SGTP: Sampling-based Game-Theoretic Planning for Real-Time Multi-Vehicle Autonomous Racing

Researchers introduced SGTP, a real-time game-theoretic planner for autonomous racing that uses GPU-accelerated sampling and rollout evaluation to generate diverse strategies.

The system then filters candidate plans for track and collision constraints, improving safety while supporting close-quarters multi-car competition.

In simulation, SGTP achieved strong win and completion rates, averaged about 0.095 seconds per run, and scaled to as many as 10 agents.

The team also released code and a benchmark, giving the field an open platform for follow-up research.