while retaining similar geometric accuracy and improving both held-out GOE scores. The external MMFS-trained 2D judge in Appendix C.5 and the blinded expert evaluation provide additional support for the quality of the generated motions. Together with the boxing and Human3.6M experiments, these results make the framework a promising application of reward-guided motion generation.
Critical weaknesses
  1. The experiments do not sufficiently isolate the contribution of the shared-noise mechanism in STSS. Table 2 includes a useful comparison between vis-GRPO with and without STSS, as well as STSS variants of other optimizers. However, STSS changes two things simultaneously: the number of timesteps receiving policy updates and the coupling of injected noise across candidates. These comparisons establish the performance of the combined procedure, but do not identify the additional benefit of sharing inactive-step noise. The reduction in within-group reward spread reported in Appendix G.3 also does not by itself establish improved optimization. A control that updates the same K timesteps while retaining independent noise at inactive steps would directly test the proposed mechanism. This matters because the claimed synergy between noise sharing and group-relative advantages is presented as a principal methodological contribution.
  2. Some central interpretations of the validation exceed what the reported evidence establishes. Appendix H concludes that the held-out judges reject reward-hacked motions and thereby validate the deployed operating point. However, the configurations identified as reward hacking still receive higher mean Trans-R/Trans-D scores than the deployed configuration: 2.60/2.00 at scoring weight 0.65 and 2.35/2.01 at 0.90, compared with 1.63/1.29 at 0.25. Trans-R does decline from its peak and Trans-D approximately plateaus, so the sweep provides evidence of partial decoupling. Nevertheless, these configuration-level averages do not establish sample-level rejection or independently exclude exploitation at the selected setting.

The external 2D scorer and expert judgments are valuable supporting evidence, but the reported evaluations do not directly resolve this particular inference. Similarly, the qualitative examples in Appendix E do not independently establish the claimed physical recovery thresholds or a “theoretical upper bound” on successful recovery. The current evidence supports model-predicted corrections and expert-perceived usefulness more directly than physically verified reachability. This distinction is consequential to the paper’s interpretation of generated motions as actionable coaching guidance.

Questions for authors
  1. Additional experimental evidence: Can you compare full STSS with timestep subsampling alone, retaining independent noise at inactive steps and using the same K, rollout length, and evaluation protocol? A comparison of their quality–compute curves would help determine whether noise sharing contributes beyond the savings from updating fewer timesteps.
  2. Clarification of computational evidence: Appendix G.2 states that Figure 3 displays every evaluated checkpoint, but its displayed range extends only to approximately 600 PFLOPs, whereas the selected PPO and DDPO checkpoints in Table 2 require 796.1 and 889.8 PFLOPs. Can you reconcile this difference and provide the complete curves with the selected checkpoints marked? This would strengthen the evidence for the claimed efficiency comparison.
  3. Clarification of validation and claim scope: What operational criterion or sample-level evidence supports the statement that the held-out judges reject the exploited motions in Table 9? Do the external 2D or expert evaluations cover the high-weight configurations as well as the deployed setting? Please also clarify what independent evidence supports the physical recovery limits asserted in Appendix E. If these statements describe the behavior of the learned predictor, their scope should be restricted accordingly.
1 comment
AuthorsOpenReview ↗

Reply to tE8Q — Comment 1/1: Q1, Q2 and Q3

We thank the reviewer for precise and actionable questions. New experiments confirm that shared inactive-step noise is the key driver of STSS, extending Figure 3 to 900 PFLOPs verifies a 3.9$\times$ speedup over PPO, and an external 2D judge validates our deployed weight (λscore=0.25\lambda_{\text{score}} = 0.25) as the optimal quality–fidelity trade-off:

Q1 and W1: Shared-noise ablation. Under the identical 50-step rollout, K=5K = 5, and Table 2 protocol, we evaluate independent versus shared noise at inactive steps.

