AIAny
Icon for item

Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving

Develops a vision-language foundation model for autonomous driving that unifies 3D BEV perception, visual question answering, and motion planning without changing the pretrained VLM architecture. Key elements include an external BEV perception head for 3D detection and occupancy, a Planning Expert using flow-matching for trajectory prediction, and a staged training recipe combining driving and general VLM data.

Introduction

Bridging large pretrained vision-language models (VLMs) with practical driving systems matters because VLMs encode broad visual and language knowledge, but do not natively expose explicit 3D scene structure or trajectory outputs required for autonomous driving. Qwen-Drive-1.0 tests whether a single pretrained VLM can serve as a shared backbone for 3D perception, driving VQA, and motion planning by attaching inspectable, task-specific modules rather than altering the VLM itself.

Key Findings
  • External BEV perception head: Fuses vision-encoder features and VLM outputs to produce bird's-eye-view (BEV) representations that jointly predict 3D object detection, semantic occupancy, and BEV map segmentation. Achieves competitive 3D-perception metrics (e.g., ~43.95 mAP and 60.99 mIoU on nuScenes; ~43.45 mAP and 71.27 mIoU on OpenScene).
  • Planning Expert and trajectories: Conditions on cached VLM keys/values and uses flow matching to generate future ego trajectories. Shows strong motion-planning performance (Predictive Driver Model Score 90.7 on NAVSIM; Rater Feedback Score 7.91 on Waymo Open Dataset end-to-end test split) and promising closed-loop results in simulation.
  • Staged training recipe: Jointly trains perception, language, and planning objectives by combining driving-specific supervision with general-purpose vision-language data, preserving broad instruction-following and multimodal understanding while adding driving competence.
Who it’s for and tradeoffs

Great fit if you study multimodal transfer to embodied driving tasks, want an inspectable BEV probe on top of a pretrained VLM, or aim to unify perception, VQA, and planning in one model. Look elsewhere if you need a purely optimized, standalone state-of-the-art detector/planner—Qwen-Drive prioritizes integration with a VLM backbone and research on unified representations over squeezing every last point on specialist leaderboards. The approach also depends on availability of multi-view driving data and careful dataset unification for trajectories.

Where it fits

Positions itself between pure vision-based driving stacks and end-to-end LLM-based agents: it retains the pretrained VLM's autoregressive decoder and attaches decoupled, inspectable modules (BEV head, Planning Expert), enabling both textual VQA outputs and structured trajectory predictions from shared representations.

Method highlights
  • Backbone: uses Qwen3.5-4B as the shared VLM backbone without architectural changes.
  • Perception head: explicit BEV construction that serves as a probe into 3D signals encoded by the VLM.
  • Planning: flow-matching for trajectory generation and unified waypoint representation across datasets to support joint training and evaluation.

Taken together, Qwen-Drive-1.0 is an exploratory but empirically grounded proof that pretrained VLM representations can be extended to explicit 3D perception and trajectory planning for driving while largely preserving general multimodal capabilities.

Information

  • Websitearxiv.org
  • OrganizationsHuazhong University of Science and Technology
  • AuthorsXin Zhou, Zongchuang Zhao, Zhibo Yang, Mingsheng Li, Humen Zhong, Shuai Bai, Du Chu, Ruizhe Chen, Zhaohai Li, Jun Tang
  • Published date2026/08/31

More Items

Learns generalizable World Action Models for robotic manipulation by scaling causal egocentric video pretraining and grounding learned dynamics with heterogeneous robot trajectories. Key features: a three-stage curriculum (video pretraining, video-action mid-training with a unified action representation, and target-robot specialization) and a Slow–Fast dual-system for 30 Hz real-time action prediction.

Converts posed indoor RGB(-D) video into editable, simulation-ready 3D scene graphs by parsing multi-view evidence into per-object bundles, generating complete object assets from that evidence, and placing them with GizmoAct, a VLM policy that refines 9-DoF poses through closed-loop GUI actions.

Proposes GenFirst, a generation-before-reconstruction end-to-end training strategy for latent generative models that avoids latent collapse by prioritizing generative objectives and then progressively strengthening reconstruction, validated with strong gFID/GenEval results on ImageNet-256 and text-to-image tasks.