CooperativeDrive AI

Inspiration

Drivers can only react to what they can see. A pedestrian may be hidden behind a truck, debris may be around a curve, or another vehicle may be approaching a blind intersection. However, another nearby vehicle may already be able to see that danger.

This inspired us to ask: What if vehicles could share what they know with each other before a hazard becomes visible to every driver?

That idea became CooperativeDrive AI.

What it does

CooperativeDrive AI allows vehicles to share important road observations with nearby vehicles.

The system combines road observations with information such as vehicle speed, lane, direction, and intended maneuver. IBM watsonx Orchestrate then reasons about whether the reported situation is relevant to another vehicle, assesses the risk, and creates a short warning.

For example, if one vehicle detects a pedestrian partially hidden behind a truck and another vehicle is about to turn into that area, CooperativeDrive AI can warn:

PEDESTRIAN AHEAD — Wait before turning right.

The goal is not to send every observation to every vehicle. The system determines which vehicles may actually be affected and sends the warning only to them.

How we built it

We built the application using Python and FastAPI with a simple web interface.

IBM Bob helped us plan and develop the initial prototype. We used IBM watsonx Orchestrate to build and deploy the CooperativeDrive AI reasoning agent.

The agent considers road observations together with vehicle context and produces:

LOW, MEDIUM, or HIGH risk
the detected hazard
a short recommended action
the reason for the warning
the vehicles that should receive the warning

We designed the system so that measurable information such as speed, distance, and time-to-conflict is not invented by the AI. These values must come from vehicle data or deterministic calculations.

Challenges we faced

One challenge was deciding what information should actually be communicated between vehicles. Sending every observation to every nearby vehicle would create unnecessary warnings and information overload.

We therefore focused on relevance: the vehicle's direction, lane, intended maneuver, and whether its path could be affected by the reported hazard.

Another challenge was keeping the interface simple. A driver does not need to see raw telemetry, model output, or long explanations. The final interface focuses on the information needed to act: risk, hazard, warning, reason, and affected vehicle.

What we learned

We learned that cooperative vehicle intelligence is not only about detecting objects. The more important problem is understanding whether an observation matters to another vehicle.

AI can help interpret the situation and communicate it clearly, while deterministic safety calculations and guardrails can provide additional validation for safety-critical information.

What's next

Future versions could connect directly to vehicle cameras and onboard sensors, process live road video, calculate real-time vehicle trajectories, and communicate through V2V/V2X systems.

Our prototype demonstrates the core idea:

What one vehicle cannot see, another vehicle can.