Compute (PFLOPs)3178109156202
Trans-R, shared (STSS)1.111.111.421.421.521.521.561.561.631.63
Trans-R, independent0.750.751.001.001.191.191.291.291.371.37
Trans-D, shared (STSS)0.760.761.081.081.201.201.221.221.291.29
Trans-D, independent0.490.490.690.690.870.870.980.981.071.07
A-MPJPE (mm), shared (STSS)145.46145.46143.05143.05143.18143.18142.60142.60142.67142.67
A-MPJPE (mm), independent152.28152.28148.86148.86147.21147.21146.73146.73146.42146.42

The 202 PFLOPs column reproduces our reported Table 2 checkpoint. This comparison demonstrates two key outcomes:

  1. Superior efficiency across all compute levels: Shared noise consistently outperforms independent noise across every evaluated checkpoint, surpassing the independent baseline's final Trans-R score (1.371.37) by 78 PFLOPs (1.421.42).
  2. Noise sharing preserves policy quality: Timestep subsampling with independent noise degrades generation quality, falling well below full-trajectory vis-GRPO without subsampling (Trans-R 1.371.37 and A-MPJPE 146.42146.42 mm at 202 PFLOPs, vs. 1.571.57 and 143.53143.53 mm at 343 PFLOPs in Table 2).

This confirms that subsampling alone degrades performance, and noise sharing is the active mechanism that sustains quality at reduced compute. We have added this ablation as Table 9 (Appendix G.3) to provide complete empirical validation alongside the internal reward-spread analysis.

Q2: Range of Figure 3. Thank you for catching this. In the earlier draft, points beyond ~600 PFLOPs were omitted to keep the low-compute region legible. The revised Figure 3 extends fully to 900 PFLOPs, plotting every measured checkpoint across all methods and explicitly marking each Table 2 operating point.

Each run was stopped once its validation reward plateaued, and Table 2 reports the first checkpoint reaching within 99% of the run's peak validation reward (the 1% margin is within one standard error). This compares each method at the compute required to reach its own performance plateau: vis-GRPO+STSS reaches its plateau at 202.1 PFLOPs, achieving competitive execution quality at 3.9×3.9\times less compute than PPO (796.1 PFLOPs).

Q3 and W2: Appendices H and E. The revised Appendix H identifies the deployed operating point via an explicit pre-defined criterion rather than retrospective inspection: (i) geometric fidelity must not degrade relative to the minimum-weight baseline, and (ii) in-loop scores must stay strictly within the annotated range [−5,+5][-5, +5]. As Table 10 shows, λscore=0.25\lambda_{\text{score}} = 0.25 is the maximal weight satisfying both (A-MPJPE 142.67 mm, 0% out-of-range scores). From 0.350.35 on, both fail sharply: A-MPJPE surges (179.08 to 324.48 mm) and predictions breach the label boundary (10.5% to 81.7%).

The held-out 3D judges already signal this transition: Trans-R peaks and declines (2.64→2.60→2.352.64 \to 2.60 \to 2.35) while Trans-D saturates (Table 10). For an orthogonal check, we evaluate the sweep using the external 2D judge (MMFS-trained, disjoint dataset and annotations):

λscore\lambda_{\text{score}}0.150.25 (deployed)0.350.650.90real motion
2D judge+0.80+0.80+0.82+0.82+0.78+0.78−0.15-0.15−0.47-0.47−0.26-0.26

The 2D judge corroborates λscore=0.25\lambda_{\text{score}} = 0.25 as the performance peak (+0.82+0.82) and sharply penalizes severe reward hacking at λscore≥0.65\lambda_{\text{score}} \geq 0.65 (dropping to −0.15-0.15 and −0.47-0.47).

We scope Appendix E to the learned predictor. E.1 and Figures 5–6 now describe predicted continuations under a fixed take-off: the model predicts successful landings for one to three rotations and a fall for the Quadruple prompt, which illustrates semantic controllability without forcing success for every prompt. On coverage, the external 2D judge is evaluated on the full sweep (Table 10), and the expert study on the deployed setting.

