acceptodds
Under review as a conference paper at ICLR 2027

PhysibleMDM: Enforcing Inference-Time Physical Feasibility Constraints on Motion Diffusion Models via Noise Optimization

Abstract

Motion Diffusion Models (MDMs) can generate diverse and visually appealing human motion; yet, the resulting trajectories frequently violate physical constraints, limiting their downstream use in simulation and control. We present *PhysibleMDM*, a framework that promotes physical feasibility of MDM output by optimizing at *inference time* the diffusion noise *without* retraining the generative model. We introduce a set of surrogate feasibility constraints that can be evaluated without full trajectory simulation. These constraints combine kinematic checks (e.g., ground penetration, slip, and self-collision) with dynamic feasibility tests computed via an inverse-dynamics inner loop formulated as a cone program. We solve the resulting black-box noise optimization problem using the Separable Covariance Matrix Adaptation Evolution Strategy (Sep-CMA-ES), thereby avoiding reliance on computationally expensive and discontinuous gradients. An autoregressive scheme then chains optimized segments to synthesize long motion sequences. Across 100 motion styles on an SMPL humanoid, PhysibleMDM reduces mean ground penetration, self-collision penetration, and ground slip by 75%, 92%, and 74%, respectively, relative to unoptimized diffusion output. Across 11 sequences for a custom bipedal robot, it achieves median reductions of 89.5% in per-sequence mean acceleration-reconstruction error and 79.5% in the mean inverse-dynamics residual.

Primary area:
applications to robotics, autonomy, planning
Keywords:
diffusion models, motion generation, physical feasibility, inference-time optimization, noise optimization, derivative-free optimization, inverse dynamics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ncztOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tries to fix the physics issues in motions produced by a pre-trained motion diffusion model, without touching the model weights. The trick is to optimize the noise instead: all the noise the sampler uses (initial noise and the noise injected at each step) gets searched over with Sep-CMA-ES, a black-box evolution strategy, to bring down a physics penalty. The penalty is computed straight from the motion, no simulator required, and covers ground penetration, foot sliding, self collision, keeping the original foot contact pattern, and in the "dynamic" version an inverse dynamics error with torque and friction limits. Long motions are assembled chunk by chunk.

On 100STYLE motions retargeted to SMPL, penetration, self collision and sliding fall sharply (75%, 92% and 74%) relative to the plain model, and the method does better than the gradient and sampling baselines they tested, although the style of the motion drifts somewhat. The dynamic version and the biped robot experiments are much smaller, and there is a single RL tracking test.

Critical strengths
  1. The idea is simple and works with a frozen model. You do not retrain anything, you only search over the sampler's noise, and the physics checks run without a simulator. The gain on SMPL model is very large and are shown against a wider set of baselines ( including gradient-based ones).

  2. The inverse-dynamics residual is a useful proxy. Solving a small constrained inverse dynamics problem per frame and using the leftover error as a "can a body actually do this" score is the most substantive part of the paper, and it is something other motion generation work could reuse.

Critical weaknesses
  1. Every number is a single run of a random optimizer. Sep-CMA-ES is stochastic, and the paper gives no variance. The percentiles in the tables are over frames, not over runs. For a method whose whole contribution is an optimizer, this is the basic control that is missing, and several comparisons are close enough that I cannot tell what is real.

  2. The dynamic setting, which is the part that goes beyond prior kinematic noise optimization, is barely evaluated. Baselines are compared on one motion (Table 3), and the method takes 75 minutes per 30-frame chunk, so it is not clear it can be really used. The claim that the method "applies to any pretrained MDM" is also untested: only one architecture with 8 steps is used, and with a normal 50 to 1000 step text-to-motion model the noise vector would be tens of times larger.

Questions for authors
  1. Please report mean and standard deviation over at least 3 to 5 optimization seeds for Tables 1 to 3.
  2. On the biped (Table 2), Motion Sep-CMA-ES beats PhysibleMDM on penetration and slip. Why is noise optimization preferred there?
  3. Are the feasibility numbers confirmed by anything other than the proxies being optimized, for example a physics simulator rollout or more tracking runs? The one RL experiment shows root drift getting about 40% worse.
2 comments
AuthorsOpenReview ↗

Initial Response to Reviewer nczt: Seed Variability, Broader Evaluation, and Improved Downstream Tracking

We thank the reviewer for their constructive feedback. All appendix references below refer to the revised manuscript.

