Fine Tuning a Model on HighD for Driver Agent on Highway

1 prepraed the training data

extracted data from HighD Dataset, assess driver's skill level, get expert data to fine tune the LLM model

input data past 3 seconds data

output data

2 fine tuning the LLM model

data

data balance

3 LLM

driver agent relection agent memory RAG decision post process

4 Validation

Simulation Environment highway-env 'highway-v0'

4 limitations

4.1 overview

• Current technology stack including E2E, VLA and WR models. E2E is widely-adopted technology with sensors input (camera, radar, lidar, gnss, map, navigation) while the output is Accelerator pedal position (APP) and steering wheel angle(SWA). End-to-end models are characterized by a lack of inter-module information loss and the removal of tedious, conflicting rules typical of conventional methods. Actually, we concentrated on prediction + decision + planning modules here. besides, we also give the chain of thoughts in our fined tuning model.

4.2 input

ignore the obstacles exists in the real scenario, car and trucks on highway, lane info, select postion, kpi as input.

4.3 output

the output is a discrete decision among in idle, accelerate, decelerate and left lane change, right lane change.

5 Todo

more scenario: intersection

6 Reference

@article{wen2023dilu,
  title={Dilu: A knowledge-driven approach to autonomous driving with large language models},
  author={Wen, Licheng and Fu, Daocheng and Li, Xin and Cai, Xinyu and Ma, Tao and Cai, Pinlong and Dou, Min and Shi, Botian and He, Liang and Qiao, Yu},
  journal={arXiv preprint arXiv:2309.16292},
  year={2023}
}

7 Usage

7.1 Requirements

For an optimal experience, we recommend using conda to set up a new environment for DiLu.

conda create -n dilu python=3.8 
conda activate dilu
pip install -r requirements.txt

7.2 Configuration

All configurable parameters are located in config.yaml.

Before running DiLu, set up your OpenAI API keys. DiLu supports both OpenAI and Azure Openai APIs.

Configure as below in config.yaml:

OPENAI_API_TYPE: # 'openai' or 'azure'
# below are for Openai
OPENAI_KEY: # 'sk-xxxxxx' 
OPENAI_CHAT_MODEL: 'gpt-4-1106-preview' # Alternative models: 'gpt-3.5-turbo-16k-0613' (note: performance may vary)
# below are for Azure OAI service
AZURE_API_BASE: # https://xxxxxxx.openai.azure.com/
AZURE_API_VERSION: "2023-07-01-preview"
AZURE_API_KEY: #'xxxxxxx'
AZURE_CHAT_DEPLOY_NAME: # chat model deployment name
AZURE_EMBED_DEPLOY_NAME: # text embed model deployment name  

7.3 Run

7.4 Visualize

We provide a visualization scripts for the simulation result.

python ./visualize_results.py -r results/highway_0.db -m memories/20_mem

Open http://127.0.0.1:7860 to view each frame's prompts and decisions!