Reviewer_SkEoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ReasonMotion proposes a framework for modifying an observed movement according to its predicted downstream outcome, without requiring paired examples of flawed and corrected executions. A diffusion-based remediation model edits the initiation segment, while a continuation model generates the subsequent movement. The framework uses condition-matched group-relative optimization (vis-GRPO) and Sub-Trajectory Step Subsampling (STSS) to improve generated motion quality and reduce training computation. Experiments cover general motion prediction, figure skating, and boxing, with additional expert assessment of skating examples.

The problem is interesting and relevant: useful coaching should consider the consequences of a correction rather than simply matching an ideal reference. My principal concern is whether the evidence supports the interpretation of these outputs as athlete-specific, consequence-aware corrections, beyond higher-scoring kinematic examples.

Critical strengths

A useful formulation of motion correction. The paper addresses the difficulty of obtaining paired corrective demonstrations and proposes evaluating edits through their predicted continuations. Preserving the original initiation while permitting bounded changes is a sensible starting point for personalized motion editing.

A thoughtful separation of prediction and optimization. Keeping the supervised continuation model frozen during remediation training is an important design choice. It helps separate improvement through editing the initiation from improvement through reward-tuning the continuation generator. The corresponding ablation provides useful evidence that initiation edits affect predicted execution quality.

Promising computational and empirical results. The optimizer comparisons use a shared backbone and reward formulation. The combination of vis-GRPO and STSS offers a favorable quality–compute trade-off. Evaluation using multiple scoring architectures and blinded experts provides evidence beyond the training reward alone.

Critical weaknesses

Kinematic initial conditions do not uniquely determine biomechanical consequences. The method represents preparation through skeletal trajectories, but similar trajectories can arise from different muscle recruitment and co-contraction strategies. Motor-control research has long examined how multiple coordination patterns can satisfy the same task constraints [1]. Antagonist coactivation can alter mechanical impedance [2], and cocontraction plays a role in movement accuracy [3]. These findings motivate a distinction between visible motion and the internal state producing it.

Consequently, remaining close to the original joint trajectory does not establish that an edit preserves the athlete’s physical capabilities or is a small neuromuscular adjustment. The noise parameter controls the magnitude of a generative perturbation; its interpretation as a pedagogical intervention-strength parameter is not yet validated. Explicit muscle modeling is not necessarily required, but stronger evidence or narrower claims are needed.

Predicted improvement is not yet validated as an intervention effect. The frozen predictor is trained on observed preparation–continuation pairs. Its accuracy on natural movements does not establish its reliability on preparations modified by an optimizing model. The remediator could exploit prediction errors or scoring-model preferences while producing plausible-looking sequences.

Independent evaluation judges reduce dependence on a single scorer, but do not establish that the edited preparation would physically produce the generated outcome. The reward-tuned predictor also intentionally favors better outcomes. Its outputs should therefore be distinguished clearly from calibrated forecasts of what would follow an athlete’s intervention.

The evidence for ecological validity is limited. Expert ratings support perceived realism and usefulness, but the study does not establish whether athletes can implement the proposed corrections, improve their performance, or retain the improvement. This limits the claims of actionable coaching, rather than invalidating the motion-generation contribution.

The boxing results also require careful interpretation: the reported hit rate is a proximity measure with a 30 cm threshold, and the entire window is remediated rather than using the same initiation–consequence decomposition. This supports applicability to another motion-editing setting, but provides a less direct test of consequence-aware correction.

The incremental algorithmic contribution needs sharper positioning. Group-relative optimization and reward-guided diffusion already exist, including the visual-generation methods cited by the authors. The paper should make clearer which components of vis-GRPO are specifically new beyond applying condition-matched group comparisons to motion. For STSS, timestep subsampling and shared-noise coupling are separate changes whose contributions should be distinguished. The strongest demonstrated claim appears to be an improved quality–compute trade-off, rather than uniformly superior execution scores.

Questions for authors

Intervention validity: How is the frozen predictor validated on edited preparations, rather than only on naturally observed prefixes? Can the authors provide evidence that improvements persist under an independently trained consequence predictor, biomechanical checks, or other tests of intervention reliability?

