Manifold AI’s WorldScape 0.2 Tops WorldArena: How MoE Drives Superior Physics and 3D Understanding

WorldArena, the first unified "function + vision" benchmark for embodied world models created by leading institutions such as Tsinghua, Peking, Hong Kong, Princeton, and others, recently released its latest leaderboard. Manifold AI’s self‑developed model WorldScape 0.2 claimed the global #1 position, surpassing foreign giants including Google and Nvidia as well as domestic competitors.

The ranking reflects a balanced superiority across multiple dimensions. In the comprehensive perception score—covering visual quality, motion quality, content consistency, and controllability—WorldScape 0.2 ranked first with no evident trade‑off among sub‑metrics. It also achieved the top physics‑compliance score, indicating that the model internalizes gravity, friction, collisions, and force feedback, producing physically plausible motions rather than merely visually plausible ones. Additionally, the model excelled in the high‑difficulty 3D accuracy metric, maintaining precise geometric structures during complex robotic arm manipulation, viewpoint changes, and occlusions.

The breakthrough is attributed to the integration of a Mixture‑of‑Experts (MoE) architecture. MoE introduces multiple specialized sub‑networks (experts) and a dynamic gating mechanism that activates only the experts most relevant to the current input, allowing parameter scaling by several orders of magnitude without proportional computational cost.

WorldScape 0.2 leverages MoE in three concrete ways:

Multi‑expert collaborative generalization: The architecture expands from a single‑task model to a unified framework that jointly learns diverse control signals, enabling fine‑grained robotic manipulation and other embodied behaviors to benefit from shared expert knowledge.

Unified spatial representation: Beyond simple geometric priors, the model aligns geometry, semantics, and physics in a shared implicit latent space, ensuring consistent spatial topology, semantic coherence, and physical plausibility across long‑range interactions.

Multi‑stage continual learning: A progressive training schedule injects massive world knowledge and couples heterogeneous control signals, shifting the focus from visual fidelity to strict adherence to physical laws, which is reflected in the top WorldArena scores.

Remarkably, WorldScape 0.2 uses only about 10 % of the parameter count of other top‑ranked models, demonstrating a superior “spatial intelligence density” and real‑time inference capability, which also supports edge‑side physical AI deployments.

Overall, the combination of MoE‑driven scaling and the model’s balanced performance across perception, physics, and 3D understanding suggests that a GPT‑3‑like era for world models may be imminent.