Skip to main content
Models & Technology

BYD AI Team Makes Debut: Hybrid World Model HyWorldVLA Scores 90.59, Completing the In-House Chip-and-Algorithm Loop

BYD’s new HyWorldVLA model has achieved a record-high score in the NAVSIM test. Together with the previously released in-house 4nm automotive-grade intelligent driving chip, BYD has completed a closed loop spanning intelligent driving algorithms and chips. #BYDAutonomousDriving#

BYD AI Team Makes Debut: Hybrid World Model HyWorldVLA Scores 90.59, Completing the In-House Chip-and-Algorithm Loop

BYD’s Automotive New Technology Research Institute AI team published its first research paper last month, introducing HyWorldVLA, a hybrid world model for autonomous driving. The model combines pixel-level and latent-space world modeling with a vision-language-action (VLA) architecture.

BYD AI Team Makes Debut: Hybrid World Model HyWorldVLA Scores 90.59, Completing the In-House Chip-and-Algorithm Loop

The model achieved a 90.59 PDMS score on the public autonomous driving benchmark NAVSIM v1, setting a new record for the benchmark. On the more comprehensive NAVSIM v2, HyWorldVLA also achieved a leading score of 89.71.

Note: The PDMS metric comprehensively evaluates real-world driving quality across multiple dimensions, including collision safety, compliance with drivable areas, safe distances, goal completion, and ride comfort.

HyWorldVLA’s core approach is to “walk on two legs.” The pixel-level world model predicts future road conditions frame by frame, offering rich detail but requiring substantial computation and being sensitive to visual noise. The latent-space model operates on compressed abstract representations, delivering high efficiency but potentially losing key details. BYD’s approach uses pixel-level supervision as a “watchdog” during training, forcing the latent space to retain sufficient environmental information, while using only the lightweight latent-space model during inference to ensure efficiency.

Ablation experiments verified that both components are indispensable: removing pixel-level prediction reduced the PDMS from 90.59 to 87.50, while removing latent-space processing lowered it to 89.91. In tests involving 655 rainy and foggy noise scenarios, HyWorldVLA scored 86.87, more than 19 points higher than the strongest baseline.

On May 28 this year, BYD released China’s first 4nm automotive-grade intelligent driving chip, the “Xuanji A3.” The chip uses a 4nm process, integrates a 16-core CPU, delivers 420K DMIPS and 273GB/s of bandwidth, and reaches the highest ASIL-D functional safety level. A single chip provides approximately 700TOPS of computing power and supports parallel operation of three chips, allowing the intelligent driving system in an entire vehicle to reach more than 2100TOPS. This complete loop from algorithms to chips is regarded as a key technological reserve for BYD’s development of L3 autonomous driving.