Seminar: Graduate Seminar
Physics-Based Character Control Through Task Encoders
Date:
February,12,2025
Start Time:
11:30 - 12:30
Location:
506, Zisapel Building
Zoom:
Zoom link
Add to:
Lecturer:
Ron Vainshtein
Research Areas:
| We present a novel approach to Physics-Based Character Control Through Task Encoders, enabling efficient task adaptation by leveraging a large, pretrained model for action inpainting. The original model synthesizes character motion by completing partial future pose information, but control over the system requires extensive computational resources and training. Our method simplifies this process by introducing task-specific task encoders. These encoders map high-level control inputs, such as a desired direction, into the latent space of the pretrained model, which remains frozen during this adaptation phase. This framework drastically reduces the computational cost of training, requiring only a single GPU for a few days to adapt to new tasks, while retaining the versatility and natural motion quality of the unified model. We evaluate this approach on a range of tasks, including steering, path-following, and object interactions, demonstrating its ability to generalize effectively with minimal retraining. By decoupling task-specific adaptation from the expensive training of the unified model, our method establishes a scalable and resource-efficient paradigm for physics-based character control in animation, gaming, and robotics. M.Sc. student under the supervision of Prof.Shie Manor.
|