W1

We agree that the core results and comparisons should account for variability in generated motions across optimizer seeds. Appendices G.2.2 and H.3.1 report motion variability and mean ± sample standard deviation over ten seeds for kinematic optimization across 100 styles and dynamic optimization of the two-foot jump, respectively. Relative standard deviations are 0.85% for the kinematic objective, 1.55% for slip, and 11.2% for dynamic acceleration-reconstruction error. These results show low variability in the kinematic scores and moderate variability in dynamic acceleration-reconstruction error, with consistent motion smoothness.

Appendix G.1.1 reports a preliminary three-seed SMPL comparison of five methods across all 100 styles. Using the updated settings documented in the appendix, Physible has the lowest mean penetration, self-collision, slip, velocity-tracking error, and motion-consistency error in each seed. Mean penetration, self-collision, and slip decrease by 74%, 87%, and 76% relative to unoptimized generation. We plan to extend the three-seed comparison to all baselines and ablations in Tables 1–3 for the final paper.

W2

Broader dynamic feasibility optimization evaluation. Appendix H.1.1 compares Physible with unoptimized generation across 100 styles and 15,000 frames per method. Compared with Table 3, this provides 100× the style coverage and 60× the evaluated frames, increasing motion duration from 5 seconds to 5 minutes. Mean acceleration-reconstruction error decreases by 98.2%, with improvements in every style.

Computational cost. Dynamic feasibility optimization remains expensive and is currently suited to offline refinement. Inverse dynamics and dynamic-constraint evaluation account for approximately 97% of runtime (Appendix F.4.1). Moreau accelerates warm inverse-dynamics solves by 3.84× on CPU and 2.89× on GPU relative to the eight-worker Clarabel interface used in our CAMDM experiments (Appendix F.4.2). Future work could amortize this cost by learning noise corrections for the frozen MDM, following Noise Hypernetworks (Eyring et al., 2025) and LeNO (Kim et al., 2026). Noise Hypernetworks learns to modulate initial noise, while LeNO learns a residual noise correction from scalar rewards and applies it in a single step at inference. These approaches offer a route to amortizing optimization through MDM post-training.

Another pretrained model. Appendix G.1.2 reports preliminary results with text-conditioned Kimodo on Unitree G1, using 100 denoising steps across 100 prompts. Relative to unoptimized motions, Physible reduces mean ground penetration by 14.9%, slip by 20.0%, and self-collision penetration by 52.1%, demonstrating effectiveness on another MDM and embodiment. Kimodo uses deterministic DDIM, so we optimize only initial noise. The search dimension is independent of the number of denoising steps, although generation cost grows. We will add preliminary dynamic-feasibility results for Kimodo on G1 during the discussion period.

Q1

As detailed in W1, Appendix G.1.1 reports mean ± sample standard deviation over three seeds for five methods from Table 1. We plan to extend this evaluation to all baselines and ablations in Tables 1–3.

Q2

Motion Sep-CMA-ES achieves lower mean penetration and slip on the biped, but reduces smoothness. Table 2 reports log dimensionless jerk (LDJ) of −9.52, compared with −7.75 for Physible and −7.52 for unoptimized motion. Higher LDJ indicates smoother motion. Direct trajectory optimization can introduce jitter, whereas noise optimization retains the frozen motion generator as a prior throughout refinement.

Q3

We expanded physics-simulation tracking to ten clips and three training seeds per policy (Appendices F.1.4 and H.5.1). The clips comprise nine walking styles and the two-foot jump. We train one policy on the optimized clips and another on their unoptimized counterparts. We gave positional tracking terms 10× the weight of orientation terms to account for their different scales and added reference state initialization. Both conditions use identical training settings.

Both policies track every clip without failure. Relative to unoptimized references, optimized references reduce root position, root orientation, body position, body orientation, and joint position errors by 6.3%, 2.4%, 5.1%, 12.9%, and 13.7%, respectively. Root position and orientation errors change from 39.6% and 41.3% higher than the baseline in the submitted single-clip experiment to 6.3% and 2.4% lower in the revised ten-clip evaluation. This improves the relative root-tracking benefit by 45.9 and 43.7 percentage points.

