AnswersPublished15 sources
BYD’s HyWorldVLA Uses Two Views of the Road to Make Autonomous Driving AI More Efficient
BYD’s HyWorldVLA reported a 90.59 PDMS score on NAVSIM v1 by combining pixel level supervision with prediction in a compressed latent space. Its training pipeline first builds a language guided video VAE, then pre trains a hybrid world model, and finally co fine tunes it with an action expert that generates trajecto...
Published byEdited with GPT-5.6 TerraImages generated with GPT Image 2
Research answer
Key takeaways
- BYD’s HyWorldVLA reported a 90.59 PDMS score on NAVSIM v1 by combining pixel level supervision with prediction in a compressed latent space.
- Its training pipeline first builds a language guided video VAE, then pre trains a hybrid world model, and finally co fine tunes it with an action expert that generates trajectories.
- Removing pixel prediction lowered PDMS to 87.50, while removing latent space processing reduced it to 89.91—evidence that the two signals serve complementary roles.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How did BYD’s newly revealed AI team develop HyWorldVLA—a Vision-Language-Action autonomous-driving foundation model that combines pixel-lev. Article summary: BYD’s first public autonomous-driving foundation-model paper presents HyWorldVLA as a hybrid VLA system: it uses pixel-level reconstruction to preserve scene detail and latent-space prediction to make planning more robus. Topic tags: general, academic, general web, government, education. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermark
BYD has introduced HyWorldVLA, a Vision-Language-Action (VLA) model for autonomous driving built around a hybrid world model. Its central idea is straightforward: use two representations of the road at once—high-detail pixel-level predictions for visual grounding and compressed latent representations for more efficient forecasting and trajectory planning. 1
The trade-off HyWorldVLA is designed to address
Predicting future driving scenes pixel by pixel can preserve fine visual detail, but it is computationally expensive. A latent-space model, by contrast, predicts within a compressed representation of the scene. That makes it more practical to run, but it can discard information that matters for driving decisions. 1
HyWorldVLA is designed not to choose one method at the expense of the other. During training, pixel-level supervision keeps the latent representation anchored to the real visual scene. During operation, however, the model predicts only latent features and sends them to an action expert that produces driving trajectories. 1
7
The three-stage training pipeline
1. Training a language-guided video VAE
First, a video variational autoencoder, or VAE, compresses sequences of driving video into latent features. Language is used as semantic context while this video compressor is trained. 2
5
The goal is not simply to shrink video data. The latent space must retain information useful for forecasting what happens next—and ultimately for making driving plans.
2. Hybrid world-model pre-training
Next, images, actions, language inputs and latent features are converted into unified discrete tokens for autoregressive prediction. At this stage, the model predicts future video latents while also reconstructing future video frames. 1
2
That creates a form of dual supervision:
- Pixel prediction and reconstruction preserve detailed visual grounding.
- Latent prediction encourages a compact signal that is useful for planning ahead.
Put simply, the pixel loss acts as a constraint against over-compressing the scene and losing important details in the latent representation. 1
7
3. Co-fine-tuning with an action expert
In the final stage, BYD jointly optimizes the world model and a specialist action—or trajectory—model. At runtime, the system predicts latent features rather than generating full pixel-level frames; those features are passed to the action expert, which generates driving trajectories. 1
This separation is central to the efficiency argument. The system benefits from costly pixel-level supervision during training without paying the same cost during online inference.
What the ablation results show
On NAVSIM v1, HyWorldVLA reported 90.59 PDMS, described in the paper and related reports as the leading result among public results at the time. 1
7
The ablation experiments suggest that the hybrid design is more than an added feature set:
- Without pixel-level prediction, PDMS fell from 90.59 to 87.50.
7
- Without latent-space processing, it dropped to 89.91.
- Without language guidance in the latent module, the score declined to 90.35.
20
These results support the researchers’ premise that pixel supervision and latent prediction solve different parts of the problem. The former preserves fidelity to the road scene, while the latter provides a more economical representation for predicting actions.
Why loss-weight tuning matters
The paper tests two weights: λ1 for video-token prediction loss and λ2 for latent-representation consistency. 20
With λ2 fixed at 0.1, increasing λ1 from 0.1 to 0.5 improved PDMS from 90.48 to 90.59. Raising λ1 further, to 1.0, reduced the score to 89.83. 20
With λ1 fixed at 0.5, increasing λ2 above 0.1 also hurt performance: at 0.2, PDMS was 90.47, and it continued to decline as the weight rose further. 20
The takeaway is not that more supervision is always better. The model needs enough pixel and latent constraints to retain useful information, but not so much that its representation becomes overly rigid.
What PDMS measures in NAVSIM v1
PDMS is a composite driving-quality metric used in NAVSIM. It includes factors such as avoiding at-fault collisions, complying with the drivable area, time to a potential collision, driving comfort and progress along the route. 2
That means the 90.59 score is not a simple image-recognition result. It reflects driving-planning performance within this particular benchmark. But it remains an evaluation in a simulated setting, not independent confirmation of safety or performance in real road traffic.
What the paper says about BYD’s AI strategy
The work is attributed to BYD’s Automotive New Technology Research Institute, making it a public sign of the company’s investment in in-house foundation-model research for autonomous driving. 1 Reports linking the team to robotics expertise at the Harbin Institute of Technology are notable, but the available primary evidence does not establish the academic background of every individual researcher.
5
Against Chinese EV rivals including NIO, XPeng and Li Auto, HyWorldVLA gives BYD a public, measurable research result in VLA systems and world modeling. It demonstrates research capability, rather than proving that BYD already holds an on-road autonomous-driving advantage.
BYD’s vehicle-development scale could become an advantage only if the company can turn this research result into a safe, validated and efficiently deployed product. The next meaningful test is not another benchmark alone, but a reliable transition from model and simulation to real-world driving.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "BYD’s HyWorldVLA Uses Two Views of the Road to Make Autonomous Driving AI More Efficient"?
BYD’s HyWorldVLA reported a 90.59 PDMS score on NAVSIM v1 by combining pixel level supervision with prediction in a compressed latent space.
What are the key points to validate first?
BYD’s HyWorldVLA reported a 90.59 PDMS score on NAVSIM v1 by combining pixel level supervision with prediction in a compressed latent space. Its training pipeline first builds a language guided video VAE, then pre trains a hybrid world model, and finally co fine tunes it with an action expert that generates trajectories.
What should I do next in practice?
Removing pixel prediction lowered PDMS to 87.50, while removing latent space processing reduced it to 89.91—evidence that the two signals serve complementary roles.
Sources
car.tom.com比亚迪推出混合世界模型HyWorldVLA
eu.36kr.comBYD Unveils Its AI Team for the First Time: Boasting Top Harbin Institute of Technology Robotics Technical Genes, Debut Large Model Hits SOTA Level