IMCS Lab

Research

Securing intelligent mobility and making physical AI dependable — from regulation and risk assessment to safe decision-making for robots and vehicles.

Physical AI and intelligent mobility

Robots and vehicles driven by AI now act in the physical world, where a wrong decision causes physical harm. We study how such systems can perceive across domains, decide with multimodal models, and know when not to act.

Capability-aware safe meta-decision: the Try-Ask-Refuse framework

Physical AI robots assess their own capability and choose to try, ask a human, or refuse, so that unsafe or infeasible commands are declined rather than attempted. See Projects for details.

Try-Ask-Refuse: a command goes through capability assessment (skill, feasibility, safety, confidence) to a meta-decision to try, ask a human, or refuse
The Try-Ask-Refuse framework for capability-aware safe meta-decision

Cross-domain perception and LLM/VLM-based multimodal decision systems

Perception that stays reliable across countries, roads and sensing conditions, and decision-making that combines language and vision models for AI-driven mobility, in joint Korea–Japan research.

Driving data from Korea and Japan feed cross-domain perception, then LLM/VLM-based multimodal decision, then a driving decision
Cross-National AI Mobility: from cross-domain perception to LLM/VLM-based decisions

Multi-sensor fusion and adaptive point cloud processing

Fusing camera, LiDAR and other sensors, and downsampling point clouds around 3-D regions of interest to keep perceptual quality at lower cost.

Automotive cybersecurity

Cybersecurity management systems, UN R.155 and ISO/SAE 21434

Laws and regulations for automobiles now require the reinforcement of security as well as safety, which reshapes how manufacturers build and certify vehicles.

Automotive regulations for safety and security
Automotive regulations for safety and security

TARA and the PIER method for cybersecurity risk assessment

The PIER approach considers cyber-resilience: beyond the risk factors of probability and impact, it also covers exposure and recovery. See Projects for details.

The PIER method for TARA
The PIER method for TARA

In-vehicle IDS and remote analysis

  • Automotive intrusion detection based on message and state learning in the in-vehicle network
  • Avoiding the classification of data that will not be used, via a hierarchical multi-class model
Intrusion detection for the in-vehicle network
Intrusion detection for the in-vehicle network
Hierarchical multi-labeled classification method for intrusion detection
Hierarchical multi-labeled classification for intrusion detection

Plug and Charge and ISO 15118

Authentication and automatic payment for electric vehicle charging, built on a public key infrastructure. See Projects for details.

Plug and Charge based on ISO 15118
Plug and Charge based on ISO 15118

Automotive OTA software update

Integrity validation of ECU software to prevent manipulated code injection attacks.

Automotive software update Over-The-Air
Automotive software update Over-The-Air (OTA)

Auto-ISAC council

Enhancing cybersecurity and sharing information on cyber threats and vulnerabilities among automotive industry members and the US government since 2015. Auto-ISAC promotes collaboration and coordination among its members, such as vehicle manufacturers, suppliers and automotive cybersecurity professionals.

The Auto-ISAC council
The Auto-ISAC council

Fundamental technology for autonomous vehicles

Perception

Recognizing moving objects from multiple sensing data fusion, including adaptive point cloud processing for perceptual quality.

Planning

Trajectory planning that predicts the movement and interference of moving objects for collision avoidance.

Software update

Automotive OTA update mechanisms aligned with UN R.156 and ISO 24089.