AuthorsOpenReview ↗

Discussion reply to nczt: Dynamical feasibility optimization on Kimodo G1 and downstream SONIC evaluation

We have further updated the revised manuscript with two additional sets of results on dynamical feasibility optimization for Kimodo G1 and downstream SONIC evaluation.

Dynamic optimization on another model (W2). Appendix H.2.2 and Table 18 evaluate dynamic feasibility optimization with Kimodo on Unitree G1 across five text prompts. Kimodo uses 100 denoising steps, and we optimize its initial noise. Both acceleration-reconstruction and inverse-dynamics force residuals decrease for every prompt. The median reductions across prompts in their mean values are 47.5% and 37.4%, respectively. We plan to evaluate more Kimodo prompts for the final paper.

Additional simulation evidence (Q3). Appendix H.2.4 and Table 20 compare inverse-dynamics torque and contact-force predictions with simulated execution using the same frozen SONIC policy for unoptimized and optimized references. Across the five prompts, average clip-level joint-torque RMSE decreases by 5.3% and net ground-force RMSE by 14.1%. Both errors improve for four prompts and increase for left-foot hopping, providing preliminary evidence beyond the optimized residuals.

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

This paper proposes PhysibleMDM, which improves the physical feasibility of motion diffusion outputs by optimizing diffusion noise at inference time. It uses Sep-CMA-ES to reduce kinematic violations and inverse-dynamics residuals without retraining the model. Experiments on SMPL motions and a bipedal robot show clear improvements in penetration, slipping, and dynamic consistency. However, the optimization is expensive, and the results do not guarantee physically feasible motion.

Critical strengths
  1. The method improves motion feasibility without retraining the diffusion model. This is a practical advantage over approaches that require additional training.

  2. The authors test both kinematic and dynamic feasibility and compare against several baselines. The results show consistent improvements across multiple metrics, rather than only on a single measure.

Critical weaknesses
  1. The optimization is very slow. In the bipedal jump experiment, the reported time per motion horizon is about 4,515 seconds. This makes it difficult to see how the method could be used in time-sensitive applications.

  2. The evaluation is limited. The robot experiments cover only a small set of motions and rely on simplified contact assumptions. It is also unclear how stable the results are across different optimization seeds. More evidence is needed to show that the method works reliably in other settings.

Questions for authors
  1. What hardware was used for the experiments, and how much of the total optimization time is spent on diffusion sampling versus inverse-dynamics solves?

  2. How many generated sequences satisfy the physical constraints within a defined tolerance before and after optimization? Average reductions alone do not show how often the method actually produces feasible motions.

  3. How much do the results vary across optimization seeds? Have the authors measured this variation quantitatively?

3 comments
AuthorsOpenReview ↗

Initial Response to Reviewer xzRh: Computational Cost, Broader Evaluation, and Consistency Across Seeds

We thank the reviewer for their constructive feedback. All appendix references below refer to the revised manuscript.

W1

We agree that dynamic feasibility optimization is currently expensive and suited to offline motion refinement. The new runtime breakdown attributes approximately 97% of runtime to inverse dynamics and dynamic-constraint evaluation (Appendix F.4.1). Our solver benchmark shows that Moreau accelerates warm inverse-dynamics solves by 3.84× on CPU and 2.89× on GPU relative to the eight-worker Clarabel interface used in our CAMDM experiments (Appendix F.4.2).

Future work could amortize this cost by learning noise corrections for the frozen MDM, following Noise Hypernetworks (Eyring et al., 2025) and LeNO (Kim et al., 2026). Noise Hypernetworks learns to modulate initial noise, while LeNO learns a residual correction from scalar rewards and applies it once at inference. This could move the optimization cost into MDM post-training, removing the need to run iterative optimization in time-sensitive applications.

W2

Broader evaluation. Appendix H.1.1 compares Physible with unoptimized generation across all 100 motion styles and 15,000 frames per method. Relative to Table 3, this provides 100× the style coverage and 60× the evaluated frames, increasing motion duration from 5 seconds to 5 minutes. Mean acceleration-reconstruction error decreases by 98.2%, with improvements in every style. Appendix H.1.1 also evaluates 11 sequences spanning styles, slopes, and friction settings, with a median reduction of 89.5% in per-sequence mean acceleration-reconstruction error. The sticking-contact assumption and extensions to sliding contacts are discussed in Appendix I.

