arxiv: 2512.03028 · v3 · submitted 2025-12-02 · 💻 cs.GR · cs.AI· cs.CV· cs.RO

Recognition: no theorem link

SMP: Reusable Score-Matching Motion Priors for Physics-Based Character Control

Yuxuan Mu , Ziyu Zhang , Yi Shi , Dun Yang , Minami Matsumoto , Kotaro Imamura , Guy Tevet , Chuan Guo

show 4 more authors

Michael Taylor Chang Shu Pengcheng Xi Xue Bin Peng

Authors on Pith no claims yet

Pith reviewed 2026-05-17 02:02 UTC · model grok-4.3

classification 💻 cs.GR cs.AIcs.CVcs.RO

keywords motion priorsscore distillation samplingphysics-based character controldiffusion modelsimitation learningreward functionshumanoid animation

0 comments

The pith

Score-Matching Motion Priors let pre-trained diffusion models serve as frozen reward functions for any downstream character control task.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that motion priors can be learned once from a large reference dataset using score distillation sampling on a diffusion model, then kept fixed while training separate policies for new tasks. This removes the need to retrain the prior or retain the original motion clips every time a new controller is introduced. A sympathetic reader would care because it turns motion data into a modular, reusable library that can still produce high-quality naturalistic movements in simulated humanoid characters. The same base prior can also be repurposed for specific styles or combined to create movements absent from the training set.

Core claim

Score-Matching Motion Priors are formed by applying score distillation sampling to a pre-trained motion diffusion model. The resulting signal acts as a general-purpose reward that encourages any policy to produce motions drawn from the learned distribution. Because the prior is trained independently of any controller and then frozen, it can be applied without modification to train policies for locomotion, jumping, and other tasks while matching the motion quality of adversarial imitation methods.

What carries the argument

Score distillation sampling applied to a motion diffusion model, which converts the model's score function into a reward signal that measures how closely a generated motion matches the pre-trained distribution.

If this is right

One SMP trained on a broad dataset can directly reward policies for many unrelated tasks without any retraining of the prior.
Style-specific priors can be derived from the same base model by conditioning or selection during reward computation.
New motion styles can be created by combining the reward signals of two or more existing SMPs.
Reference motion clips no longer need to be stored or accessed once the SMP has been trained.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Animation pipelines could maintain a small collection of SMPs as interchangeable modules rather than task-specific models.
The reusability might allow iterative design where an artist first trains a general prior and later specializes it for new characters without restarting from raw data.
Composition of priors suggests a route to generating motions that blend characteristics from separate datasets in a single training run.

Load-bearing premise

That score distillation sampling applied to the diffusion model will produce stable reward signals that generalize across tasks without introducing artifacts or collapsing to limited motion modes.

What would settle it

A clear failure would be a policy trained with an SMP that produces visibly jittery, unstable, or stylistically incorrect motions on a held-out control task when compared side-by-side with motions from a retrained adversarial prior.

Figures

Figures reproduced from arXiv: 2512.03028 by Chang Shu, Chuan Guo, Dun Yang, Guy Tevet, Kotaro Imamura, Michael Taylor, Minami Matsumoto, Pengcheng Xi, Xue Bin Peng, Yi Shi, Yuxuan Mu, Ziyu Zhang.

**Figure 1.** Figure 1: Our framework constructs reusable and modular motion priors. A general motion prior can be trained on a large dataset spanning 100 styles, and then [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Figure 2: Schematic overview of the system. The dashed arrows indicate com [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: A qualitative illustration of the probability density maps of [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 5.** Figure 5: Score-matching motion priors can be trained on datasets of varying sizes, independently of any task or control policy. Once trained, an SMP provides a [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

**Figure 6.** Figure 6: A pretrained 100-style motion prior can also be adapted to synthesize [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

**Figure 7.** Figure 7: Comparison of normalized task returns across motion control tasks. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗

**Figure 9.** Figure 9: Visual snapshots of humanoid characters trained via SMP imitating [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗

