Commit History
-
Add the crowd models: pedestrian predictors that drive the crowd, not the robot
Verified signatureSeven checkpoints from humanmodel/train_cb_models.py (pooled ETH/UCY, 36,197 windows) and humanmodel/train_jrdb.py (JRDB alone, 43,750 train / 6,923 test windows). phys / slstm / sgan / tf are our re-implementations of a CV+residual model, Social-LSTM (Alahi CVPR'16), the Social-GAN generator (Gupta CVPR'18) and a transformer residual, all behind one heading-normalized CV-residual interface so they are directly comparable. They are the CROWD, not the controller: they step the pedestrians in the crowd-behaviour study while a scripted robot traverses the corridor. The controller weights in the repo root are a separate stack and none of them load these.65f383e