Interaction-Aware Model Predictive Decision-Making for Socially-Compliant Autonomous Driving in Mixed Urban Traffic Scenarios
New AI decision-making model uses real-time pedestrian intent prediction to shorten crossing negotiations.
A research team led by Balint Varga has published a new paper on arXiv detailing their Interaction-Aware Model Predictive Decision-Making (IAMPDM) framework, designed to enable autonomous vehicles to navigate complex urban scenarios with pedestrians more effectively. The core innovation is the integration of a pedestrian intention-prediction module—inspired by gap-acceptance theory—with a real-time Model Predictive Control (MPC) system. This allows the vehicle to jointly reason about a pedestrian's likelihood of crossing (via a continuous 'crossing-propensity' signal) and its own control actions, aiming for socially compliant behavior rather than just collision avoidance. The system was implemented using the CasADi/IPOPT optimization framework and tested in a motion-tracked simulator.
The study compared IAMPDM against a rule-based intention-aware controller (RBDM) and a conservative non-interactive baseline (NIA) in a human-in-the-loop experiment with 25 participants. Results showed that both intention-aware methods (IAMPDM and RBDM) significantly shortened negotiation and scenario completion times compared to the non-interactive NIA, though they operated with tighter time-to-collision and distance margins. Critically, participants rated the intention-aware algorithms higher for subjective comfort, safety, and trust. The paper discusses crucial implications for real-world deployment, including proposed safety guardrails like minimum surrogate-safety margins and deadlock prevention mechanisms to balance the efficiency gains with robust safety assurances in mixed traffic.
- The IAMPDM framework predicts pedestrian intent using a continuous crossing-propensity signal based on time-to-collision with an intention discounting mechanism.
- Human-in-the-loop tests with 25 participants showed intention-aware methods reduced crossing time and improved ratings of comfort, safety, and trust vs. a non-interactive baseline.
- The system is implemented for real-time use with CasADi/IPOPT and proposes deployment guardrails like minimum safety margins to balance efficiency and safety.
Why It Matters
This research moves AVs beyond simple obstacle avoidance toward nuanced, human-like social negotiation, a critical step for deployment in dense urban environments.