WorldArena Challenge 2.0 is built around the next-generation WorldArena 2.0 benchmark for embodied world models.
The benchmark extends the original WorldArena along three coordinated axes: modality, functionality, and platform.
WorldArena 2.0 introduces visuo-tactile evaluation, evaluates world models as online reinforcement learning environments,
and expands evaluation from simulator-only settings toward simulated and real robotic platforms.
Following this benchmark direction, the challenge is organized into three competition tracks:
- Track 1: Video Quality Evaluation - a continuation of the 1.0 video quality task with increased difficulty and new OOD scenes.
- Track 2: World Model as RL Environment - evaluating whether world models can serve as interactive environments for reinforcement learning and policy optimization.
- Track 3: Real-World WAM Manipulation - evaluating real-world manipulation ability across tactile WAM and vision-only WAM settings.