⏤ Our Driving IntelligenceDeep Active Inference
Our Deep Active Inference approach enables our autonomous system to build an internal world model, continuously interpreting its surroundings and anticipating what might happen next.
"The system doesn't just react to what it sees but rather asks itself what it might be missing before it decides what to do."
Instead of simply reacting, it's designed to recognize uncertainty and understand its own limitations. The system operates epistemically, meaning it recognizes when it doesn't have enough information, and actively seeks it out before acting.
In unclear situations, this explorative behaviour lets us handle complex and unexpected traffic scenarios with greater awareness and adaptability, producing safer, more predictable outcomes.
Practical Examples
"Creep-and-Peek"
At blind intersections Unlike reactive systems that either stall or risk blind maneuvers, our system recognizes when its view is obstructed. It actively executes micro-movements to position sensors effectively, systematically reducing uncertainty before making a turn.
Anticipating Hidden Hazards.
When passing stationary buses or high-density parking zones, the environmental representation model accounts for risks it cannot yet see. By projecting potential hidden pedestrians, it proactively adjusts speed and posture before a conflict even appears.
Handling Edge Cases & Unknown Objects
When faced with rare or unfamiliar obstacles (out-of-distribution scenarios), standard AI often misclassifies or panics. Our approach quantifies its own uncertainty, relying on underlying physics and motion dynamics to pass safely without unnecessary phantom braking.
Interactive "Negotiation"
In tight spaces In narrow street bottlenecks, the system uses subtle probing actions to evaluate the intentions of oncoming drivers by resolving potential deadlocks through continuous, dynamic mutual adjustment.