MapAgent: An Industrial-Grade Agentic Framework for City-scale Lane-level Map Generation

The paper figures and bad cases are put in the paper folder.

This repository contains the official implementation of:

MapAgent: An Industrial-Grade Agentic Framework for City-scale Lane-level Map Generation

---

📋 Overview

MapAgent is an agentic refinement framework built on top of a frozen BEV (Bird's Eye View) vectorization backbone. It introduces a bounded Judge–Planner–Worker loop to verify mapping specifications and perform deterministic, tool-grounded edits for lane-level map refinement in autonomous driving systems.

Key Features

• 🎯 Vision-Language Judge Agent for automated quality assessment of lane predictions
• 🔄 Bounded Refinement Loop with deterministic worker tools
• 🚗 Industrial-Grade design for city-scale deployment
• 🧠 GRPO-based reinforcement learning for model optimization
• 📊 Comprehensive error taxonomy and structured reasoning

---

🗂️ Repository Structure

./
├── data_example/          # Training data and examples
│   ├── dataset/
│   │   ├── map_agent_dataset.jsonl
│   │   └── images/        # Lane prediction visualization
│   └── prompt/
│       └── train_example.txt
├── EasyR1/                # GRPO reinforcement learning
│   ├── generate_grpo_data.py
│   ├── grpo_config.yaml
│   ├── grpo.sh
│   ├── format_prompt/
│   │   └── map.jinja
│   └── reward/
│       └── map_reward.py
└── LlamaFactory/          # Model training configurations
    ├── inference/
    │   └── infer_qwen_lora_sft.yaml
    ├── merge/
    │   └── merge_qwen_lora_sft.yaml
    └── train/
        └── train_qwen_lora_sft.yaml

---

📦 Components

1. Data Example (data_example/)

Contains training datasets and prompt examples for the vision-language judge agent.

• Dataset Structure:

  • map_agent_dataset.jsonl: Annotated lane prediction examples with error labels
  • images/: Visualization of predicted lane lines overlaid on road images
  • prompt/train_example.txt: Complete prompt template with reasoning examples

• Error Taxonomy:

  • extra_lane_line: Predicted line on non-existent markings
  • category_error: Incorrect lane category classification
  • geometry_error: Local geometric defects (kinks, spikes)
  • structure_error: Major structural failures (wrong direction, large deviations)
  • no_error: Correct prediction

• Ground Truth Categories:

  1: Long dashed line      6: Short dashed line
  2: Double solid line     7: Double dashed line
  3: Single solid line     9: Virtual line
  4: Dashed-solid line    10: Stop line
  5: Solid-dashed line
  

2. GRPO Training (EasyR1/)

Group Relative Policy Optimization (GRPO) for reinforcement learning-based model refinement.

• Core Scripts:

  • generate_grpo_data.py: Generate training pairs for GRPO
  • grpo.sh: Training pipeline execution script
  • grpo_config.yaml: Hyperparameters and training configuration

• Components:

  • format_prompt/map.jinja: Jinja2 template for prompt formatting
  • reward/map_reward.py: Reward function for RL training

• Purpose: Fine-tune the judge agent to improve error detection accuracy through reinforcement learning.

3. Model Training (LlamaFactory/)

Configuration files for supervised fine-tuning (SFT) and model deployment using LLaMA-Factory.

• Training (train/):

  • train_qwen_lora_sft.yaml: LoRA-based SFT configuration for Qwen models

• Inference (inference/):

  • infer_qwen_lora_sft.yaml: Inference configuration with LoRA adapters

• Merge (merge/):

  • merge_qwen_lora_sft.yaml: Configuration for merging LoRA weights into base model

• Supported Models: Qwen-VL, Qwen2-VL with LoRA parameter-efficient fine-tuning

---

🚀 Quick Start

Prerequisites

# Clone the repository
git clone https://github.com/eadst/KDD-2026-MapAgent.git

Training the Judge Agent

Option 1: Supervised Fine-Tuning (SFT)

# Using LlamaFactory
cd LlamaFactory/train
llamafactory-cli train train_qwen_lora_sft.yaml

Option 2: GRPO Reinforcement Learning

# Generate GRPO training data
cd EasyR1
python generate_grpo_data.py

# Run GRPO training
bash grpo.sh

Inference

# Inference with trained model
cd LlamaFactory/inference
llamafactory-cli chat infer_qwen_lora_sft.yaml

Model Merging

# Merge LoRA weights into base model
cd LlamaFactory/merge
llamafactory-cli export merge_qwen_lora_sft.yaml

---

📊 Dataset Format

The training data follows a structured format:

{
  "image": "path/to/image.png",
  "predicted_category": "double_solid_line",
  "bounding_box": [0.0, 1473.0, 1536.0, 11.0],
  "reasoning": "<think>Analysis process...</think>",
  "error_type": "no_error",
  "gt_category_id": 2
}

---

🎯 Usage Example

# Example: Evaluate a lane prediction
from mapagent import JudgeAgent

# Initialize judge agent
judge = JudgeAgent(model_path="path/to/model")

# Evaluate prediction
result = judge.evaluate(
    image_path="data_example/dataset/images/557_5950_pred.png",
    predicted_category="double_solid_line",
    bounding_box=[0.0, 1473.0, 1536.0, 11.0]
)

print(f"Error Type: {result['error_type']}")
print(f"Reasoning: {result['reasoning']}")
print(f"GT Category: {result['gt_category_id']}")

---

📈 Evaluation Protocol

The judge agent follows a strict priority-based evaluation:

1. Priority Order: extra_lane_line → category_error → geometry_error → structure_error → no_error
2. Short-Circuit Logic: Stop evaluation once an error is confirmed
3. Visual Decoding Strategy: Distinguish between prediction mask (hypothesis) and road texture (fact)
4. Structured Reasoning: Output reasoning process within <think> tags

---

🔧 Configuration

GRPO Training Parameters

# grpo_config.yaml
model_name: Qwen2-VL-7B-Instruct
learning_rate: 5.0e-6
num_train_epochs: 2
per_device_train_batch_size: 2
reward_model: map_reward

SFT Training Parameters

# train_qwen_lora_sft.yaml
model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
lora_rank: 8
lora_alpha: 16
lora_dropout: 0.05
learning_rate: 5.0e-5

---

📝 Citation

If you find this work useful, please cite our paper:

@article{mapagent2025,
  title={MapAgent: An Industrial-Grade Agentic Framework for City-scale Lane-level Map Generation},
  author={Author Names},
  journal={Conference/Journal Name},
  year={2026}
}

---

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

---

🙏 Acknowledgments

• LLaMA-Factory for training infrastructure
• Qwen-VL for vision-language foundation models
• EasyR1 framework for GRPO implementation

---

📧 Contact

For questions and discussions, please open an issue:

• Project Page: https://github.com/eadst/KDD-2026-MapAgent

---

🚧 Status

Current Version: v0.1.0 (Initial Release)

• ✅ Data examples and prompt templates
• ✅ GRPO training pipeline
• ✅ LlamaFactory configurations
• 🚧 Full pipeline integration (coming soon)
• 🚧 Pre-trained model weights (coming soon)
• 🚧 Evaluation scripts (coming soon)

---

🗺️ Roadmap

• Release pre-trained judge agent weights
• Add full pipeline integration code
• Provide evaluation benchmarks
• Add worker tools implementation
• Release technical report
• Add interactive demo

---

Last Updated: February 17, 2026