BYD's AI team published HyWorldVLA, scoring 90.59 PDMS on NAVSIM v1. The model uses a hybrid pixel-latent world model and VLA architecture.
Key facts
- 90.59 PDMS on NAVSIM v1 benchmark.
- Hybrid pixel-latent world model architecture.
- VLA (Vision-Language-Action) design.
- Team includes HIT robotics researchers.
- BYD's first autonomous driving foundation model.
BYD's AI team published HyWorldVLA, achieving 90.59 PDMS on the NAVSIM v1 benchmark — a new state-of-the-art in autonomous driving planning [According to Pandaily]. The model combines a hybrid pixel-latent world model with a Vision-Language-Action (VLA) architecture, marking BYD's entry into foundation models for self-driving.
Key Takeaways
- BYD's HyWorldVLA achieved 90.59 PDMS on NAVSIM v1, a new SOTA, using a hybrid pixel-latent world model.
- It marks BYD's entry into autonomous driving foundation models.
Hybrid World Model Design
HyWorldVLA uses a hybrid pixel-latent world model that predicts future driving scenes at both pixel level (for fine-grained perception) and latent level (for long-horizon reasoning). This dual representation aims to overcome the limitations of purely pixel-based or purely latent world models — the former being computationally expensive, the latter losing spatial detail. The model then conditions its action policy on these predicted future states.
State-of-the-Art on NAVSIM v1
The 90.59 PDMS (Planning Decision-Making Score) on NAVSIM v1 surpasses prior SOTA methods. NAVSIM v1 evaluates planning performance across diverse driving scenarios, including intersections, lane changes, and obstacle avoidance. BYD has not disclosed the model size, training data volume, or compute budget, making independent reproducibility difficult.
Team and Institutional Context
The HyWorldVLA team includes researchers from the Harbin Institute of Technology (HIT) robotics lab, signaling BYD's collaboration with academic robotics groups. This is BYD's first public release of an autonomous driving foundation model, positioning the company against Tesla's FSD, Waymo's Unified Agent, and Chinese competitors like Huawei's ADS and Baidu's Apollo.
Why This Matters
BYD's entry into autonomous driving foundation models is significant because the company is the world's largest EV manufacturer by volume. If BYD deploys HyWorldVLA across its production vehicles, it could rapidly scale autonomous driving data collection and real-world validation, potentially leapfrogging competitors who have smaller fleets. However, the paper does not address deployment plans, safety validation, or regulatory approvals.
What to watch
Watch for BYD's deployment timeline of HyWorldVLA in production vehicles, and whether the company publishes training details or open-sources the model. Also track NAVSIM v1 leaderboard updates for competing methods from Tesla, Waymo, Huawei, and Baidu.
Source: pandaily.com