Athlete-specific feasibility: What supports the claim that corrections remain feasible for the particular athlete? Can the authors distinguish preservation of visual identity from preservation of physical capability, and clarify what the intervention-strength parameter means for actual coaching?

Attribution of improvement: Could the main results more explicitly separate improvements obtained by editing preparation under the frozen predictor from those obtained by reward-tuning the continuation generator? Reporting the distribution of improvements, including failures and degraded cases, would help assess reliability.

Novelty and efficiency: What is the precise algorithmic difference between vis-GRPO and existing diffusion GRPO formulations? Can STSS’s timestep subsampling and shared-noise mechanism be ablated separately under matched compute?

Coaching evidence: Were experts assessing whether the movement looked better, whether a specific correction was understandable, or whether it was feasible for the original athlete? Without direct athlete testing, would the authors qualify the outputs as candidate kinematic corrections and idealized continuations?

The paper could strengthen its treatment of the boundary between edited preparation and generated consequence. Boundary-specific velocity, acceleration, and jerk measures would complement average smoothness metrics. M2D2M: Multi-Motion Generation from Text with Discrete Diffusion Models (Chi et al., ECCV 2024) is relevant background on coherent motion composition and transition evaluation.

InfoGCN++: Learning Representation by Predicting the Future for Online Human Skeleton-based Action Recognition provides complementary context for using future prediction to interpret partial movement. Discussing it could help distinguish predictive representation learning from intervention optimization. These are optional positioning suggestions, not requirements for acceptance.

Overall assessment

The work is promising as reward-guided motion editing, with a useful computational contribution and encouraging generated-motion evaluations. The central uncertainty is whether its improvements support the stronger claims of consequence-aware, personalized athletic coaching. Clarifying the algorithmic novelty and providing evidence for intervention reliability—or narrowing the claims accordingly—would materially strengthen the paper.

References supporting the biomechanical concern

[1] Scholz, J. P., and Schöner, G. (1999). The uncontrolled manifold concept: identifying control variables for a functional task. Experimental Brain Research, 126, 289–306. https://doi.org/10.1007/s002210050738

[2] Hogan, N. (1984). Adaptive control of mechanical impedance by coactivation of antagonist muscles. IEEE Transactions on Automatic Control, 29(8), 681–690. https://doi.org/10.1109/TAC.1984.1103644

[3] Gribble, P. L., Mullin, L. I., Cothros, N., and Mattar, A. (2003). Role of Cocontraction in Arm Movement Accuracy. Journal of Neurophysiology, 89(5), 2396–2405. https://doi.org/10.1152/jn.01020.2002

5 comments
AuthorsOpenReview ↗

Reply to SkEo — Comment 1/3: Q1 and W2

We thank the reviewer for a precise review. New evaluations across unseen predictors, physics simulation, and individual jumps confirm that our remediated initiations preserve execution gains across predictor architectures, stay within real-world kinematic and physics-tracking bounds, and improve over 40% of jumps (≤1.0\le 1.0 \\% degradation):

Q1 and W2: Do the improvements hold beyond the frozen predictor? We ran three checks.

(a) Independently trained predictors. We kept the remediated initiations (t0=15t_0 = 15, remediator trained against the frozen predictor). We then generated continuations with three predictors never used in remediator training, all trained on the FineFS training split: a second DDPM run with the same architecture, a flow-matching model, and a DDPM conditioned on the full jump name.

