Skip to content

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

🕒 Published (v1): 2025-12-02 18:54 UTC · Source: Arxiv · Venue: SIGGRAPH 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Score-Matching Motion Priors (SMP) repurposes frozen, pretrained motion diffusion models as reusable reward functions for physics-based character control via score distillation sampling (SDS). Unlike adversarial imitation learning, SMP decouples prior training from policy training, so a single prior can train many downstream controllers without retaining the original dataset. The method matches state-of-the-art adversarial baselines while enabling style composition and style-specific specialization through prompting.

Problem

Adversarial motion priors (e.g., AMP) must be retrained jointly with each new control policy and require perpetual access to the reference motion dataset, limiting modularity and reusability. No prior method produces a single frozen motion prior that can be applied across arbitrary downstream tasks and policies without further updates.

Method

SMP trains a motion diffusion model (DDPM-style) on an unstructured motion capture dataset independently of any task or policy. Once trained, the diffusion model is frozen and used as a stationary reward function via score distillation sampling: for a given simulated motion trajectory, Gaussian noise is added at a sampled diffusion timestep to produce a noisy sample, then the frozen denoiser predicts the clean sample, and the SDS gradient signal drives the RL policy toward the learned motion distribution. The SDS reward is r_smp = exp(−w_s ‖ε̂ − ε‖²), where ε̂ is the diffusion model's noise prediction and ε is the added noise, measuring how well the simulated motion aligns with the reference distribution. Style-specific priors are derived from the general prior via conditioning/prompting (e.g., classifier-free guidance) without retraining; multiple style priors can be composed to synthesize novel styles absent from the training data. The overall policy is trained with standard RL (PPO), combining the SMP reward with task-specific objectives in a goal-conditioned framework.

Key Contributions

  • A frozen, task-agnostic motion prior built from a pretrained diffusion model that requires no retraining or dataset access when applied to new controllers or tasks.
  • Application of SDS as a stationary RL reward function for physics-based character animation.
  • Demonstration that one general prior (trained on 100 motion styles) can be repurposed into 100 style-specific priors via conditioning, with no additional model updates.
  • Style composition: priors for distinct styles can be blended to produce novel styles not present in the original dataset.
  • Release via MimicKit codebase.

Results

  • SMP achieves motion quality comparable to state-of-the-art adversarial imitation learning (AMP and variants) across a diverse suite of humanoid control tasks (locomotion, goal-reaching, terrain traversal, etc.).
  • A single general prior trained on 100 motion styles can be directly repurposed into 100 style-specific priors without any re-training or dataset access.
  • Style composition produces novel behavioral styles not contained in the original 100-style dataset.
  • Specific quantitative metrics (numeric scores, FID, task success rates) are not included in the provided text excerpt.

Limitations

  • Quantitative performance is described as "comparable" to adversarial baselines, not clearly superior; no numeric advantage is stated in the provided text.
  • SDS-based reward gradients can be noisy; the method requires careful design choices (mentioned but not fully detailed in the excerpt) to achieve high-quality motions.
  • Style conditioning and composition rely on the diffusion model's generalization; out-of-distribution styles may degrade quality.
  • Closest prior work (SMILING) also uses a score-matching objective but requires task-specific diffusion models; SMP's advantage over SMILING is modularity, not necessarily motion quality.
  • Physical plausibility is enforced by the RL simulator, but the diffusion prior itself was trained on kinematic mocap data, so distribution mismatches remain possible.

Relevance to Agentic AI / LLM Agents

This paper is only tangentially relevant to the LLM-agent tracking topic: it addresses RL-based embodied agents in physics simulation rather than language-grounded or tool-using agents. The core contribution — using a frozen generative model as a reusable, modular reward function that decouples prior learning from policy learning — is architecturally analogous to debates in agentic AI about separating world models or value functions from task-specific policies. The SDS-as-reward paradigm could inform future work on reusable reward shaping in LLM-agent training pipelines, and the style-composition mechanism parallels skill or capability composition in multi-agent and skill-chaining systems. Researchers tracking agentic AI should note this mainly as background on generative-model-as-reward techniques rather than a direct advance in LLM agents.