PhysMoDPO: Physically-Plausible Humanoid Motion with Preference Optimization

In recent years, text-conditioned human motion generation has made significant progress, largely due to diffusion models trained on large-scale human motion data. These models can generate complex humanoid motions, offering new possibilities for character animation and robot control. However, existing methods face challenges in generating physically consistent motions, especially when applying these generated motions to real robot control. To overcome these challenges, researchers have proposed various methods, including using a Whole-Body Controller (WBC) to convert generated motions into executable trajectories. However, these methods often rely on handcrafted physics-aware heuristics, such as foot-sliding penalties, which may lead to deviations between generated and original motions.

Existing diffusion model-based human motion generation methods face challenges in generating physically consistent motions. Specifically, when applying these generated motions to character animation and real robot control, a Whole-Body Controller (WBC) is often required to convert generated motions into executable trajectories. While WBC-generated trajectories comply with physical laws, they may deviate substantially from the original motion. This deviation not only affects the physical realism of generated motions but may also reduce task execution accuracy and efficiency. Therefore, balancing physical consistency and task instructions in generated motions is a pressing issue that needs to be addressed.

The core innovation of PhysMoDPO lies in its Direct Preference Optimization framework, which integrates a Whole-Body Controller (WBC) into the training pipeline to achieve dual optimization of physical consistency and textual instructions. Specifically, PhysMoDPO does not rely on traditional physics-aware heuristics but employs physics-based and task-specific rewards to assign preferences to synthesized trajectories. This innovation not only enhances the physical realism of generated motions but also improves task execution accuracy and efficiency. Additionally, PhysMoDPO represents a fundamental methodological innovation, offering new ideas and methods for future robotics and animation production.

The methodology of PhysMoDPO includes the following key steps:

• �� Integrate a Whole-Body Controller (WBC) into the training pipeline to optimize the diffusion model, ensuring that WBC outputs are compliant with both physics and original text instructions.

• �� Employ physics-based and task-specific rewards to assign preferences to synthesized trajectories, ensuring dual optimization of physical consistency and task instructions.

• �� Utilize a Direct Preference Optimization framework that does not rely on traditional physics-aware heuristics, such as foot-sliding penalties, to enhance the physical realism and task execution accuracy of generated motions.

• �� Validate PhysMoDPO's performance on simulated robots in text-to-motion and spatial control tasks, demonstrating consistent improvements in physical realism and task-related metrics.

To validate the effectiveness of PhysMoDPO, the researchers designed a series of experiments. In text-to-motion and spatial control tasks, tests were conducted using simulated robots. Multiple benchmark datasets were used in the experiments, including large-scale human motion datasets and task-specific datasets. The researchers also conducted ablation studies to verify the effectiveness of each component within PhysMoDPO. The experimental results showed that PhysMoDPO performed excellently in terms of physical realism and task-related metrics, significantly improving performance in zero-shot motion transfer tasks.

The experimental results demonstrated that PhysMoDPO performed excellently in text-to-motion and spatial control tasks. In simulated environments, physical realism metrics improved by 15%, and task success rates increased by 20%. Additionally, PhysMoDPO significantly improved performance in zero-shot motion transfer tasks, particularly in real-world deployment on a G1 humanoid robot, where task success rates increased by 20%. Ablation studies confirmed the effectiveness of each component within PhysMoDPO, highlighting the critical role of preference optimization in enhancing physical consistency.

PhysMoDPO's application scenarios include character animation production and robot control. In character animation production, PhysMoDPO can generate physically consistent humanoid motions, enhancing the realism and appeal of animations. In robot control, PhysMoDPO can generate motion trajectories that comply with physical laws, improving the accuracy and efficiency of robot task execution. Additionally, PhysMoDPO performs excellently in zero-shot motion transfer tasks, enabling rapid adaptation of motions across different environments.

Despite PhysMoDPO's excellent performance in generating physically consistent humanoid motions, it may experience performance degradation in handling extremely complex motion scenarios. Additionally, the method incurs high computational costs during real-world robot deployment, necessitating further optimization for real-time applications. In certain specific tasks, the preference optimization mechanism may require adjustments based on task characteristics to achieve optimal performance. Future research directions include optimizing the computational efficiency of PhysMoDPO to meet real-time application demands and exploring its application across more diverse motion scenarios and tasks.