Below is the performance gain (Δ\Delta = Remediated −- SFT initiation) across 905 test clips, averaged over 3 generation seeds (95% bootstrap CIs exclude 0 for every entry except the full-name 2D judge):

PredictorΔ\Delta Trans-RΔ\Delta Trans-DΔ\Delta 2D judge
frozen (used in training)+0.72+0.72+0.72+0.72+0.35+0.35
second DDPM run+0.38+0.38+0.53+0.53+0.14+0.14
flow matching+0.36+0.36+0.45+0.45+0.14+0.14
full-name conditioning+0.12+0.12+0.19+0.19−0.01-0.01

Gains remain positive across nearly all settings and seeds, with only one marginal exception (full-name 2D judge at −0.01-0.01). Although the improvement magnitude narrows to 17%–74% of that on the training predictor under the two 3D judges (40% under the 2D judge for the second DDPM run and flow matching), edited initiations do not increase disagreement among predictors (116 mm vs. 117 mm for real prefixes), demonstrating algorithmic robustness across different outcome predictor architectures.

(b) Kinematic checks (skating). To assess physical and geometric plausibility (Sun et al., 2017; Shi et al., 2024), we establish empirical kinematic bounds from the real training distribution (0.1st–99.9th percentiles) and evaluate generated motions across all joints:

  • Range validity: At most 0.054% of generated frames exceed the empirical range in joint velocity or acceleration (vs. 0.13% in real test sequences), and at most 0.16% exceed valid knee or elbow angle ranges.
  • Junction continuity: At the boundary between the edited initiation and predicted continuation (frames 29 to 30), generated motions exhibit no seam spike: joint acceleration is 55% to 86%, and jerk is 50% to 83% of real motion at the same timestep.

Generated motions remain kinematically smooth, though the reward-tuned predictor unnaturally shortens leg bones (86% to 92% of real length) and reduces post-landing speed to exploit scoring models, noted as a limitation in the revision.

(c) Physics-based tracking (boxing). To evaluate physical plausibility in simulation, we test generated motions using PHC+, a physics-based humanoid controller, tracking 60 windows (5 samples each). Tracking failure rates are comparable across sources: 22% for real motion, 23% for the supervised model, and 28% for ReasonMotion. Crucially, failures concentrate heavily in the 13 challenging windows where real motion itself fails tracking, reflecting the inherent limits of the physics controller and kinematic tracking rather than generative artifacts. Sweeping t0t_0 from 5 to 25 yields stable failure rates (22%–28%) aligned with real motion benchmarks (see W4 for our task framing of boxing). For figure skating, full physics tracking is currently precluded because existing controllers cannot simulate gliding mechanics, causing even clean real jump recordings to fail.

AuthorsOpenReview ↗

Reply to SkEo — Comment 2/3: Q2, Q3 and W4

Q2: Feasibility for this athlete. Our objective is kinematic personalization rather than neuromuscular simulation. While joint-trajectory proximity does not model muscle recruitment or internal forces, preserving the athlete's observable movement characteristics is essential for actionable visual guidance.

The noise level t0t_0 operates as a geometric regularization bound: it explicitly confines the deviation from the athlete's recorded takeoff (with initiation A-MPJPE strictly bounded at 53.4, 72.4, and 89.0 mm for t0=5t_0 = 5, 1010, and 1515; Table 3). This ensures the refined preparation corrects technical flaws while remaining anchored to the athlete's unique execution style, preventing collapse onto a generic template (Appendix I). We will clarify this distinction in the revision to ensure t0t_0 is understood strictly as a geometric constraint rather than an internal biomechanical state.

Q3: Where the improvement comes from. Below is the t0=15t_0 = 15 part of Table 3, with the missing cell added in bold (905 clips, mean ±\pm std over 3 seeds).

RemediationPredictionTrans-RTrans-D2DInit. A-MPJPECons. A-MPJPEAcc.
SFTSFT0.146±.0500.146 \pm .050−0.081±.051-0.081 \pm .051−0.393±.009-0.393 \pm .00991.2791.27178.25178.250.05000.0500
vis-GRPOSFT0.866±.0380.866 \pm .0380.643±.0320.643 \pm .032−0.048±.011-0.048 \pm .01189.0289.02180.01180.010.04090.0409
SFTvis-GRPO1.923±.029\mathbf{1.923 \pm .029}1.513±.016\mathbf{1.513 \pm .016}0.199±.018\mathbf{0.199 \pm .018}91.27\mathbf{91.27}163.61\mathbf{163.61}0.0500\mathbf{0.0500}
vis-GRPOvis-GRPO2.302±.0292.302 \pm .0291.923±.0271.923 \pm .0270.457±.0110.457 \pm .01189.0289.02169.39169.390.04090.0409

