Page 17 - IJAMD-1-3

International Journal of AI for
Materials and Design
Smart cockpit design with generative models

allow for comprehensive state monitoring (e.g., emotion deep learning algorithms. In this framework, the MAS
recognition, fatigue detection, stress monitoring, comprises a perception agent, a decision-making agent,
and health assessment) and enable personalized and an execution agent. The perception agent employs
adjustment to the passenger’s condition. This generative models to collect visual information and
significantly enhances the mobility experience, safety, recognize the surrounding environment. The decision-
and efficacy of personalized regulation strategies. The making agent generates personalized driving strategies
collection of user data through sensors (e.g., wearables, based on the driver’s identified driving style and real-time
cameras, microphones) supports the construction and environmental conditions. Multi-modal feedback, such as
maintenance of the HDT model, ensuring continuous voice reminders and visual cues, is provided to the driver.
updates that accurately reflect users’ physiological and For example, if the driver’s HDT model identifies an
mental status in real time. aggressive driving style and the perception agent detects
that the driver is currently navigating in rainy weather with
MASs comprise a perception agent, a decision-making
agent, and an execution agent. The perception agent is care and several pedestrians crossing the road ahead, the
decision-making agent generates a deceleration command
responsible for monitoring and collecting user status and to remind the driver to stay focused. The executive agent
intensity information. The decision-making agent analyzes then issued a voice alert to the driver, stating, “There are
this status information and devises regulation strategies. The pedestrians ahead; please slow down to 30 km/h and pay
execution agent is tasked with the actual implementation attention to safety.”
of conditioning measures, such as adjusting music or
providing health reminders and rest advice to drivers or 4. Case study
passengers. The HDT detects user states and transmits this
information to the perception agent of the MAS, ensuring In this section, we leverage the proposed CockpitGemini
that the decision-making and execution agents base their personalized design methodology to execute a personalized
decisions and operations on accurate and up-to-date product design and present a case study focusing on the
information. For instance, the perception agent detects the personalized automobile seat design to showcase the
frequency of the driver’s eye and head movement in real viability of the proposed approach within a targeted
time and determines that the driver may be experiencing context. It is demonstrated from the following two aspects:
fatigue. This status information is then broadcast to individual HDT construction and generative model-based
all related agents. Based on the recognized fatigue and MAS for personalized seat design, respectively.
intensity of the driver, the decision-making agent decides 4.1. Individual HDT construction
to implement measures such as adjusting the seat angle,
playing soothing music, initiating proactive, empathetic We employed a vision-based HDT modeling approach
speech, and adjusting the vehicle’s interior lighting. These to capture customers’ RGB-D data in real time using an
tasks are assigned to different execution agents, such as the Azure Kinect camera, establish individualized human
seat control agent, the entertainment system agent, and meshes, and perform ergonomic analysis in a virtual
the environment control agent. Each execution agent then environment. The implementation details of the vision-
takes actions to adjust interior temperature and lighting, based HDT modeling approach are depicted in Figure 4.
play soothing music, and adjust the seat angle according The model comprises an advanced convolution neural
to the instructions. The perception agent continuously network backbone, a body part segmentation branch, and
monitors the driver’s status and transmits feedback to the an ergonomics branch. The input data consists of RGB-D
decision-making intelligence, which adjusts the strategy information on the human’s various gestures. Subsequently,
51
based on this feedback to ensure that the passengers regain a high-resolution network (HRNet-W32 ) is utilized to
their comfort. develop the body segmentation and ergonomics branches,
fulfilling the perceptual requirements of the HDT
3.2.4. Personalized driving strategy recommendations modeling process. This results in the creation of a highly
The integration of the HDT model and the MAS can personalized human model. The HDT perception model
recommend personalized driving strategies to drivers. is implemented utilizing the widely adopted PyTorch deep
Historical driving behavior data, including metrics such learning framework, with Nvidia RTX 3090 GPU providing
as driving speed, acceleration, gas pedal angle, brake pedal enhanced hardware acceleration. The Adam optimizer is
-
frequency and force, and steering angle, are collected to employed with a learning rate of 5 × 10 ⁵, a batch size of 2,
construct an HDT model. These data are then classified and a total of 100 training epochs.
into different driving styles (conservative, aggressive, Regarding the part segmentation branch, we introduced
and moderate) using advanced machine learning and a body part attention mechanism to construct an HDT

Volume 1 Issue 3 (2024) 11 doi: 10.36922/ijamd.4220