SELF-IMPROVING VISION-LANGUAGE-ACTION MODELS WITH DATA GENERATION VIA RESIDUAL RL

PLD   2025   2511.00091  
arXiv Code Project

论文的三那个阶段都展示在以下的algo中 需要注意的点是environment step返回的reward,real world考虑训练一个binary reward classifier:输入是相机图像和 proprioception,输出当前状态是否成功。论文 appendix 里说:

My concern

作者其实也有比较long-horizon tasks,但是得出的结论是:the fine-tuning of the data PLD exceeds the tuning of the data from the baseline policy roll-out (self-bootstrapped), but still falls short of the performance achieved with demonstrations by human experts 因此从这个角度看,简单的RL进行修正并不能完全解决问题,且DAgger还是针对long-horizon任务的一个非常有效的数据采集手段。



← Home · All Readings