SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning

A new model named SOLE-R1 has been introduced, designed to enhance robot learning by utilizing video-language reasoning as the sole reward signal in reinforcement learning. This model processes raw video observations alongside natural-language goals, enabling it to perform spatiotemporal reasoning and provide dense estimates of task progress.

The development of SOLE-R1 is significant as it addresses limitations faced by existing vision-language models in reinforcement learning, particularly under conditions of partial observability and distribution shift, potentially leading to more effective and reliable robotic learning systems.