Generative Infrastructure for
Autonomous Driving

We are redefining how autonomous systems learn by creating infinite, high-fidelity synthetic worlds and unified multimodal brains.

Core Pillar I

Generative World Models

We use Generative AI to construct a data feedback loop framework. This creates large-scale simulation scenes including complex high-level rules (e.g. police gestures) and complete 3D ground truth.

  • ✓ Static World Construction (BEVControl)
  • ✓ Dynamic Video Synthesis (Unleashing)
  • ✓ Closed-loop Self-Correction
Generative World Simulation

Scroll horizontally to view full workflow

Selected Research

BEVControl

ICCV 2023

Accurate geometric generation of 3D scenes from bird's-eye view sketches, enabling controllable autonomous driving simulation.

OmniGen

Generative AI

Unified generation of multimodal sensor data through shared BEV space, ensuring consistency across cameras and LiDARs.

DriveMRP

Safety Prediction

Predicting potential risks of planned trajectories using large language models enhanced by synthetic data.

DualToken

Visual Tokenizer

A unified visual tokenizer for both understanding and generation, achieving state-of-the-art performance in MLLM tasks.