**Figure 8.** Figure 8: Learning curves for single-clip imitation tasks over three random [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗

read the original abstract

Data-driven motion priors that can guide agents toward producing naturalistic behaviors play a pivotal role in creating life-like virtual characters. Adversarial imitation learning has been a highly effective method for learning motion priors from reference motion data. However, adversarial priors, with few exceptions, need to be retrained for each new controller, thereby limiting their reusability and necessitating the retention of the reference motion data when applied to downstream tasks. In this work, we present Score-Matching Motion Priors (SMP), which leverages pre-trained motion diffusion models and score distillation sampling (SDS) to create reusable task-agnostic motion priors. SMPs can be pre-trained on a motion dataset, independent of any control policy or task. Once trained, SMPs can be kept frozen and reused as general-purpose reward functions to train new policies to produce naturalistic behaviors for downstream tasks. We show that a general motion prior trained on large-scale datasets can be repurposed into a variety of style-specific priors. Furthermore, SMP can compose different styles to synthesize new styles not present in the original dataset. Our method can create reusable and modular motion priors that produce high-quality motions comparable to state-of-the-art adversarial imitation learning methods. In our experiments, we demonstrate the effectiveness of SMP across a diverse suite of control tasks with physically simulated humanoid characters. Video available at https://youtu.be/jBA2tWk6vzU

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SMP turns pre-trained motion diffusion models into frozen, reusable rewards via score distillation, which solves a real retraining headache in character control but leaves the kinematic-to-dynamic consistency question open.

read the letter

The main new piece is the construction of score-matching motion priors that get trained once on a motion dataset and then stay fixed while serving as general rewards for new policies and tasks. This sidesteps the usual need to retrain an adversarial prior or keep the reference data around every time the controller changes. They also show the same prior can be repurposed for specific styles and composed to generate combinations not seen in the original data, which is a clean modular angle for animation work.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes Score-Matching Motion Priors (SMPs) that pre-train on motion datasets using diffusion models and score distillation sampling (SDS) to produce reusable, task-agnostic motion priors. These priors are kept frozen and applied as general-purpose reward functions to train physics-based control policies for diverse downstream tasks on simulated humanoid characters. The work further demonstrates repurposing a general prior into style-specific ones and composing multiple styles to generate novel motions, claiming results comparable in quality to state-of-the-art adversarial imitation learning methods.

Significance. If the central claims hold with supporting evidence, the reusability of frozen SMPs as modular rewards would constitute a meaningful advance over task-specific adversarial priors, reducing the need to retain reference motion data or retrain for each controller. The style composition capability adds further value for synthesizing behaviors not present in the original dataset.

major comments (3)

[Method (SDS gradient term)] The SDS reward formulation (described in the method for distilling from the pre-trained kinematic motion diffusion model) does not include any physics-aware correction or contact-aware adjustment. Because the diffusion model is trained on mocap-style kinematic trajectories, the resulting score estimate can assign high rewards to dynamically infeasible poses (e.g., foot-skate or momentum violations), which the policy optimizer may exploit rather than converge to stable locomotion under the simulator's forward dynamics.
[Experiments] The experiments claim that SMP produces motions 'comparable to state-of-the-art adversarial imitation learning methods' across a diverse suite of control tasks, yet the manuscript provides no quantitative metrics, error bars, ablation studies, or direct comparisons (e.g., success rates, motion quality scores, or stability measures) to support this central claim.
[Experiments / Results] The reusability argument—that a single pre-trained SMP can be frozen and directly reused as a reward for new policies and tasks without degradation—rests on unshown experimental outcomes; no results demonstrate cross-task generalization or stability when the prior is held fixed while only the policy is optimized.

minor comments (2)

[Abstract] The abstract states that SMPs 'can compose different styles to synthesize new styles not present in the original dataset' but does not clarify whether this composition occurs at the prior level or only at the policy level.
[Method] Notation for the score function and the precise form of the SDS loss term should be introduced earlier and used consistently to aid readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. We have carefully considered each major comment and provide point-by-point responses below. Where appropriate, we have revised the manuscript to address the concerns raised.

read point-by-point responses

Referee: [Method (SDS gradient term)] The SDS reward formulation (described in the method for distilling from the pre-trained kinematic motion diffusion model) does not include any physics-aware correction or contact-aware adjustment. Because the diffusion model is trained on mocap-style kinematic trajectories, the resulting score estimate can assign high rewards to dynamically infeasible poses (e.g., foot-skate or momentum violations), which the policy optimizer may exploit rather than converge to stable locomotion under the simulator's forward dynamics.

