Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ReinDiffuse: Crafting Physically Plausible Motions with Reinforced Diffusion Model

Gaoge Han, Mingjiang Liang, Jinglei Tang, Yongkang Cheng, Wei Liu, Shaoli Huang

TL;DR

ReinDiffuse is presented, which combines reinforcement learning with motion diffusion model to generate physically credible human motions that align with textual descriptions, and adapts Motion Diffusion Model to output a parameterized distribution of actions, making them compatible with reinforcement learning paradigms.

Abstract

Generating human motion from textual descriptions is a challenging task. Existing methods either struggle with physical credibility or are limited by the complexities of physics simulations. In this paper, we present \emph{ReinDiffuse} that combines reinforcement learning with motion diffusion model to generate physically credible human motions that align with textual descriptions. Our method adapts Motion Diffusion Model to output a parameterized distribution of actions, making them compatible with reinforcement learning paradigms. We employ reinforcement learning with the objective of maximizing physically plausible rewards to optimize motion generation for physical fidelity. Our approach outperforms existing state-of-the-art models on two major datasets, HumanML3D and KIT-ML, achieving significant improvements in physical plausibility and motion quality. Project: https://reindiffuse.github.io/

ReinDiffuse: Crafting Physically Plausible Motions with Reinforced Diffusion Model

TL;DR

ReinDiffuse is presented, which combines reinforcement learning with motion diffusion model to generate physically credible human motions that align with textual descriptions, and adapts Motion Diffusion Model to output a parameterized distribution of actions, making them compatible with reinforcement learning paradigms.

Abstract

Generating human motion from textual descriptions is a challenging task. Existing methods either struggle with physical credibility or are limited by the complexities of physics simulations. In this paper, we present \emph{ReinDiffuse} that combines reinforcement learning with motion diffusion model to generate physically credible human motions that align with textual descriptions. Our method adapts Motion Diffusion Model to output a parameterized distribution of actions, making them compatible with reinforcement learning paradigms. We employ reinforcement learning with the objective of maximizing physically plausible rewards to optimize motion generation for physical fidelity. Our approach outperforms existing state-of-the-art models on two major datasets, HumanML3D and KIT-ML, achieving significant improvements in physical plausibility and motion quality. Project: https://reindiffuse.github.io/

Paper Structure

This paper contains 16 sections, 8 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Motion Diffusion Model
  5. Reinforcement Learning
  6. Method
  7. Adapting MDM with Reparameterization
  8. Reinforcement Learning Training
  9. Reward Functions for Physical Plausibility
  10. Loss Functions and Optimization
  11. Experiments
  12. Datasets and Evaluation Metric
  13. Implementation Details
  14. Comparison to State-of-the-art Approaches
  15. Ablation Studies
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Our ReinDiffuse can generate physically plausible motion, effectively eliminating common physical issues such as floating, penetration, foot clipping, and skating. ReinDiffuse enables MDM to learn physical commonsense with reinforcement learning.
  • Figure 2: Overview of our ReinDiffuse training framework. Given the condition $c$, time step $t$ and noised motion $\epsilon_t$. Initial MDM output the $x_a$ as the mean of Gaussian distribution and use the fixed $\sigma$ as the standard deviation. We sample $a$ from the distribution to compute the physically plausible rewards and obtain the motion likelihood $q(a)$ and $p(a)$. We perform RL training using the combined loss of $\mathcal{L}_{PPO}$ and $\mathcal{L}_{simple}$ with PPO to update tuned MDM.
  • Figure 3: Visual results of ReinDiffuse against the MDM. The darker colors indicate the later frame in time.