Both stages improve execution quality, with the predictor yielding a larger score increase: optimizing remediation alone raises Trans-R from 0.1460.146 to 0.8660.866, whereas tuning prediction alone raises it from 0.1460.146 to 1.9231.923, reaching 2.3022.302 when both are applied. This matches how figure skating is scored in practice: the most severe GOE deductions target landing failures (a fall incurs −5-5; two-footed landing or stepping out incurs −3-3 to −4-4; International Skating Union [ISU], 2024), which physically manifest in the predicted continuation. Nonetheless, take-off kinematics establish the governing boundary conditions: a jump's parabolic trajectory and angular momentum are fixed when leaving the ice. Accordingly, effective remediation provides the foundation for successful execution, adding further gains on top of the tuned predictor (Trans-R increases from 1.9231.923 to 2.3022.302; 2D judge from 0.1990.199 to 0.4570.457) while improving initiation geometry (A-MPJPE drops from 91.2791.27 to 89.0289.02 mm). The remediation gain is consequence-aware by design, as initiation edits are guided strictly through their simulated downstream outcome.

Across 905 test clips under the frozen predictor (t0=15t_0 = 15), we evaluate whether remediation yields consistent sample-level improvements rather than outlier-driven gains. The distribution of score changes is detailed below (all mean gains have 95% bootstrap CIs excluding 0; "unchanged" denotes changes within the random generation noise floor):

JudgeMean gainImproveUnchangedDegradeFlawed jumps (GOE <0< 0)Successful jumps (GOE >0> 0)Unseen athletes
Trans-R+0.72+0.7241%59%0.6%+0.86+0.86+0.65+0.65+0.80+0.80
Trans-D+0.72+0.7246%53%0.6%+0.88+0.88+0.65+0.65+0.75+0.75
2D judge+0.35+0.3528%71%1.0%+0.44+0.44+0.30+0.30+0.39+0.39

Across all judges, at most 1.0% of clips degrade beyond generation noise. Crucially, flawed jumps gain the most (+0.86+0.86 vs. +0.65+0.65 for successful ones), and improvements transfer robustly to unseen athletes (+0.80+0.80 across 160 clips).

W4: Boxing as a consequence-aware task. The two stages of ReasonMotion are modular. Because boxing represents a purposive task where the outcome lies outside the body (hitting the target), we remediate the full window directly. Please refer to our detailed response in the reply to Reviewer HZok (W2) for the complete discussion on task modularity and aesthetic versus purposive formulations.

AuthorsOpenReview ↗

Reply to SkEo — Comment 3/3: Q4, Q5 and suggestions

Q4: Novelty and efficiency. vis-GRPO extends standard GRPO (Shao et al., 2024) to multi-step diffusion policies, arriving at a policy gradient objective that aligns with recent visual generation formulations (DanceGRPO; Flow-GRPO). The core algorithmic contribution lies in what the group shares: rather than grouping candidates by discrete text prompts, our groups share a continuous motion condition (the observed initiation, or the edited one). Normalising within the group naturally factors out the inherent execution difficulty of that specific attempt, resolving the credit assignment bottleneck without requiring PPO's complex value network or DDPO's prompt-keyed running statistics (Appendix G.3).

STSS shares conceptual motivation with prior efficiency-oriented schemes, but is uniquely tailored for stochastic DDPM samplers. While DanceGRPO subsamples a random subset of timesteps and MixGRPO restricts stochastic updates to a localized sliding window while reverting to deterministic ODE steps elsewhere, STSS updates non-contiguous steps while sharing inactive-step injected noise across candidates as common random numbers. This ensures each candidate remains a strictly valid draw from the generative policy while concentrating trajectory variance on the updated transitions. Reviewer tE8Q's ablation directly validates this mechanism: at identical compute, replacing shared noise with independent noise drops Trans-R from 1.631.63 to 1.371.37 and increases A-MPJPE from 142.67142.67 to 146.42146.42 mm (see reply to tE8Q). Our primary claim remains this favorable quality–compute trade-off, achieving competitive execution quality with 41% less cumulative compute (202 vs. 343 PFLOPs in Table 2).

