WorldDreamer V4 Leads Benchmarks, Paving the Way for Collective Intelligence in World Models

"One World One Model" is presented as the guiding principle behind the next‑generation AI architecture, arguing that the future of automation lies in coordinated groups of robots rather than a single all‑purpose machine.

WorldDreamer V4 departs from traditional single‑agent world models by training on multi‑agent tasks from the outset. Its core is a Multi‑Agent End‑to‑End Pretraining Architecture that ingests visual observations, state information, task commands, and action trajectories from multiple agents with differing viewpoints, enabling the model to learn a shared representation of the world.

The system introduces Masked Agent Modeling , where during training random agents are hidden and the model must predict the hidden agents' current state, future actions, and collective coordination strategy, forcing the model to reconstruct a complete world view from partial observations.

Key capabilities of WorldDreamer V4 include:

Understanding complex physical environments;

Modeling relationships among multiple agents;

Predicting future world states;

Generating joint actions for multiple agents;

Supporting knowledge transfer across different robot morphologies.

Instead of relying on pairwise communication (Agent A → Agent B → Agent C), the model adopts a Shared Latent World State , allowing all agents to depend on a continuously evolving world model, thereby reducing communication overhead and shifting collaboration from "information exchange" to "information prediction".

Benchmark evaluations demonstrate the model's strength: on the RoboCasa robot‑control benchmark, WorldDreamer V4 achieved the top rank without task‑specific fine‑tuning, proving that its multi‑agent pretraining captures general world‑action regularities that excel in complex and unseen tasks. On the WorldScore world‑generation benchmark, it also secured first place, excelling in world consistency, camera control, object control, 3D stability, and dynamic evolution.

These results support the authors' claim that the model is not merely a video generator but a controllable, consistent, dynamic world simulator.

The article further proposes a Multi‑Agent Scaling Law (referred to as the "MELKOV robot law"), suggesting that as the number of robots grows, data richness, model strength, robot capability, and deployment scenarios all increase, enabling a shared world model to connect diverse platforms such as industrial arms, AGVs, quadrupeds, humanoids, drones, service robots, and space robots.

Ultimately, WorldDreamer V4 is positioned as a step toward a future where a single shared world model links millions of agents, moving from "one brain per robot" to "one brain controlling all" and heralding the onset of the collective intelligence era for physical AI.