Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving

CVPR 2024


Multiview visual forecasting and planning by world model. At time step \( T \), the world model imagines the multiple futures at \(T+K\), and finds it is safe to keep going straight at \(T\). Then the model realizes that the ego car will be too close to the front car according to the imagination of time step \(T + 2K\), so it decides to change to the left lane for a safe overtaking.


In autonomous driving, predicting future events in advance and evaluating the foreseeable risks empowers autonomous vehicles to better plan their actions, enhancing safety and efficiency on the road. To this end, we propose Drive-WM, the first driving world model compatible with existing end-to-end planning models. Through a joint spatial-temporal modeling facilitated by view factorization, our model is the first to generate high-fidelity multiview videos in driving scenes. Building on its powerful generation ability, we showcase the potential of applying the world model for safe driving planning for the first time. Particularly, our Drive-WM enables driving into multiple futures based on distinct driving maneuvers, and determines the optimal trajectory according to the image-based rewards. Evaluation on real-world driving datasets verifies that our method could generate high-quality, consistent, and controllable multiview videos, opening up possibilities for real-world simulations and safe planning.

Method Overview

Overview of the proposed framework. (a) illustrates the training and inference pipeline of the proposed method. (b) visualizes the unified conditions leveraged to control the generation of multi-view video. (c) represents the probabilistic graph of factorized multiview generation. It takes the 3-view output from (a) as input to generate other views, enhancing the multi-view consistency.

Generation of Diverse Multiview Videos

Generation of Multiple Plausible Futures

Scene 1 (Turn left)

Scene 1 (Turn right)

Scene 2 (Turn left)

Scene 2 (Go straight)

Scene 3 (Turn left)

Scene 3 (Go straight)

Visual Element Control

Light Changes

Weather Changes

Control Foreground Object Layouts

Action Condition

Other Datasets

World Model for End-to-End Planning

End-to-end planning pipeline with our world model. We display the components of our planning pipeline at the top and illustrate the decision-making process in the planning tree using image-based rewards at the bottom.

GPT-4V as Reward Function for Non-vectorized Observations