Page 809 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

compression and update frequency reduction, targeting an order-of-magnitude reduction
in required bandwidth versus naive approaches. Privacy-preservation will be substantially
reinforced: DP will be tuned to provide quantifiable privacy guarantees. The chosen DP noise
levels will be calibrated so that the model’s utility remains high while ensuring that individual Transport 4.10: Intelligent
data points from any single vehicle cannot be inferred from the shared updates. In addition,
more advanced privacy-preserving techniques will be prototyped. The mid-term goal is a
demonstrable federated learning system where, for example, communication costs are
drastically lower than central training and privacy metrics meet regulatory guidelines. By this
stage, we expect to report quantitative metrics such as the bytes communicated per training
round, the convergence time to reach a target performance, and the achieved privacy level
alongside any trade-offs in accuracy.

Beyond 2 years, the vision is to evolve the intelligent mobility system from advanced prototyping
to real-world deployment, while pushing the boundaries of autonomy, collaboration, and
security:

Real-World Deployment & Continuous Learning: Transition the LLM-integrated autonomous
driving system from the digital twin and test fleets to real-world operating environments. This
will likely involve phased deployments in controlled settings before broader rollout. A key
long-term goal is to enable continuous learning and adaptation: as vehicles encounter new
situations on the road, they will leverage federated learning to continuously improve the shared
models without requiring downtime or centralized data collection. The digital twin orchestration
platform will play a pivotal role in this deployment. By mirroring each vehicle with a live digital
twin, the system can perform real-time shadow evaluations of novel scenarios in the virtual
space before the vehicle takes action, thus bridging the gap between theoretical modeling
and safe real-world behavior. The long-term outcome is a fleet of LLM-enabled autonomous
agents that improve over time, providing evidence of learning in the wild while new knowledge
is instantly shared across the fleet through federated updates.

Advanced LLM–Agent Capabilities: In the long run, the integration of LLMs will be extended
to more complex and critical driving tasks. We envision LLMs handling higher-level reasoning
such as cooperative maneuvers between vehicles, strategic route planning under congested
or hazardous conditions, and semantic understanding of traffic rules and human gestures. The
agent architecture will evolve into a hierarchical model: the LLM acting as a high-level cognitive
layer for reasoning and communication, and the low-level controllers handling execution. This
arrangement will be refined to ensure real-time performance and fail-safe operation. Moreover,
the LLMs could be specialized or fine-tuned on mobility-specific corpora by this stage, giving
them a richer understanding of traffic narratives and regulations. By more than 2 years, we
aim for the LLM-augmented system to be capable of complex multi-agent interactions, which
would be validated first in the simulator and then in field trials. Throughout this, rigorous testing
and verification will be conducted to certify that the LLM’s behavior meets safety and reliability
standards required for autonomous driving deployments.

Federated Learning 2�0 – City-Scale and Fully Privacy-Preserving: As the number of deployed
autonomous vehicles grows, the federated learning framework will be scaled to city-scale fleets
and beyond. The long-term goal is a globally distributed learning network where thousands
of vehicles and infrastructure units collaborate to train robust AI models for perception and
decision-making. Emphasis will be on maintaining system scalability, fairness, and rigorous
security despite the scale. Techniques such as hierarchical federated learning may be employed
to manage communication loads. We will target stringent privacy guarantees, aiming for strong

773