PAIWorld: 3D-consistent world model for robotic manipulation

PAIWorld is a 3D-consistent world foundation model designed for robotic manipulation, addressing the critical limitation of current multi-view world models that lack explicit geometric reasoning. Developed by a team of 28 researchers, the framework builds on a DiT-based architecture and introduces three key innovations: Geometry-Aware Cross-View Attention blocks that establish explicit communication channels across camera views, Geometric Rotary Position Embedding that encodes ray directions and camera poses directly into attention, and Latent 3D-REPA which distills 3D-aware features from frozen 3D foundation models. These components jointly solve cross-view object drift, depth inconsistency, and texture misalignment.

Benchmark results show PAIWorld achieving state-of-the-art multi-view 3D consistency, currently ranking 1st on the WorldArena leaderboard and 2nd on the AgiBot-Challenge2026 leaderboard for robotic manipulation. Beyond simulation accuracy, the model enables practical downstream applications including model-based planning, world action models that predict action outcomes, and multi-view policy post-training for real robot systems. This work provides a strong foundation for robots to perceive and interact with 3D environments through multiple cameras—a core requirement for real-world manipulation tasks.