PAIWorld Tops WorldArena Ranking, Showcasing Industrial Embodied AI Breakthroughs

PAIWorld, developed by the PAI Lab at the Chinese Academy of Sciences' Institute of Industrial AI, recently secured the top position on the WorldArena leaderboard with a total score of 72.31, reflecting comprehensive strength across visual quality, motion quality, content consistency, physical compliance, 3D accuracy and controllability.

WorldArena is the most authoritative evaluation suite for embodied world models, aggregating submissions from leading groups such as WorldLab (led by Fei‑Fei Li), Google, NVIDIA, Stanford, ZhiYuan Robotics, Beijing Humanoid Robot Innovation Center, Gaode, Xiaomi, and others, making the competition extremely fierce.

In the detailed breakdown, PAIWorld leads the Motion Smoothness metric with 95.41 points, demonstrating superior temporal consistency in generated motions. It also outperforms the second‑place model by 7.4 points on the Trajectory Accuracy metric, indicating accurate long‑term prediction of object and camera trajectories with minimal drift.

Technical Highlights

3D Geometry Prior Injection: A foundational 3D model supplies explicit depth, surface geometry and occlusion constraints, enabling stable structural consistency over long sequences and complex interactions.

Geometric Rotational Position Encoding (Geo‑RoPE): Attention heads are split into ray‑space and pose‑space sub‑spaces, encoding pixel‑level 3D ray directions and camera pose information, which grants inherent cross‑view 3D geometric perception.

Multi‑View Attention Mechanism: Integrated into the backbone video generation network, this mechanism aligns geometric and appearance information across viewpoints for each generated frame, achieving precise physical world simulation.

The accompanying figures illustrate PAIWorld’s performance in multi‑object interaction scenes, hinge‑type manipulation scenarios, and high‑quality spatio‑temporal reconstruction, confirming the model’s ability to maintain consistent geometry and motion.

Earlier, a predecessor of PAIWorld earned the runner‑up spot in the AGIBOT World Challenge@ICRA 2026 world‑model track and topped the “scene consistency” metric, a key indicator of physical environment understanding. The challenge attracted 336 top teams worldwide.

Building on these successes, the PAI Lab plans to combine the world model with a World Action Model to create a closed‑loop embodied data pipeline, ultimately enabling self‑improving, continuously evolving embodied intelligence in real‑world settings.