- 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
-
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).
-
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
-
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.
-
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
- Please report mean and standard deviation over at least 3 to 5 optimization seeds for Tables 1 to 3.
- On the biped (Table 2), Motion Sep-CMA-ES beats PhysibleMDM on penetration and slip. Why is noise optimization preferred there?
- 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
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.
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.