Variability across seeds. Appendices G.2.2 and H.3.1 report motion variability and mean ± sample standard deviation over ten seeds for kinematic optimization across 100 styles and dynamic optimization of the two-foot jump, respectively. Relative standard deviations are 0.85% for the kinematic objective, 1.55% for slip, and 11.2% for dynamic acceleration-reconstruction error. Average pairwise root-trajectory RMSE is 3.51 cm and 7.57 cm in the kinematic and dynamic settings, respectively.

Appendix G.1.1 adds a preliminary three-seed SMPL comparison of five methods across all 100 styles. Physible has the lowest mean penetration, self-collision, slip, velocity-tracking error, and motion-consistency error in each seed. Mean penetration, self-collision, and slip decrease by 74%, 87%, and 76% relative to unoptimized generation. We plan to extend this comparison to all baselines and ablations in Tables 1–3 for the final paper.

Q1

The CAMDM experiments use an Intel i9-12900K CPU, NVIDIA RTX 4090 GPU, and 64 GB RAM (Appendix F.4). In the new profiling results (Appendix F.4.1), diffusion generation accounts for 43.6% of kinematic optimization runtime, which requires no inverse-dynamics solves. For dynamic optimization, generation accounts for 1.5%, while inverse dynamics and dynamic-constraint evaluation account for 96.8%.

Appendix F.1 describes the custom biped with 20 actuated joints. We also added preliminary kinematic results with text-conditioned Kimodo on Unitree G1 across 100 prompts and 100 denoising steps (Appendix G.1.2). Relative to unoptimized motions, mean ground penetration, slip, and self-collision penetration decrease by 14.9%, 20.0%, and 52.1%, respectively. We will add preliminary dynamic-feasibility results for Kimodo on G1 during the discussion period.

Q2

We added empirical cumulative distribution plots for kinematic feasibility on both embodiments (Appendix G.1.1), Kimodo on G1 (Appendix G.1.2), and dynamic feasibility across 100 styles (Appendix H.1.1). These show the fraction of evaluated frames whose violation is below any chosen tolerance for each metric.

For example, kinematic optimization increases the fraction of frames with zero self-collision penetration from 90.7% to 94.9% on SMPL and from 96.4% to 98.4% on the biped. Dynamic optimization reduces the 99th-percentile acceleration-reconstruction error from 19.0 to 0.256, demonstrating improvement beyond the mean.

Q3

Please see W2 and Appendices G.2.2, H.3.1, and G.1.1 for the ten-seed variability measurements and the preliminary three-seed comparison with baselines. For each seed, we first average each metric over the evaluated frames and styles. We then report the mean and standard deviation of these averages across seeds.

AuthorsOpenReview ↗

Discussion reply to xzRh: Dynamical feasibility optimization on Kimodo G1

We have further updated the revised manuscript with two additional sets of results. The new Kimodo G1 dynamical feasibility results address your concern in W2 about limited evaluation coverage.

Broader dynamic evaluation (W2). To address the concern about limited evaluation across motions and settings, Appendix H.2.2 and Table 18 now evaluate dynamic feasibility optimization with text-conditioned Kimodo on Unitree G1 across five prompts. Both acceleration-reconstruction and inverse-dynamics force residuals decrease for every prompt. The median reductions across prompts in their mean values are 47.5% and 37.4%, respectively. This extends the dynamic evaluation to another motion generator and robot embodiment. We plan to include more Kimodo prompts in the final paper.

Thank you for the detailed response. The rebuttal and additional experiments have addressed my concerns, so I will keep my positive recommendation.

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

PhysibleMDM improves the physical feasibility of frozen motion-diffusion-model (MDM) output at inference time, without retraining or a tracking controller. It optimizes the diffusion noise (jointly initial and posterior) with the gradient-free Sep-CMA-ES algorithm, evaluating objectives on fully denoised motions. Feasibility is enforced through kinematic constraints (ground penetration, slip, self-collision, contact pattern) and dynamic constraints computed by a constrained inverse-dynamics cone program that checks whether accelerations are realizable under underactuated floating-base dynamics, torque bounds, and friction cones. Noise standardization preserves style, and an autoregressive scheme chains segments. On SMPL (100 styles) it cuts ground penetration, self-collision, and slip by 75%, 92%, and 74%; on a bipedal robot (11 sequences) it cuts median acceleration-reconstruction error and inverse-dynamics residual by 89.5% and 79.5%.

