fudan-generative-vision/WAM-Flow

What it solves

WAM-Flow addresses the challenge of accurate and safe motion planning for autonomous driving. It aims to generate future trajectories that are compliant with road rules and safety constraints while following natural-language navigation commands.

How it works

The system takes a front-view image, a natural-language navigation command (with a system prompt), and the ego-vehicle's current state as inputs. It uses a parallel coarse-to-fine motion planning approach based on discrete flow matching to output an 8-waypoint future trajectory spanning 4 seconds.

Training occurs in two stages:

1. Supervised Fine-Tuning (SFT): The model is first trained to learn accurate trajectory prediction.
2. RL Optimization: Simulator-guided Group Relative Policy Optimization (GRPO) is used to optimize closed-loop behavior. The reward function for GRPO integrates safety constraints—such as collision avoidance and drivable-area compliance—with performance goals like ego-progress, comfort, and time-to-collision.

Who it’s for

Researchers and developers working on autonomous driving, Vision-Language-Action (VLA) models, and motion planning using generative AI.

Highlights

• Parallel Denoising: Uses discrete flow matching for efficient trajectory generation.
• Cose-to-Fine Planning: Implements a parallel coarse-to-fine strategy for motion planning.
• VLA Integration: Combines visual perception, language commands, and vehicle state for action output.
• Simulator-Guided RL: Employs GRPO to align the model's behavior with safety and performance metrics.