Researchers at the Max Planck Institute for Intelligent Systems and ETH Zurich have developed WANDR, an innovative model that generates realistic human motions for 3D avatars. This technology, set to be presented at the upcoming CVPR 2024 conference, aims to enhance interactions in virtual environments, such as the metaverse and digital entertainment.
WANDR utilizes a conditional Variational Autoencoder to unify various data sources, allowing avatars to reach specified goals—like grabbing objects—through fluid and natural movements. "Our research focuses on understanding what it takes to create virtual humans that behave like us," said Markos Diomataris, the study's first author.
Key Features of WANDR
Intention-Driven Motion: The model generates actions based on intention features, which guide the avatar's orientation and movements toward a goal.
Data-Driven Learning: WANDR leverages large datasets of general human motions alongside smaller, specialized datasets focusing on reaching tasks. This hybrid approach enables the avatar to learn nuanced skills without requiring extensive reinforcement learning (RL).
Active Feedback Loop: WANDR operates with an autoregressive mechanism, predicting frame-by-frame actions that adjust continuously, akin to human motion dynamics. This allows the avatar to adapt and reach moving or sequential goals even without prior training for those specific tasks.
Comparison with Traditional Approaches
While traditional methods often rely heavily on reinforcement learning, which can be computationally expensive and require extensive trial and error, WANDR's data-driven framework simplifies the process. By hallucinating potential future positions as goals during training, the model can effectively combine various datasets, enhancing its ability to generalize across different tasks.
Diomataris noted, "By teaching avatars to replicate human movements through observational learning, we can create more lifelike interactions in virtual spaces." This groundbreaking work paves the way for more realistic avatars that can navigate and interact within their environments seamlessly.