Critical strengths
  1. Well-motivated and genuinely training-free (no retraining, no simulator, no tracking controller).
  2. Gradient-free Sep-CMA-ES is a sound fit for non-smooth contact constraints and high-dimensional noise (78,840-103,680 variables).
  3. Constrained inverse dynamics (cone program) is a careful dynamic-feasibility surrogate that respects underactuated floating-base dynamics.
  4. Noise standardization is simple yet essential (without it style accuracy collapses to 57.5%).
  5. Thorough evaluation: ten baselines/ablations, honest limitations.
Critical weaknesses
  1. Downstream RL validation is mixed: body/joint tracking improves (16.4%/11.9%/7.4%) but root position/orientation tracking worsens by 39.6%/41.3%.
  2. No comparison to PhysDiff (simulation-based projection), the most direct "physics at inference" competitor; yet the paper already uses a simulator and a tracker in the RL experiment (Sec. 8.7), so the no-simulator boundary is inconsistent, and the surrogate ID residual is never validated against actual simulated feasibility.
  3. "Applicable to any pretrained MDM" is too broad. The method requires decoding to full-body link poses for forward kinematics, a robot-specific dynamics/contact model, and a sampler whose posterior noise can be injected deterministically — a narrower class than "any MDM."
Questions for authors
  1. Why does dynamic-feasibility optimization worsen global root tracking by ~40%? Does re-weighting the root objective or relaxing root-DOF feasibility help?
  2. Since the dynamic setting optimizes the residual it reports, what independent evidence shows actual dynamic feasibility improved beyond the mixed RL result?
  3. Since you already use a simulator and a tracker in Sec. 8.7, why not run PhysDiff on the same motions? Does a reduction in your ID residual correspond to improved simulated feasibility, or merely to a reduction in the residual you optimized?
7 comments
AuthorsOpenReview ↗

Initial Response to Reviewer Wkyo: Improved Tracking, Comparison with PhysDiff, and Model Applicability

We thank the reviewer for their constructive feedback. All references below refer to the revised manuscript.

W1

We expanded downstream physics-simulation tracking to ten clips and three training seeds per policy (Appendices F.1.4 and H.5.1). The clips comprise nine walking styles and the two-foot jump. We train one policy on optimized clips and another on their unoptimized counterparts. We gave positional tracking terms 10× the weight of orientation terms to account for their different scales and added reference state initialization. Both conditions use identical training settings.

Both policies track every clip without failure. Relative to unoptimized references, optimized references reduce root position, root orientation, body position, body orientation, and joint position errors by 6.3%, 2.4%, 5.1%, 12.9%, and 13.7%, respectively. Root position and orientation errors change from 39.6% and 41.3% higher than the baseline in the submitted single-clip experiment to 6.3% and 2.4% lower in the revised ten-clip evaluation. This improves the relative root-tracking benefit by 45.9 and 43.7 percentage points.

W2

Comparison with PhysDiff. PhysDiff requires a pretrained general-purpose motion tracker capable of following the varied intermediate motions produced during denoising. This tracker projects those motions through simulator rollouts at selected diffusion steps. Our method evaluates physical feasibility through constrained inverse dynamics, without requiring a tracking policy during motion optimization. Appendix A discusses this distinction.

Our downstream trackers are trained after motion optimization and specialize in the resulting reference clips. Using them for simulation-based evaluation is consistent with refining motions before a controller is available. A PhysDiff comparison would require training a separate general-purpose tracker for the intermediate denoising motions.

Validation in simulation. The revised RL evaluation provides downstream evidence beyond the optimized inverse-dynamics residual. Under identical training and evaluation settings, optimized references improve all five tracking measures in physics simulation (W1 and Appendix H.5.1). We use closed-loop tracking because bipedal balance requires feedback to correct deviations from the reference. Replaying feedforward inverse-dynamics torques alone would not provide that stabilizing feedback.

W3

We agree that the applicability claim should state the required interfaces. Section 1 now specifies pretrained MDMs with decodable robot configuration sequences and controllable sampling noise. Section 3 clarifies that the generated features must allow recovery of root position, root orientation, and joint coordinates over the motion chunk. The MDM need not output full-body link poses directly, as these follow from forward kinematics. Section 6.1 and Appendix C.1 explain that the robot-specific kinematics, dynamics, and contact properties come from a separately supplied robot model, as in our Newton implementation for G1.

