PMP: Learning to Physically Interact with Environments using Part-wise Motion Priors

Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, Young Min Kim

Introduction

Physics-based character animation aims to create virtual agents that move realistically while interacting with dynamic environments. Achieving this requires balancing physical constraints with natural-looking motion. While Adversarial Motion Priors (AMP) have significantly advanced the field by leveraging motion capture (mocap) data, they often struggle with complex, multi-faceted tasks where specific full-body reference data is unavailable. This paper introduces Part-wise Motion Priors (PMP), a framework designed to overcome data scarcity by decomposing motion into modular, part-specific components.

The Challenge of Interaction Data

Traditional methods rely on a global discriminator to evaluate whether a character's movement matches human-like patterns. However, capturing high-fidelity data for every possible interaction—such as carrying a heavy object while navigating stairs—is impractical. Mocap datasets are often fragmented: one clip might show walking, while another shows grasping. Furthermore, capturing intricate hand-object interactions is technically difficult due to finger occlusions and high degrees of freedom. This "data scarcity" prevents agents from learning complex tasks that require simultaneous locomotion and manipulation.

Methodology

The PMP framework addresses these challenges through a compositional approach to motion learning, breaking the "whole-body" problem into manageable segments.

1. Part-wise Motion Priors (PMP)

Instead of a single discriminator for the entire body, PMP employs multiple discriminators assigned to specific joint sets (e.g., upper body, lower body, and hands). This allows the agent to synthesize motion by "mixing and matching" styles from different reference clips. For instance, an agent can adopt a "soldier walk" for its legs while performing a "lifting" motion with its arms, even if no single mocap clip contains both behaviors.

2. Interaction Prior and Interaction Gym

To solve the specific difficulty of hand-grasping, the authors developed an Interaction Gym. This synthetic environment is used to train a standalone hand model to grasp a simple cylindrical rod. The resulting "expert" behavior serves as a pre-trained Interaction Prior. By using simulation-generated data, the framework bypasses the need for difficult-to-capture hand mocap, providing the agent with the necessary kinematic insights for dexterous manipulation that generalizes to various objects.

3. Multiplicative Style Reward

To ensure that the character coordinates its various parts effectively, PMP utilizes a multiplicative reward structure. The total style reward is the product of individual part-wise rewards. This mathematical choice forces the agent to satisfy the stylistic requirements of all body parts simultaneously; if the legs move correctly but the hands fail to grasp, the total reward remains near zero, pushing the agent toward a unified solution.

4. Demo Blend Technique

Learning complex interactions can lead to training stalls, particularly when the multiplicative reward is low. The authors introduced "Demo Blend," a technique that randomly mixes successful trajectories for different body parts during the training process. This maintains a consistent reward signal, helping the agent navigate difficult learning phases where one part of the body might otherwise fail to find a valid solution.

Experimental Setup and Datasets

The framework was evaluated using a combination of real-world and synthetic data:

  • Mixamo: A standard library used for fundamental locomotion styles such as walking, jumping, and idling.
  • Interaction Gym: A custom-built simulation environment for generating synthetic hand-interaction priors.
  • Custom Scenarios: The system was tested on seven high-difficulty tasks, including Upstair Carrying, Sight Locomotion (maintaining gaze while moving), and Rope Climbing.

Results and Discussion

PMP demonstrates significant improvements over standard AMP baselines across challenging interaction tasks. In the Barbell Lifting task, PMP achieved a normalized return of 0.55, whereas the baseline failed almost entirely with a score of 0.022. Similarly, in Rope Climbing, PMP reached a score of 0.63, nearly doubling the performance of the baseline (0.31).

Beyond performance metrics, PMP excels in Style Augmentation. The system can successfully combine distinct walking styles (like "Hopping" or "Soldier") with upper-body tasks without losing the characteristic "flavor" of the movement. Furthermore, the agents showed remarkable Robustness, adapting to varying object weights (up to 30kg) and different handle shapes without requiring task-specific mocap for every iteration.

Limitations and Future Work

Despite its success, PMP has areas for improvement. Currently, body part segmentation (defining which joints belong to which discriminator) is defined manually. Future research could explore automated segmentation to make the framework more flexible. Additionally, while PMP reduces the need for motion data, it still requires manual "task reward" engineering to define the agent's goals. Integrating visual inputs and developing explicit global coordination modules could further enhance the agent's ability to navigate complex, unlabeled environments.

Summary updated

BibTeX
@inproceedings{bae2023pmp,
  author = {Jinseok Bae and Jungdam Won and Donggeun Lim and Cheol-Hui Min and Young Min Kim},
  title = {{PMP: Learning to Physically Interact with Environments using Part-wise Motion Priors}},
  booktitle = {SIGGRAPH (Conference Paper Track)},
  year = {2023},
  url = {https://openreview.net/forum?id=xFHIYShf0t},
}
Figure