BYD launches hybrid world model HyWorldVLA! Self-driving benchmark test scores 90.59 points, ranking highest in the world

Kuai Technology reported on August 12 that according to Gasgoo, the artificial intelligence team of BYD Automobile New Technology Research Institute published its first research paper "HyWorldVLA" on July 29, introducing a hybrid world model HyworldVLA. This model combines pixel-level and latent space world modeling and adopts a vision-language-action (VLA) architecture for autonomous driving. It achieved a record-breaking PDMS score of 90.59 in the NAVSIM v1 public benchmark test, setting a new record. It is reported that the core idea of ​​HyWorldVLA is "walking on two legs". Generally speaking, the pixel-level world model predicts future road conditions frame by frame like a movie. It is rich in details but expensive to calculate and sensitive to visual noise; the latent space model operates in a compressed abstract representation, which is efficient but easily loses key details. BYD's solution uses pixel-level supervision as a "watchdog" during the training phase to force the latent space to retain sufficient environmental information. During actual reasoning, only a lightweight latent space model is used to ensure efficiency. The ablation experiment also verified that both are indispensable. When pixel-level prediction is removed, PDMS drops from 90.59 to 87.50; when latent space processing is removed, it drops to 89.91. NAVSIM is currently one of the most watched public benchmarks in the field of autonomous driving. Its PDMS indicators do not simply compete in prediction accuracy, but comprehensively examine responsible collisions, drivable area compliance, safe distances, target completion, and driving comfort. The model cannot rely on "slowly driving" to gain points, nor can it only focus on efficiency at the expense of safety. In addition to v1 reaching the top, HyWorldVLA also scored 89.71 points on NAVSIM v2 with more comprehensive indicators, nearly 1 point ahead of the second place ExploreVLA; it scored 86.87 in 655 rain and fog noise scene tests, more than 19 points higher than the strongest baseline, showing a clear advantage. In fact, the significance of this paper goes beyond the score itself. The outside world's inherent impression of BYD has always been that it "emphasizes manufacturing and neglects R&D." Intelligentization relies more on God's Eye and supply chain integration. This is the first time BYD has made a public appearance in the field of autonomous driving basic models.