Regarding control over posterior noise, our method requires access to the sampler’s noise inputs so that each candidate motion can be reproduced for fixed conditioning. Section 1 now includes controllable sampling noise among the applicability requirements.

We also demonstrate this applicability with text-conditioned Kimodo on Unitree G1 (Appendices B.2 and G.1.2). Across 100 text prompts, Physible reduces mean ground penetration by 14.9%, slip by 20.0%, and self-collision penetration by 52.1% relative to unoptimized motions. These results extend the evaluation to another pretrained MDM and robot embodiment. We will add preliminary dynamic-feasibility results for Kimodo on G1 during the discussion period.

Q1

The revised RL training addresses root tracking through reweighted rewards and reference state initialization (Appendix F.1.4). As detailed in W1, optimized references now reduce root position and orientation errors by 6.3% and 2.4%, respectively.

Q2

The revised simulation experiment shows lower tracking errors for optimized references across all five measures, with three training seeds per policy and ten clips (Appendix H.5.1). This evaluates the optimized motions through closed-loop execution beyond the residual used during optimization. We will seek additional independent validation and report it during the rebuttal period.

Q3

Please see W2 for the distinction between the tracker required during PhysDiff refinement and our trackers trained afterward. The revised downstream results also show improved simulated tracking. We will follow up with additional independent evidence as it becomes available during the rebuttal period.

I thank the authors for their detailed responses to my questions. I have gained a fairly comprehensive understanding of this Paper. First, PhysibleMDM is an optimization method that enforces physical constraints through diffusion-noise optimization. Its computational cost, particularly in the dynamic setting, appears substantial, making it more suitable for offline or non-real-time applications. Second, the effectiveness of such methods depends heavily on the design of physical constraints and optimization objectives, while their generalization to complex scenarios, such as human–scene interactions and various unseen terrains, remains largely unvalidated. Finally, since PhysibleMDM optimizes diffusion noise rather than motion directly, it relies heavily on the motion prior learned by the pretrained MDM to preserve the style and structure of generated motions. Finally, I’m curious — what digital human and embodied AI tasks do the authors see PhyssibleMDM being useful for?

AuthorsOpenReview ↗

Reply to Reviewer Wkyo: Applications of PhysibleMDM in Digital Animation and Embodied AI

We thank the reviewer for this thoughtful assessment. We see the clearest near-term applications in digital-human animation and offline preparation of motion references for embodied-agent training.

For digital humans, PhysibleMDM could refine generated animation clips for games, film, and virtual environments, reducing artifacts that undermine perceived physical realism. Our experiments demonstrate reductions in ground penetration, foot sliding, and self-collisions, which could support future integration into animation-production workflows.

For embodied agents, we demonstrate that refining reference motions before training a tracking policy improves downstream tracking in physics simulation. This workflow accommodates the optimization cost offline, while the trained controller executes without repeating the noise optimization.

Beyond these demonstrated results, we see promising extensions to manipulation, underactuated systems, and nonholonomic systems. In bimanual manipulation, adapting a familiar carrying motion to a heavier payload requires coordinated hand trajectories, collision avoidance, and forces consistent with grasp and actuator limits. Underactuated systems offer further applications, including quadrotors with coupled translation–attitude dynamics and limited thrust, and robots handling suspended or deformable objects whose passive swing or deformation must be considered. For nonholonomic systems, such as wheeled mobile robots, generated trajectories could be refined to respect rolling and steering constraints. These applications would combine the learned behavior with explicit models of the relevant robot, object, and interaction dynamics.

More broadly, the same sampling-based noise optimization could incorporate task completion, safety, stability, robustness, and efficiency objectives when suitable evaluators are available. Its gradient-free formulation is particularly useful for hybrid or discontinuous dynamics, such as contact-rich interactions, where gradients may be unavailable or poorly informative. We expect this approach to be most useful when the pretrained motion prior already captures relevant behavior for the intended task and environment, and refinement can adapt that behavior to the specific physical and task requirements.

