BYD's AI Team Revealed for the First Time, Harbin Institute of Technology's Robotics Genes Shine, Large Model Achieves SOTA on Debut

BYD's AI Ambition Takes Center Stage

For years, BYD's prowess in intelligent vehicles was largely associated with its robust supply chain integration, large-scale deployment, and hardware capabilities, often leaving its "code-level self-developed" AI technology and academic contributions in the shadows. The recent unveiling of HyWorldVLA, a multimodal foundational model from BYD Automotive New Technology Research Institute, marks a significant departure from this perception. This paper, achieving state-of-the-art (SOTA) results on the NAVSIM v1 autonomous driving benchmark, signals BYD's serious commitment to foundational physical AI research, a domain previously thought to be dominated by tech giants like Tesla and Waymo.

The HyWorldVLA model specifically targets the cutting-edge Vision-Language-Action (VLA) and World Model paradigms in autonomous driving. This approach aims to move AI beyond mere road recognition towards a deeper "understanding of the world," enabling it to predict future scenarios and make informed driving decisions. This public disclosure not only showcases BYD's advanced research capabilities but also challenges the narrative that its intelligence strategy was primarily focused on integration rather than deep, proprietary AI development.

HyWorldVLA: A Hybrid Approach to World Models

The core innovation of HyWorldVLA lies in its hybrid world modeling approach, designed to resolve the inherent tension between understanding real-world details and achieving efficient inference in autonomous driving. Traditional methods often struggle, with pixel-level models offering detail but high computational cost, and latent world models providing efficiency but potentially losing crucial real-world nuances. BYD's solution integrates both, leveraging a three-step training process.

Initially, a video compressor transforms continuous video frames into compact latent features, guided by textual information to enhance reconstruction quality. This is followed by a pre-training phase where the model learns to predict future frames and latent features simultaneously, ensuring the latent space retains sufficient information. Finally, a joint fine-tuning stage focuses solely on outputting latent features, which are then used by an action generation module to produce driving trajectories. This phased approach, validated by ablation studies, demonstrates that the hybrid model's superior performance, particularly in challenging conditions like rain and fog, stems from its ability to "feed" the latent space with pixel-level supervision during pre-training, then switch to efficient latent-space inference during fine-tuning, thereby gaining robustness against visual noise.

The Harbin Institute of Technology Robotics Pedigree

The independent development of HyWorldVLA by BYD Automotive New Technology Research Institute, without external academic co-authorship, is noteworthy. A deeper look into the team behind this and other recent BYD AI papers, such as EMoE-Planner (published in 2025 according to the article), reveals a distinct "robotics AI faction" within the company. Key figures like Liulong Ma, a corresponding author on multiple BYD AI papers, and Hongbiao Zhu, the first author of EMoE-Planner, share strong academic backgrounds from the Harbin Institute of Technology (HIT) and Carnegie Mellon University (CMU).

HIT, particularly its State Key Laboratory of Robotics and System, is renowned for its robotics research. This strong academic lineage in robotics, rather than traditional automotive electronics or ADAS engineering, suggests a strategic recruitment of top-tier computer science and robotics talent. This organizational philosophy aligns closely with Tesla's integrated "robotics + autonomous driving" model, where the underlying principles of physical AI for robots are directly applied to autonomous vehicles. This foundational approach, rooted in first-principles thinking for physical AI, positions BYD to compete at the forefront of the evolving autonomous driving landscape, where the focus is shifting from mere feature stacking to comprehensive physical AI capabilities.