On October 19, Interface News learned that WorldOracle, a world model evaluation platform jointly developed by researchers from Zhiyuan, Nvidia, and many universities, was officially launched and open for evaluation submission. The platform uses 335 actual operation trajectories, and about 40% of the interactive fragments include failed or suboptimal operations to test the model's predictive ability with actual execution results. Testing of 7 world models revealed that although some models can restore robotic arm movements more accurately, there are still large deviations in predicting the state of the object, and the long-term prediction performance also declined markedly.

On October 19, Interface News learned that WorldOracle, a world model evaluation platform jointly developed by researchers from Zhiyuan, Nvidia, and many universities, was officially launched and open for evaluation submission. The platform uses 335 actual operation trajectories, and about 40% of the interactive fragments include failed or suboptimal operations to test the model's predictive ability with actual execution results. Testing of the 7 world models revealed that although some models can restore robotic arm movements more accurately, there are still large deviations in predicting the state of the object, and the long-term prediction performance also declined markedly.