难得啊，比亚迪也开始卷世界模型和VLA了。
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR - BYD has published HyWorldVLA, a vision-language-action driving model that pairs pixel-level future-video prediction with latent world modeling, signaling the volume-leader carmaker is now competing publicly on frontier autonomous-driving algorithms. It matters because it shows a hybrid supervision recipe that markedly improves robustness under rain/fog conditions rather than just topping clean benchmarks.
- Hybrid world modeling: A text-guided Video VAE (spatial + temporal compression, scene captions injected via cross-attention) produces driving-aware latents; during pretraining the model predicts language, action, future visual tokens, and future VAE latents together over an Emu3 backbone, so dense pixel supervision "calibrates" the latents.
- Latent query: a dedicated internal slot aggregates history frames, ego motion, and navigation intent into a compact summary of the next few seconds — the same future is learned twice (fine-grained visual tokens vs. stable high-level latent).
- Inference doesn't generate video: at co-fine-tuning/deployment the action expert consumes VLM context, ego history, navigation, and predicted latents only; the method was plugged into both trajectory-vocabulary selection and flow-matching continuous action heads, so it isn't tied to one planner.
- Results: 90.59 PDMS on NAVSIM v1 and 89.71 EPDMS on NAVSIM v2 (incremental gains), but on 655 rain/fog noisy OpenScene cases it reached 86.87 PDMS vs. WoTE 60.65, DriveLaW 67.49, and DriveVLA-W0 61.18; qualitatively it holds a stop at blurry red lights where baselines creep forward.
Sources (1)
难得啊，比亚迪也开始卷世界模型和VLA了。
TL;DR - BYD has published HyWorldVLA, a vision-language-action driving model that pairs pixel-level future-video prediction with latent world modeling, signaling the volume-leader carmaker is now competing publicly on frontier autonomous-driving algorithms. It matters because it shows a hybrid supervision recipe that markedly improves robustness under rain/fog conditions rather than just topping clean benchmarks.
- Hybrid world modeling: A text-guided Video VAE (spatial + temporal compression, scene captions injected via cross-attention) produces driving-aware latents; during pretraining the model predicts language, action, future visual tokens, and future VAE latents together over an Emu3 backbone, so dense pixel supervision "calibrates" the latents.
- Latent query: a dedicated internal slot aggregates history frames, ego motion, and navigation intent into a compact summary of the next few seconds — the same future is learned twice (fine-grained visual tokens vs. stable high-level latent).
- Inference doesn't generate video: at co-fine-tuning/deployment the action expert consumes VLM context, ego history, navigation, and predicted latents only; the method was plugged into both trajectory-vocabulary selection and flow-matching continuous action heads, so it isn't tied to one planner.
- Results: 90.59 PDMS on NAVSIM v1 and 89.71 EPDMS on NAVSIM v2 (incremental gains), but on 655 rain/fog noisy OpenScene cases it reached 86.87 PDMS vs. WoTE 60.65, DriveLaW 67.49, and DriveVLA-W0 61.18; qualitatively it holds a stop at blurry red lights where baselines creep forward.