Authors: We agree that the pre-trained diffusion model is kinematic and does not explicitly model physics. However, the SMP is used as a reward signal within a physics-based reinforcement learning framework, where the simulator's forward dynamics and contact forces inherently penalize dynamically infeasible actions. Policies that exploit kinematic rewards leading to instability (e.g., foot-skating) would receive low cumulative rewards due to falling or poor task performance. In our experiments, we observe stable, physically plausible motions without such artifacts. To strengthen the manuscript, we have added a discussion in Section 3.2 clarifying the interplay between the kinematic prior and physics constraints, and included qualitative analysis showing absence of common artifacts. revision: yes
Referee: [Experiments] The experiments claim that SMP produces motions 'comparable to state-of-the-art adversarial imitation learning methods' across a diverse suite of control tasks, yet the manuscript provides no quantitative metrics, error bars, ablation studies, or direct comparisons (e.g., success rates, motion quality scores, or stability measures) to support this central claim.

Authors: We acknowledge that the current version relies primarily on qualitative comparisons and video demonstrations. To provide stronger evidence, we have added quantitative evaluations in the revised manuscript, including task success rates, motion similarity metrics (such as average joint position error to reference motions where applicable), and stability measures (e.g., center of mass deviation). We also include direct comparisons with adversarial methods like AMP, with results averaged over multiple random seeds and reported with standard deviations. revision: yes
Referee: [Experiments / Results] The reusability argument—that a single pre-trained SMP can be frozen and directly reused as a reward for new policies and tasks without degradation—rests on unshown experimental outcomes; no results demonstrate cross-task generalization or stability when the prior is held fixed while only the policy is optimized.

Authors: The reusability is a core contribution, and our experiments do demonstrate using the same frozen SMP across different tasks (e.g., walking, running, jumping) by only training new policies. However, to make this more explicit, we have added a new subsection in the experiments detailing the fixed prior setup, with performance metrics showing consistent quality across tasks without retraining the SMP. This includes comparisons of policy performance with and without the prior to highlight generalization. revision: yes

Circularity Check

0 steps flagged

Minor self-citation present but derivation remains independent of fitted inputs or self-defined loops

full rationale

The paper pre-trains SMPs on motion datasets using external pre-trained diffusion models and applies score distillation sampling (SDS) to produce reusable frozen rewards for downstream policies. No load-bearing step reduces by construction to a parameter fitted on the target result or to a self-citation chain whose cited result itself depends on the present claims. The central reusability argument relies on established SDS and diffusion techniques from outside the authors' prior work, with the kinematic-to-physical transfer assumption left as an empirical claim rather than a definitional identity. This yields a low but non-zero circularity score for routine self-citation that does not carry the main result.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The method assumes that pre-trained diffusion models capture sufficient motion statistics and that SDS can be repurposed as a stable reward without additional learned components.

axioms (1)

domain assumption Pre-trained motion diffusion models provide a useful score function for guiding physics-based policies toward naturalistic motion.
Invoked when stating that SMPs can be kept frozen and reused as general-purpose reward functions.

pith-pipeline@v0.9.0 · 5590 in / 1180 out tokens · 25747 ms · 2026-05-17T02:02:24.527970+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

ExpertGen: Scalable Sim-to-Real Expert Policy Learning from Imperfect Behavior Priors
cs.RO 2026-03 conditional novelty 6.0

ExpertGen generates high-success expert policies in simulation from imperfect priors by freezing a diffusion behavior model and optimizing its initial noise via RL, then distills them for real-robot deployment.

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages · cited by 1 Pith paper · 1 internal anchor

[1]

https://github

threestudio: A unified framework for 3D content generation. https://github. com/threestudio-project/threestudio. Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. 2020. Robust motion in-betweening.ACM Transactions on Graphics (TOG)39, 4 (2020), 60–1. Jonathan Ho and Stefano Ermon. 2016. Generative adversarial imitation learning. Adva...

work page arXiv 2020
[2]

InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 6517–6526. Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. 2023. Magic3d: High- resolution te...

work page doi:10.48550/arxiv.2503.01143 2023
[3]

High-Dimensional Continuous Control Using Generalized Advantage Estimation

High-dimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438(2015). John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347(2017). Agon Serifi, Ruben Grandia, Espen Knoop, Markus Gross, and Moritz Bächer. 2024...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1609/aaai.v38i14.29470 2015