This project addresses the challenge of motion planning in autonomous driving by reformulating it as a language modeling problem, leveraging the reasoning capabilities of Large Language Models (LLMs). It targets researchers and engineers in autonomous driving, offering a novel approach to improve generalization in unseen scenarios.

How It Works

GPT-Driver transforms motion planning into a language modeling task by representing planner inputs and outputs as text tokens. It utilizes OpenAI's GPT-3.5 model to generate driving trajectories based on natural language descriptions of coordinate positions. A key innovation is the prompting-reasoning-finetuning strategy, designed to enhance the LLM's numerical reasoning and enable it to articulate its decision-making process.

Quick Start & Requirements

• Install dependencies: pip install -r requirements.txt
• Data preparation: Download cached nuScenes data and UniAD pretrained models from Google Drive and place them in the data directory.
• Fine-tuning: Requires an OpenAI API account and key. The gpt-driver/create_data.py script generates train.json for fine-tuning. Fine-tuning is performed via OpenAI API calls, with costs associated.
• Evaluation: Run python gpt-driver/test.py -i your_model_id -o your_output_file_name using a fine-tuned model ID.
• Official Project Page: [Project Page] (link not provided in README)

Highlighted Details

• Evaluated on the nuScenes dataset for motion planning performance.
• Demonstrates effectiveness, generalization ability, and interpretability.
• Enables LLM to describe precise trajectory coordinates and internal decision-making.

Maintenance & Community

• The project is associated with an arXiv pre-print. No specific community channels or active maintenance signals are detailed in the README.

Licensing & Compatibility

• The README does not explicitly state a license. The use of OpenAI's API implies adherence to OpenAI's terms of service and pricing.

Limitations & Caveats

• Requires an OpenAI API account and incurs costs for fine-tuning.
• The project relies on external, potentially proprietary, LLM services.
• Fine-tuning can take several hours and requires careful API key management.