Q5: Coaching evidence. The expert study measures whether a generated correction serves effectively as a visual demonstration, rather than testing whether athletes can immediately execute or retain the trajectory in physical practice (which we clarify in the revision). In Phase 3 (Appendix J.3), evaluators assessed whether the output could serve as a demonstration to help an athlete understand how to improve. The six competitive athletes in our cohort (national/international competitors, Table 14) answered affirmatively in 79.2% of trials (38 of 48), rating executability 3.84 out of 5 and reporting an improved understanding of the underlying technical flaw in 78.9% of cases (Table 16).

As suggested, we formally position the outputs as candidate kinematic corrections paired with idealized continuations that provide visual cues for movement pattern acquisition.

Suggestions. Thank you for both pointers. We evaluated velocity, acceleration, and jerk at the initiation–continuation boundary (frames 29–30), finding no seam spike (Q1(b)). In the revision, we will cite M2D2M for transition evaluation and discuss InfoGCN++ in Related Work.

References

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response

We thank the reviewers for their careful and constructive reviews.

ReasonMotion is designed as an actionable visual cue that helps an athlete imagine an improved version of their own movement, coupling a remediated preparation with an idealized continuation. Motor learning research supports observational demonstrations for pattern acquisition (Ashford et al., 2006; Boyer et al., 2009), and in our user study, 79.2% of responses from competitive athletes confirmed these videos helped them understand how to improve (Table 16; reply to SkEo, Q5). To substantiate this scope and evaluate empirical reliability across the reviewers' concerns, we report extensive new experiments summarized below.

Additional experiments addressing reviewer concerns

  1. Shared-noise ablation of STSS (tE8Q Q1, SkEo Q4): with the same KK , rollout length and protocol, shared noise consistently outperforms independent noise across compute levels (GOE 1.631.63 vs. 1.371.37 with lower MPJPE at 202 PFLOPs).
  2. Three independently trained outcome predictors (SkEo Q1): the remediation gain keeps its sign on both 3D judges for all three.
  3. Per-sample distribution of the remediation gain and the missing cell of Table 3 (SkEo Q3): 41% of samples improve and at most 1% get worse, failed jumps gain the most, and both stages contribute.
  4. Kinematic checks and physics-based tracking (SkEo Q1, Q2): generated motion stays within the range of real motion with no spike at the junction, and boxing motions are tracked by PHC+ at a rate close to real motion.
  5. Figure 3 over the full compute range, with the Table 2 checkpoints marked (tE8Q Q2).
  6. An external 2D judge on the reward-weight sweep (tE8Q Q3): it stays flat around the deployed weight and drops at the high weights where geometry degrades.

References

  • Ashford, D., Bennett, S. J., & Davids, K. (2006). Observational modeling effects for movement dynamics and movement outcome measures across differing task constraints: A meta-analysis. Journal of Motor Behavior, 38(3), 185–205. https://doi.org/10.3200/JMBR.38.3.185-205
  • Boyer, E., Miltenberger, R. G., Batsche, C., & Fogel, V. (2009). Video modeling by experts with video feedback to enhance gymnastics skills. Journal of Applied Behavior Analysis, 42(4), 855–860. https://doi.org/10.1901/jaba.2009.42-855
AuthorsOpenReview ↗

General Response (Update): Additional Experiment on Baseball Pitching

We share an additional experiment applying ReasonMotion to baseball pitching where the downstream consequence is radar-measured pitch velocity.

