Practical overview
When it fits
Use Alphadrive when you need a research codebase for staged training and evaluation of vision-language-model planning in autonomous-driving tasks.
Who it is for
- Researchers studying vision-language models for autonomous-driving planning.
- Developers exploring supervised fine-tuning and GRPO training workflows.
- Teams interested in generating planning-reasoning data or evaluating meta-action planning accuracy.
Documented uses
- Fine-tune a vision-language model for driving-planning reasoning.
- Continue training with the repository's GRPO reinforcement-learning stage.
- Use supplied prompt templates and an example to support planning-reasoning data preparation.
- Evaluate meta-action planning accuracy with the supplied script.
Requirements & limits
The repository is public and identifies its license as Apache-2.0.
- The documented setup creates a Conda environment with Python 3.11 and runs setup.sh.
- Training and evaluation are provided as separate shell scripts.
- Prompt templates and an example QA file are included for planning-reasoning data preparation.
- The README does not specify hardware requirements.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up the Alphadrive research environment
Clone the repository and prepare the Conda environment specified in its README.
Before you start
- Linux environment
- Git
- Conda
Steps
Clone the repository with the documented Git command.
Create a Conda environment named alphadrive with Python 3.11.
Activate the environment.
Run the repository's setup.sh script.
Expected outcome
The documented environment setup sequence has been completed; the README proceeds to data preparation, training and evaluation.
Guide limitations
- The setup instructions do not specify hardware requirements.
- This guide covers environment setup only, not successful training or evaluation.
Run the documented training and evaluation stages
Follow the repository's listed supervised fine-tuning, GRPO training and planning-accuracy evaluation stages.
Before you start
- Alphadrive repository and setup completed
- Planning-reasoning data prepared using the provided prompt templates and example
- A configured training environment
Steps
Use the prompt templates in data_tools/prompt_hub.py and consult data_tools/example.json for the provided example.
Run sh train_tools/run_train_sft.sh for supervised fine-tuning.
Run sh train_tools/run_train_grpo.sh for the reinforcement-learning phase.
Run sh eval_tools/qwen2vl_plan_cmd_eval.sh to evaluate meta-action planning accuracy.
Expected outcome
The README provides a staged training workflow and a script for assessing meta-action planning accuracy.
Guide limitations
- The README does not specify hardware requirements or guarantee that a particular environment will complete these stages.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- hustvl/AlphaDrive · Official homepage title · retained in its source language
Related options
open-r1 — Acknowledged in the README as one of the projects on which Alphadrive is built.
R1-V — Acknowledged in the README as one of the projects on which Alphadrive is built.