PRTS: A VLA Foundation Model Using Goal-Conditioned Reinforcement Learning

Researchers have created an innovative foundation model known as PRTS, which stands for Primitive Reasoning and Tasking System. This model utilizes Goal-Conditioned Reinforcement Learning for its pretraining, setting it apart from conventional Vision-Language-Action models that rely heavily on supervised behavior cloning. Instead, PRTS views language instructions as goals and uses contrastive reinforcement learning to form a unified embedding space. Within this framework, it calculates the likelihood of reaching a goal from a given state-action by approximating the log-discounted goal occupancy. This approach addresses a significant limitation of existing VLA models, which often overlook the importance of goal orientation and task progression. You can find this research on arXiv under the identifier 2604.27472.