On held-out test pitchers, ReasonMotion increases independently estimated pitch velocity by +4.0 mph (85.4 to 89.4 mph) and model-free peak hand speed from 19.4 to 21.3 m/s while staying closer to the original trajectory than the supervised baseline. This shows that ReasonMotion applies directly whenever downstream consequences are physically measurable, without needing human judging rubrics.

Experimental Setup. We use the public OpenBiomechanics Project pitching dataset (Driveline Baseball Research and Development & Boddy), comprising 411 fastballs from 100 pitchers with motion capture and radar velocity, split by pitcher into 70/15/15 train/val/test sets (63 test pitches) with no pitcher overlap. As discussed in our reply to Reviewer HZok (W2) on aesthetic versus purposive sports, baseball pitching, like boxing, is a purposive task governed by an extrinsic physical outcome rather than judged form. Following our boxing setup, we therefore apply the remediation stage only using the dataset's 18 joints (60 Hz) and the same three-term reward (proximity to original motion, smoothness, and predicted pitch velocity).

Table A: Velocity and trajectory deviation on held-out test pitches (n=63n = 63).

MethodPitch velocity (mph)*Peak hand speed (m/s)Deviation from original (cm)
Original motion85.4 (radar)19.40.0
Supervised remediation only86.319.223.6
ReasonMotion89.421.321.8

*Original velocity is radar-measured; post-remediation velocity is estimated by an independent velocity evaluator trained separately from the in-loop reward model.

Supervised remediation alone leaves estimated velocity nearly unchanged (86.3 vs. 85.4 mph) and does not increase hand speed (19.2 vs. 19.4 m/s). ReasonMotion raises estimated velocity to 89.4 mph and model-free peak hand speed to 21.3 m/s while requiring a smaller trajectory deviation (21.8 cm vs. 23.6 cm). Thus, the velocity gain stems from RL guiding edits toward higher-velocity mechanics rather than larger perturbations (validation results follow the same trend).

Table B: Biomechanical kinematics on held-out test pitches (n=63n = 63).

MethodMax COM velocity toward home (m/s)Trunk forward tilt at release (deg)Peak pelvis rotation (deg/s)Lead knee extension vel. at release (deg/s)
Original motion3.0725.4789240
ReasonMotion3.3930.6690211

All biomechanical metrics follow literature and dataset definitions and were verified to match official dataset measurements when computed solely from joint positions. ReasonMotion increases forward center-of-mass (COM) velocity toward home plate and trunk forward tilt at release, both established kinematic correlates of higher fastball velocity (Matsuo et al., 2001; Stodden et al., 2001; Stodden et al., 2005; Ramsey et al., 2014). While peak pelvis rotation and lead knee extension velocity decrease slightly, both remain well within the typical range of the 411 real pitches, and arm slot angle at release is virtually unchanged (48.5° vs. 50.2°). The speed gain therefore arises from physically plausible linear-momentum adjustments rather than altering the pitcher's style or generating infeasible kinematics.

References

  • Driveline Baseball Research and Development, & Boddy, K. The OpenBiomechanics Project (OBP) [Dataset]. https://github.com/drivelineresearch/openbiomechanics
  • Matsuo, T., Escamilla, R. F., Fleisig, G. S., Barrentine, S. W., & Andrews, J. R. (2001). Comparison of kinematic and temporal parameters between different pitch velocity groups. Journal of Applied Biomechanics, 17(1), 1–13.
  • Ramsey, D. K., Crotin, R. L., & White, S. (2014). Effect of stride length on overarm throwing delivery: A linear momentum response. Human Movement Science, 38, 185–196.
  • Stodden, D. F., Fleisig, G. S., McLean, S. P., Lyman, S. L., & Andrews, J. R. (2001). Relationship of pelvis and upper torso kinematics to pitched baseball velocity. Journal of Applied Biomechanics, 17(2), 164–172.
  • Stodden, D. F., Fleisig, G. S., McLean, S. P., & Andrews, J. R. (2005). Relationship of biomechanical factors to baseball pitching velocity: within pitcher variation. Journal of Applied Biomechanics, 21(1), 44–56.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.