PAIWorld: 3D-Consistent World Foundation Model Advances Robotic Manipulation

Researchers have introduced PAIWorld, a world foundation model designed to overcome a critical limitation in robotic manipulation systems: the lack of multi-view 3D consistency. While existing world foundation models are powerful simulators, they predominantly operate in single-view settings and fail to maintain geometric coherence across the multiple camera perspectives that robotic systems require for effective policy learning. Current multi-view world models simply concatenate view tokens without explicit geometric reasoning, leading to cross-view object drift, depth inconsistency, and texture misalignment.

PAIWorld addresses these deficiencies through a three-component framework built upon a diffusion-transformer architecture. The model incorporates Geometry-Aware Cross-View Attention blocks that establish explicit communication pathways between different camera views, Geometric Rotary Position Embedding that encodes camera ray directions and extrinsic poses directly into the attention mechanism, and Latent 3D-REPA, which distills 3D-aware features from frozen 3D foundation models to enforce 3D consistency. This integrated approach simultaneously resolves the absence of inter-view communication mechanisms and the lack of 3D geometric priors.

The framework has demonstrated state-of-the-art performance on robotic manipulation benchmarks. PAIWorld achieved first place on the WorldArena leaderboard and second place on the AgiBot-Challenge2026 leaderboard, validating its multi-view 3D consistency capabilities. The model supports downstream applications including model-based planning, world action models, and multi-view policy post-training, enabling more robust and accurate robotic control systems. The research was submitted to arXiv on June 16, 2026, within the Computer Science > Robotics category.