The training regularized 60 simulation years with human data
Trained with PufferLib + PufferDrive! Really awesome method as well. Considering integrating into our native trainer in a future update.
New Paper: Human-like Autonomy Emerges from Self-Play and a Pinch of Human Data.
We trained self-play RL on 60 years of simulation on 1 GPU in ~15 hours. Regularizing with 30 minutes of demonstration data produces much more human-like driving policies!
Users are excited by self-play RL producing human-like driving policies with minimal data, praising the approach as superhuman and thanking collaborators for the achievement.
No Digg Deeper questions have been answered for this story yet.
Project page: https://spiced-self-play.com/ Arxiv: https://arxiv.org/abs/2606.19370
New Paper: Human-like Autonomy Emerges from Self-Play and a Pinch of Human Data.
We trained self-play RL on 60 years of simulation on 1 GPU in ~15 hours. Regularizing with 30 minutes of demonstration data produces much more human-like driving policies!
@daphne_cor Unregularized is a better driver. Deploy it in real life. Superhuman
A big thank you to my collaborators for their unique contributions to this work: @julianh651, Zixu Zhang, Waël Doulazmi, @kev_joseph_, Jaime Fernández Fisac, and @EugeneVinitsky!
@daphne_cor Cool!!!
@daphne_cor The reason unregularized drives the way it does is because it has god sight. Give it only raytraced pixels by simulating lidar
