Koala Youran's Youran Wujie World Model Takes Second in WorldArena 2.0, Leads Two Core Metrics

The evaluation gathered 80 models from Alibaba, the Chinese Academy of Sciences, other top technology companies, research institutions and embodied-AI unicorn companies, Leiphone reported. The models were assessed on video generation quality, temporal consistency and physical evolution. Track 1 does not ask whether a generated image looks similar to a target; it tests whether a model can maintain the physical consistency of scene geometry, texture appearance and object and interaction states over long rollouts. Background Consistency measures whether the model can keep the global environment and scene layout stable throughout generation, addressing a common failure in long video generation known as scene drift. JEPA Similarity measures distance from the real world's evolution process in a high-level representation space rather than at the pixel level, examining whether scene semantics, object states and action changes are correctly preserved. The two first-place results led to the conclusion that the model had not merely generated realistic images but had learned how the physical world evolves over time.

Koala Youran's technical route combines a world model, an agent and scene replication. The base is the Youran Wujie world model, the upper layer is the Geek Mind embodied brain, and a three-cross capability connects spaces, tasks and bodies. Leiphone reported that the decisive factor for world models is architecture rather than compute. The team divided the problem into two challenges: error accumulation during long-horizon rollouts and distribution shift in unseen environments. It designed two architectural responses. Structured 4D world modeling jointly models first-frame 3D scene geometry and motion trajectories, linking static spatial structure with dynamic action processes and providing continuous geometry and motion constraints for each frame. This suppresses object position drift, trajectory divergence and error accumulation over time. Dynamic scene memory continuously maintains and updates key environmental features such as texture, appearance and layout. During long generation, a model can gradually forget the scene itself and regress toward its training distribution; dynamic memory refreshes its understanding of the current environment, improving long-range consistency and out-of-distribution generalization. Training uses a coarse-to-fine two-stage strategy. The first stage learns general scene representations and motion rules from large-scale, diverse data. The second stage fine-tunes on selected high-quality data to strengthen geometry consistency, temporal coherence and motion accuracy.

On Track 1, the Youran Wujie world model showed highly consistent future rollouts in scene, motion and interaction, mitigating state drift, error accumulation and physical distortion in complex long-horizon generation. It handled out-of-distribution generalization tasks by using an initial scene state and a given action sequence to roll out the future while preserving scene geometry, texture appearance and physical consistency in robot-object interactions. For long-horizon tasks involving long, multi-stage continuous operations, it continued to roll out future states along the action sequence, suppressing error accumulation and state drift and maintaining long-term consistency among the scene, objects and interactions. In complex comprehensive tasks that combined out-of-distribution scenes, dual-arm coordination and long-horizon operations, it produced stable, continuous and physically consistent future rollouts, according to Leiphone.

Geek Mind, built on the Youran Wujie world model, has already been deployed in industry. In industrial inspection, manufacturing faces high labor costs, restrictions on work in dangerous areas and difficulty standardizing manual records. Manufacturers need an inspection system with autonomous perception and decision-making, not a single-point detection tool. Koala Youran integrates Geek Mind, a general embodied agent, into a quadruped robot platform for inspection solutions in large central state-owned enterprise manufacturing workshops. The solution has three core capabilities. Rapid deployment does not require production-line modification; pretrained models and scene adaptation allow quick configuration of inspection routes and task switching, shortening the period from entry to operation. Deep reasoning uses a world-model-driven perception-reasoning loop. The system collects status data such as temperature, vibration and instrument readings, performs causal analysis of anomalies and outputs traceable judgments to reduce false positives and false negatives. Spatial coverage is expanded by the quadruped platform's ability to move across floors and terrain and reach high-altitude, narrow and dangerous areas that are difficult for humans to access. Leiphone reported that the value of embodied intelligence in industrial settings is not replacing a particular job but building a replicable, scalable unmanned inspection paradigm. Central state-owned enterprise manufacturing workshops are becoming an early field for validating that paradigm.

The competition in spatial intelligence has just begun. Leiphone reported that three conditions are maturing in the same window: falling compute costs make edge deployment feasible, embodied hardware supply is becoming sufficient, and industrial customers are shifting from watching demonstrations to evaluating output. Koala Youran currently holds an architecture-original world model base validated by an international benchmark, a three-cross mechanism for reusing the technology across multiple bodies, and a platform intended to turn the technology into a low-threshold product. The real test is whether the upper limit of cross-body generalization can remain stable in more extreme scenarios. That will determine how quickly world models move from being able to infer to being able to work and be replicated.