BYD's HyWorldVLA sets new record in autonomous driving benchmark

BYD's AI team published its first research paper on HyWorldVLA, a hybrid world model for autonomous driving. It combines pixel-level and latent space world modeling with a vision-language-action (VLA) architecture. On the NAVSIM v1 benchmark, it scored 90.59 PDMS, setting a new record, and also led NAVSIM v2 with 89.71.

The model uses a two-pronged approach: pixel-level supervision during training to preserve details, and a lightweight latent model during inference for efficiency. Ablation tests showed that removing either component reduces performance. In rainy and foggy scenarios, it outperformed the strongest baseline by over 19 points.

Additionally, BYD launched China's first 4nm automotive-grade chip, Xuanji A3, with 700 TOPS per chip and support for three-chip parallel, reaching over 2100 TOPS in total. This algorithm-to-chip integration is seen as key preparation for L3 autonomous driving.