Project Page | Paper | GitHub
L-IVA (Long-horizon Interactive Visual Avatar) is a benchmark for evaluating goal-directed planning in stochastic generative environments. It is designed to evaluate the ability of video avatars to autonomously pursue long-term goals through adaptive environmental interaction.
The benchmark accompanies the paper "Active Intelligence in Video Avatars via Closed-loop World Modeling", which introduces ORCA… See the full description on the dataset page:
https://huggingface.co/datasets/Alexhe101/L-IVA.
