Physical AI in Cyber-Physical Systems
Card Grid View — Paper by Gurdur Broo (2025)
1. Why Digital AI Fails
- Why can't LLMs / Digital AI be used directly?
- Lack physical grounding: mass, friction, torque, energy, timing
- Internet data lacks industrial domain context
- Tacit operator knowledge not captured by text
- LLMs unsuited for irreversible physical processes
- Main limitations
- Cannot account for all real-world constraints
- Physical phenomena seen as noisy digital signals
- Cannot learn in real-time during operation
2. Fundamental Bottleneck
- Main bottleneck
- Information integration across cyber-physical boundaries
- Fragmented, siloed, heterogeneous data
- Qualitative gap (not quantitative)
- Tacit knowledge cannot be codified into rules
- Why is integration difficult?
- Data from different systems, formats, domains
- Different units, unsynchronized timing
- Legacy systems and non-uniform protocols
3. Digital vs Physical
- Digital Environment
- Unlimited computational resources
- Perfect reproducibility
- Instant state transitions
- Constraints can be violated and corrected
- Physical Environment
- Irreversible processes
- Energy conservation
- Material fatigue, thermal expansion
- Real consequences: waste, cost, safety
4. Physical Intelligence
- What defines it?
- Understanding mechanics, materials, energy, timing
- Understanding real-world constraints
- Adaptation to physical phenomena
- Tacit knowledge + data-driven insight
- Digital AI is best at
- Pattern recognition, text generation
- Image classification, prediction
5. Path Forward
- 5 Development Pillars
- Physics-Informed AI Architectures
- Real-Time Multi-Modal Data Integration
- Domain-Adaptive Foundation Models
- Human-AI Collaborative Intelligence
- Integrated Digital Twin Ecosystems
- Key Priorities
- Reorient to integration, not model scaling
- AI must be purpose-built on industrial data
- Build info infrastructure across cyber-physical
6. Human vs AI Roles
- AI provides
- Data-driven insights
- Pattern recognition
- Humans provide
- Contextual understanding
- Safety judgment
- Practical experience
- Tacit knowledge
- Result
- AI + Human is better than either alone
- Complementary, not replacement
7. CPS Definition
- Definition
- Integration of computation, networking and physical processes
- Characteristics
- High complexity
- More connected = more unpredictable
- Human-centered design is important