Comments on the paper

2 comments

Dear Reviewers,

A friendly reminder to check the authors’ rebuttals and engage in the discussion, as the deadline is approaching. Authors still have time to address any further questions or concerns you may have, so your timely feedback would be greatly appreciated.

Thank you for your time and effort!

Your AC

AuthorsOpenReview ↗

Rebuttal summary for the Reviewers and Area Chair

We thank all reviewers for their thorough study of our paper and their constructive feedback, and the Area Chair for overseeing the discussion. The reviewers' comments have helped us substantially improve the manuscript, particularly the breadth and rigor of the evaluation. As the discussion period draws to a close, we summarize the main additions below. All references refer to the revised manuscript.

Broader dynamic-feasibility evaluation. We expanded the CAMDM evaluation to 100 motion styles and 15,000 frames per method, providing 100× the style coverage and 60× the evaluated frames of the submitted single-motion comparison. Physible reduces mean acceleration-reconstruction error by 98.2% relative to unoptimized generation, with improvements in every style. We also evaluate 11 sequences spanning selected styles, slopes, and friction settings, and report error distributions alongside averages (Appendix H.1.1).

Evaluation across random seeds. We added a preliminary three-seed comparison of five methods across all 100 SMPL styles. Physible achieves the lowest mean ground penetration, self-collision, slip, velocity-tracking error, and motion-consistency error among these methods in each seed; mean penetration, self-collision, and slip decrease by 74%, 87%, and 76% relative to unoptimized generation (Appendix G.1.1). Additional ten-seed studies quantify variability in kinematic optimization across 100 styles and dynamic optimization of the two-foot jump (Appendices G.2.2 and H.3.1).

Another pretrained model and robot embodiment. We added text-conditioned Kimodo on Unitree G1, using 100 denoising steps. Across 100 prompts, kinematic optimization reduces mean ground penetration, slip, and self-collision penetration by 14.9%, 20.0%, and 52.1%, respectively (Appendix G.1.2). A preliminary dynamic evaluation across five prompts reduces both acceleration-reconstruction and inverse-dynamics force residuals for every prompt, with median reductions in their per-prompt means of 47.5% and 37.4% (Appendix H.2.2).

Comparison with PhysDiff. We added a preliminary Kimodo G1 comparison across five prompts using 2, 4, and 8 physics projections (Appendices H.2.2–H.2.3). Physible achieves slightly lower average acceleration-reconstruction error at all three budgets and lower average inverse-dynamics residuals at four and eight projections. It also preserves the source motion more closely, with 78–81% lower average joint deviation. PhysDiff achieves lower average SONIC tracking errors, ground slip, and self-collision penetration. SONIC also supplies PhysDiff's internal projections, whereas Physible requires no pretrained tracking policy during refinement.

Downstream evidence in physics simulation. We expanded the CAMDM tracking evaluation to ten clips and three training seeds per policy. Under identical training settings, optimized references improve all five tracking measures, including reductions of 6.3% in root-position error and 13.7% in joint-position error (Appendix H.5.1). Separately, executing the five Kimodo G1 motions with the same frozen SONIC policy reduces average clip-level joint-torque and net ground-force RMSE relative to inverse-dynamics predictions by 5.3% and 14.1%. Both errors decrease for four prompts and increase for left-foot hopping. This provides evidence beyond the optimized residuals (Appendix H.2.4).

Computational cost and applicability. New profiling identifies inverse dynamics and dynamic-constraint evaluation as approximately 97% of dynamic optimization runtime. We evaluated the recently released Moreau solver, which achieves 3.84× faster warm inverse-dynamics solves on CPU and 2.89× on GPU relative to the eight-worker Clarabel interface used in the CAMDM experiments (Appendix F.4). Dynamic refinement remains primarily suited to offline use. We have also clarified the required MDM interfaces, robot models, and application-specific constraints.

Some additions remain preliminary because of the computational resources and time available during this short rebuttal period, particularly the five-prompt Kimodo dynamic and PhysDiff studies and the three-seed baseline comparison. For the camera-ready version, we plan to broaden the evaluations to more motion styles and text prompts and extend the seed comparison to the remaining baselines and ablations.

We would welcome any remaining questions or concerns from the reviewers and Area Chair and would be happy to discuss them during the final hours of the rebuttal period. If the revised manuscript and additional evidence address your concerns, we would be grateful if you would consider updating your assessments and ratings accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.