PARC: Physics-based Augmentation with Reinforcement Learning for Character Controllers

SIGGRAPH 2025

Michael Xu (1)    Yi Shi (1)    KangKang Yin (1)    Xue Bin Peng (1, 2)

(1) Simon Fraser University    (2) NVIDIA

---

Abstract

Humans excel in navigating diverse, complex environments with agile motor skills, exemplified by parkour practitioners performing dynamic maneuvers, such as climbing up walls and jumping across gaps. Reproducing these agile movements with simulated characters remains challenging, in part due to the scarcity of motion capture data for agile terrain traversal behaviors and the high cost of acquiring such data. In this work, we introduce PARC (Physics-based Augmentation with Reinforcement Learning for Character Controllers), a framework that leverages machine learning and physics- based simulation to iteratively augment motion datasets and expand the capabilities of terrain traversal controllers. PARC begins by training a motion generator on a small dataset consisting of core terrain traversal skills. The motion generator is then used to produce synthetic data for traversing new terrains. However, these generated motions often exhibit artifacts, such as incorrect contacts or discontinuities. To correct these artifacts, we train a physics-based tracking controller to imitate the motions in simulation. The corrected motions are then added to the dataset, which is used to continue training the motion generator in the next iteration. PARC’s iterative process jointly expands the capabilities of the motion generator and tracker, creat- ing agile and versatile models for interacting with complex environments. PARC provides an effective approach to develop controllers for agile terrain traversal, which bridges the gap between the scarcity of motion data and the need for versatile character controllers.

Arxiv       Code: [coming soon]      

Video

Method

PARC iteratively trains a motion generator and motion tracker while expanding a physically accurate motion-terrain dataset. At the beginning of each iteration, the motion generator is trained on the current dataset. The trained motion generator is then used to generate kinematic motions, which is used to supply a physics-based motion tracking controller with reference data. The motion tracker is trained using reinforcement learning, and is then used to physically correct the generated kinematic motions. The physically corrected motions are added back into the dataset for the next iteration.

Original Data

PARC's original dataset contains of core terrain-traversal skills captured at Beyond Capture Studios.

PARC Generated Data

As the augmentation loop progresses, PARC is able to discover interesting phyiscally accurate behaviors for terrain traversal. For example, combining a jumping motions with a climbing motion, even when the original dataset did not contain this combination. Another example is dropping from a higher ledge and catching onto a lower ledge. The original dataset contained motions for climbing up and down, but now this specific behavior.

Long Horizon Results

Using the motion generator and motion tracking controller together, we can generate physically realistic and interesting motions for traversing complex terrain.

Bibtex

@inproceedings{xu2025parc,
    author = {Xu, Michael and Shi, Yi and Yin, KangKang and Peng, Xue Bin},
    title = {PARC: Physics-based Augmentation with Reinforcement Learning for Character Controllers},
    year = {2025},
    booktitle={SIGGRAPH 2025 Conference Papers (SIGGRAPH '25 Conference Papers)}
}