Embedded Files
System Overview
System Overview
In our experiments, we utilize instances of OpenAI's Generative Pre-trained Transformer (GPT) 4 module for both language models.
We conducted a number of different experiments in which the passenger varies requirements and level of vagueness in preferences. A few scenarios and their resultant trajectories are described below.
The inputs and outputs detailed below are from the first language model in the system pipeline.
We employ chain-of-thought prompting, encouraging the language model to reason through what the human is requiring before coming to a decision about optimal parameter values. This provides substantially better results, as the model is able to output responses that are more interpretable.
The output responses and a detailed function API template describing the cost function in an encapsulated class format will be sent as input to the second language model.
An instance of this created class coupled with the executable of the model-predictive control program will be the final system.
Experiments
Experiments
Scenario 1:
Scenario 1:
Passenger Preferences:
I prefer a smooth ride, so please avoid abrupt accelerations and decelerations.
GPT-4 Output:
Given the human's preference for a smooth ride and the request to avoid abrupt accelerations and decelerations, the importance values for each term should prioritize minimizing jerk and ensuring comfort while still maintaining some adherence to speed.
Here is the completed template:
The desired speed along straight segments should be 25.0. This is vLine.
The desired speed along curved segments should be 15.0. This is vCurve.
desiredSpeedAdherence = 0.25
longitudinalJerkMinimization = 0.45
turningComfortImportance = 0.30
These values reflect the passenger's preference for a smooth and comfortable ride, placing a significant emphasis on minimizing jerk while still maintaining a reasonable speed during turns and straight segments.
CAV Trajectory Plots:
Scenario 2:
Scenario 2:
Passenger Preferences:
Please drive fast.
GPT-4 Output:
Given the human's preference to drive fast, the importance values for each term should be set to prioritize speed adherence while maintaining some level of comfort and safety, especially during turns.
Here is the completed template:
The desired speed along straight segments should be 35.0. This is vLine.
The desired speed along curved segments should be 30.0. This is vCurve.
desiredSpeedAdherence = 0.60
longitudinalJerkMinimization = 0.15
turningComfortImportance = 0.25
These values reflect the passenger's preference for a fast drive while ensuring turns are handled safely and there is some level of comfort regarding acceleration changes.
CAV Trajectory Plots:
Scenario 3:
Scenario 3:
Passenger Preferences:
I have motion sickness.
GPT-4 Output:
Given the human's preference to drive slow, the importance values for each term should prioritize comfort and safety over adhering to a specific speed.
Here is the completed template:
The desired speed along straight segments should be 25.0. This is vLine.
The desired speed along curved segments should be 15.0. This is vCurve.
desiredSpeedAdherence = 0.20
longitudinalJerkMinimization = 0.40
turningComfortImportance = 0.40
These values reflect the passenger's preference for a slow and careful drive, ensuring high importance on minimizing jerk for comfort and safely handling turns.
CAV Trajectory Plots:
Author of the Page: Shruti Banerjie
Page updated
Report abuse