Efficient VLA Models
Parallel generation, block diffusion, discrete diffusion, and efficient decoding for Vision-Language-Action models.
ZZ
Academic Homepage
Zhihao Zhu / 朱志浩
Master Student, Fudan University
I work on efficient multimodal generative models for autonomous driving, with a focus on Vision-Language-Action models, diffusion/flow-based planning, and AR-to-diffusion distillation.
Vision-Language-ActionAutonomous DrivingDiffusion ModelsFlow MatchingReinforcement LearningMotion Planning
About Me
Efficient generative models for autonomous driving.
I am a master's student at Fudan University, working on multimodal generative models and autonomous driving. My research focuses on efficient Vision-Language-Action models, especially diffusion- and flow-based methods for motion planning.
Recently, I have been exploring how to convert autoregressive VLAs into parallel generative models through block diffusion, hierarchical distillation, and reinforcement learning. I am broadly interested in efficient multimodal reasoning, embodied intelligence, and safety-critical decision making.
News
Recent Updates
- WAM-Flow was accepted to CVPR 2026.
- WAM-Flow pretrained models and evaluation resources were released.
- WAM-Diff and WAM-Flow were released as preprints.
- Started exploring efficient diffusion-based VLA models for autonomous driving.
Research Interests
Efficient Generative Vision-Language-Action Models for Autonomous Driving
My current work centers on efficient VLA planning systems, generative trajectory modeling, and reliable decision making for autonomous driving.
Autonomous Driving Planning
Trajectory-as-language, closed-loop planning, NAVSIM / Bench2Drive evaluation, and safety-critical decision making.
AR-to-Diffusion Distillation
Progressive block-wise adaptation, block-wise distillation, and cross-scale distillation from autoregressive VLAs to diffusion models.
Reinforcement Learning for Planning
GRPO, PDMS reward optimization, simulator-guided feedback, and safe trajectory alignment.
Selected Publications
Representative papers and ongoing research
Teaser coming soon
WAM-Flow: Parallel Coarse-to-Fine Motion Planning via Discrete Flow Matching for Autonomous Driving
A discrete flow matching framework for parallel coarse-to-fine motion planning in autonomous driving, enabling efficient bidirectional trajectory refinement with simulator-guided alignment.
Teaser coming soon
WAM-Diff: A Masked Diffusion VLA Framework with MoE and Online Reinforcement Learning for Autonomous Driving
A masked diffusion Vision-Language-Action framework for autonomous driving, integrating non-causal trajectory decoding, MoE scaling, and online reinforcement learning for closed-loop planning.
Teaser coming soon
Hierarchical AR-to-Diffusion Distillation for Efficient Autonomous Driving VLAs
A research direction on transforming pretrained autoregressive VLAs into efficient block diffusion models through progressive block-wise adaptation, block-wise distillation, and cross-scale model distillation.
Selected Projects
Research code and project materials
WAM-Flow
Parallel coarse-to-fine motion planning via discrete flow matching.
A VLA-based planning framework that casts future trajectory generation as discrete flow matching over structured trajectory tokens.
WAM-Diff
Masked diffusion VLA framework for autonomous driving.
A diffusion-based VLA framework that iteratively refines discrete future trajectory tokens using masked denoising and reinforcement learning.
AR-to-Diffusion Distillation
Turning autoregressive VLAs into efficient parallel generative models.
A research line on progressive block-wise adaptation, block-wise teacher-student distillation, and cross-scale diffusion model transfer.
Experience & Education
Academic background
2025 - Present
Master Student
Fudan University · Shanghai, China
Research on multimodal generative models, Vision-Language-Action models, and autonomous driving.
2019 - 2023
B.Eng. / B.S. in Information Management and Information Systems
South China University of Technology · Guangzhou, China
Undergraduate study in the School of Mathematics.
Contact
Research discussions and collaborations
I am always open to research discussions and collaborations on efficient multimodal generative models, autonomous driving, and embodied intelligence.