Generative World Models

Generative World Model for 7V Closed-Loop Driving Simulation

A real-time streaming 7-camera world model powered by causal DiT architecture, 3-stage training paradigm, and interactive vector-to-neural closed-loop simulation.

Timeline
2025.06–Present
Context
Frontier World Model Lab
Role
Core Algorithm & System Architect
Stage
Production & Research

Overview

What is this project about?

Problem

Real-world test drives cannot efficiently scale to rare long-tail hazards, while classical graphics simulators suffer from severe sim-to-real domain gaps. Existing surround video diffusion models are non-causal and require 30+ iterative denoising steps, making real-time closed-loop policy evaluation impossible.

What I built

A real-time streaming 7V surround world model with block-causal spatio-temporal attention, cross-view epipolar conditioning, and a 3-stage training pipeline (Bidirectional Teacher → Causal Student → DMD Self-Forcing Distillation) that generates photorealistic 7-camera video streams from interactive BEV vector layouts.

Result

Achieved continuous streaming rollout with only 4 denoising steps per chunk, supporting interactive closed-loop simulation across highway merging, dense night traffic, cutting-in maneuvers, and controllable weather synthesis (dusk, rain, storm, snow).

My role. Designed the 7V streaming DiT network architecture, formulated the cross-view causal attention mechanism, and developed the end-to-end 3-stage distillation and closed-loop vector-to-video rollout platform.

research world-model generative e2e deployment

01 · System Architecture & Methodology

Streaming 7V Generative Simulation Architecture

A complete closed-loop neural simulation framework translating high-level policy actions into photorealistic, spatially and temporally coherent 7-camera surround video streams.

7V Closed-Loop Simulation System Workflow
Figure 1: End-to-end 7V closed-loop simulation pipeline. From policy action & BEV vector dynamics, through multi-camera layout projection, to streaming neural world model rendering and closed-loop perception-action feedback.
Model Architecture

Streaming World Model DiT Block

Streaming World Model DiT Architecture

28-layer DiT block with AdaLN modulation, intra-view temporal attention with KV-caching, localized cross-view spatial attention, and cross-attention text/vector conditioning.

Training Paradigm

Three-Stage Progressive Training

Three-Stage Training Paradigm

Stage 1 (L1b): Bidirectional Teacher pretraining.
Stage 2 (L2a): Causal Student conversion with block-causal mask.
Stage 3 (L3): Self-Forcing DMD distillation (35 steps → 4 steps).

02 · Model Evolution & Data Engine

Teacher Denoising & Causal Student Conversion

Comparing full-window bidirectional teacher generation against streaming causal student rollout and rare-case safety scenario synthesis.

Stage 1 · Bidirectional Teacher

Surround Generation Walkthrough — Case 1

Teacher Case 1 (93 frames @ 10Hz). Top: Condition Layout | Mid: World Model Output | Bottom: Ground Truth.
Stage 1 · Bidirectional Teacher

Surround Generation Walkthrough — Case 2

Teacher Case 2 (93 frames @ 10Hz). Top: Condition Layout | Mid: World Model Output | Bottom: Ground Truth.
Smart Agent · Long-tail Generation

Rare Interactive Safety Case — Example 1

Synthetic long-tail traffic injection: aggressive cut-in and emergency braking behavior generation.
Smart Agent · Long-tail Generation

Rare Interactive Safety Case — Example 2

High-density interaction and corner-case scenario synthesis for stress-testing planner policies.
Stage 2 · Causal Student

Causal Block Rollout — Example 1

Streaming causal generation with KV cache (2-frame block autoregression).
Stage 2 · Causal Student

Causal Block Rollout — Example 2

Long-horizon temporal consistency without future-frame leakage.

03 · Stage 3 Self-Forcing & Interactive Rollout

Streaming Closed-Loop Traffic Scenarios

Final self-forcing distilled model executing continuous real-time neural rollouts across interactive driving maneuvers in complex traffic environments.

Vector layout interaction & 7V camera projection. Interactive policy maneuvers in BEV vector space mapped dynamically to 7 surround-view camera layout conditions.
Closed-Loop Scenario 01

Daytime Highway Overtaking & Lane Change

High-speed interactive overtaking with continuous cross-camera object tracking.
Closed-Loop Scenario 02

Nighttime Congested Highway Traffic

Challenging low-light glare, tail-light reflections, and multi-agent interaction.
Closed-Loop Scenario 03

Highway Confluence & Ramp Merging

Multi-lane ramp merging scenario rendered in real-time streaming mode.
Benchmark Validation

Dusk Generation vs Ground Truth Benchmark

Top: Input Layout | Mid: World Model Output | Bottom: Ground Truth Reference.

04 · Zero-Shot Weather Transfer

Multi-Weather Controllable Generation

Demonstrating zero-shot style transfer on the identical trajectory layout across harsh weather conditions to test perception robustness.

Condition A · Atmospheric Lighting

Dusk / Sunset Simulation

Low-angle golden sunlight and long shadow rendering.
Condition B · Precipitation

Heavy Rain Storm Simulation

Wet road specular reflections and windscreen droplet distortion.
Condition C · Complex Lighting + Water

Night Rain Weather Simulation

Complex headlight bloom and wet asphalt light reflections.
Condition D · Extreme Winter

Blizzard Snow Weather Simulation

Heavy snowfall occlusion and whiteout surface texture synthesis.
7 CamerasSynchronized 360° surround video
4 StepsSelf-forcing distilled real-time rollout
Causal DiTStreaming block autoregression + KV cache
Multi-WeatherDay, night, dusk, rain, storm, snow
Confidentiality note. All displayed scenarios, models, and evaluation benchmarks are based on sanitized open research configurations. Proprietary commercial calibration and confidential fleet telemetry are strictly excluded.