BYD Publishes Autonomous Driving Paper HyWorldVLA, Achieves SOTA on NAVSIM v1 with PDMS 90.59

BYD's Automotive New Technology Institute has publicly released its autonomous driving foundation model HyWorldVLA, which achieved a PDMS score of 90.59 on the public autonomous driving benchmark NAVSIM v1, reaching a leading level among current public results. HyWorldVLA adopts a hybrid world modeling and Vision-Language-Action (VLA) architecture, aiming to combine the environmental detail understanding capability of pixel-level world models with the reasoning efficiency of latent space models. The model training is divided into three steps: first, train a video compressor to compress continuous video frames into latent features; second, pre-train the model to simultaneously predict future frame tokens and latent features; finally, joint fine-tuning, outputting only latent features and generating driving trajectories. Ablation experiments show that the hybrid strategy contributes the most to performance; removing frame prediction reduces PDMS to 87.50, and removing latent space reduces it to 89.91.

In rain-fog scenario tests, HyWorldVLA's PDMS reached 86.87, significantly outperforming pure pixel methods.

HyWorldVLA is BYD's first multimodal foundation model paper, completed entirely independently by BYD's Automotive New Technology Institute with no external co-authors. The team has previously published autonomous driving-related papers such as EMoE-Planner (2025). Team members' backgrounds are highly concentrated at Harbin Institute of Technology and Carnegie Mellon University, with Liulong Ma serving as corresponding author on multiple papers, whose research trajectory covers autonomous driving planning, robot navigation, multimodal perception, and world models. Hongbiao Zhu is the first author of EMoE-Planner, also with HIT and CMU background. This indicates that BYD's New Technology Institute AI team is not an extension of automotive electronics or ADAS engineering, but has absorbed top computer and robotics talent, forming a "robotics AI school" research approach.

BYD's self-developed AI team timeline points to a start in 2024 or earlier, with EMoE-Planner in 2025 and HyWorldVLA in 2026, during which multiple collaborative papers were published at top conferences such as ICLR, CVPR, and NeurIPS. The team composition shows distinct characteristics: members' backgrounds are highly concentrated at HIT and CMU, forming a technical echo with BYD's concurrently advancing self-developed robotics projects. The VLA plus world model approach represented by HyWorldVLA is essentially an autonomous driving version of the robotics "perception-reasoning-action" loop, sharing the same technology stack and talent gene. This organizational logic is similar to Tesla AI's "robotics plus autonomous driving" integrated architecture, reflecting the first-principles approach of physical AI starting from technology.