Research project associated with Chinese Conference on Pattern Recognition and Computer Vision (PRCV 2025); proceedings published in 2026.
Core approach
Uses a diffusion-based preference score as a reward for trajectory prediction, together with rejection sampling.
A pretrained diffusion backbone scores trajectory features for alignment with human behavior. The score guides reinforcement learning and selection of training trajectories.
Associated publication
Enhancing Human Trajectory Prediction with Reinforcement Learning from Quantified Human Preferences