AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-23
A Transformer-based Approach for Real-time Multi-agent Path Planning with Varying Initial/Final Conditions
Real-time path planning for many agents with varying start and goal configurations becomes intractable for classical solvers as the agent count grows. The authors propose OCFormer, a transformer that maps a decoupled latent space to the coupled physical space, trained with a physics-informed loss and no precomputed trajectories. It produces near-optimal trajectories for unseen configurations in one forward pass (about 220 ms), adapts to out-of-distribution perturbations in seconds via L-BFGS fine-tuning, and scales to 128 agents in 2D and 3D navigation experiments.