Abstract
In computer animation, driving a simulated character with lifelike motion is challenging. Current generative models, though able to generalize to diverse motions, often pose challenges to the responsiveness of end-user control. To address these issues, we introduce RACon: Retrieval-Augmented Simulated Character Locomotion Control. Our end-to-end hierarchical reinforcement learning method utilizes a retriever and a motion controller. The retriever searches motion experts from a user-specified database in a task-oriented fashion, which boosts the responsiveness to the user's control. The selected motion experts and the manipulation signal are then transferred to the controller to drive the simulated character. In addition, a retrieval-augmented discriminator is designed to stabilize the training process. Our method surpasses existing techniques in both quality and quantity in locomotion control, as demonstrated in our empirical study. Moreover, by switching extensive databases for retrieval, it can adapt to distinctive motion types at run time.
Introduction
The quest for realistic character animation in physics-based simulations often encounters a fundamental trade-off between motion quality and user controllability. Traditional motion tracking methods excel at realism by strictly following Motion Capture (MoCap) data but lack the flexibility to adapt to dynamic user inputs or novel environments. Conversely, modern generative models offer high diversity but frequently behave as "black boxes," making precise control via gamepads or other interfaces difficult. Furthermore, many existing systems are rigid, requiring extensive retraining to incorporate new motion styles.
RACon (Retrieval-Augmented Locomotion Control) addresses these challenges by introducing a "search engine" paradigm for character motion. By integrating a learnable retriever with a physics-based controller, the framework achieves high responsiveness to user commands while maintaining natural, physically plausible movement.
Methodology
RACon utilizes a Hierarchical Reinforcement Learning (HRL) framework organized into a manager-worker structure.
Task-Oriented Learnable Retrieval (TOLR)
The "manager" component is a neural network-based retriever. Unlike traditional systems that rely on hand-crafted heuristics, TOLR generates adaptive queries based on the character's current state and the user's target goal. It performs a k-Nearest Neighbors (kNN) search within a motion database to identify the "expert" motion clip that best facilitates the desired transition.
Simulated Character Controller
The "worker" component is a physics-based controller operating within the Isaac Gym simulator. It receives the retrieved reference motion and translates it into joint torques. To ensure robustness, the controller prioritizes end-effector constraints (e.g., foot placement) to maintain balance while allowing for the flexibility needed to satisfy physical laws.
Retrieval-Augmented Discriminator (RA-GAIL)
To maintain visual fidelity, the system employs a specialized discriminator. By comparing simulated transitions against retrieved expert states, the discriminator provides a strong motion prior. This "retrieval-augmented" term stabilizes the adversarial training process, preventing the character from adopting unnatural or "glitchy" poses common in standard reinforcement learning setups.
Datasets and Experimental Setup
The framework was trained and evaluated using several large-scale datasets:
- AMASS & BABEL: Used for general locomotion (walking, running, turning), comprising approximately 40,000 seconds of motion data.
- Specialized Databases: Custom sets were developed for complex tasks, including a Cartwheel database (~6,590 clips) and a Zombie Walk database (142 clips). The latter served as a test for the system's ability to handle small, "plug-and-play" datasets at runtime.
Results and Discussion
RACon demonstrates superior performance across both quantitative metrics and qualitative evaluations.
Quantitative Performance
In comparison with state-of-the-art methods such as DReCon, AMP, and ASE, RACon showed marked improvements in stability and accuracy:
- Stability: RACon achieved a Termination Rate (TRate) of 0.44%, significantly outperforming ASE (1.65%) and DReCon (28.06%), indicating a much lower likelihood of the character falling.
- Control Accuracy: The system recorded the lowest Mean Velocity Error (MVE) of 2.709 m/s, demonstrating high precision in following user-defined velocity vectors.
Qualitative Versatility
A key strength of RACon is its ability to switch motion styles "zero-shot" at runtime. By simply swapping the retrieval database, the character can transition from standard locomotion to specialized movements like cartwheels or zombie walks without retraining.
Conclusion and Future Work
RACon successfully bridges the gap between retrieval-based realism and generative flexibility. While the system exhibits slight latency during extreme accelerations—a byproduct of realistic physics—it offers a robust and extensible solution for interactive character control. Future research will focus on scaling the system to handle complex object interactions and even larger, more diverse motion libraries.
Summary updated