Episode Details

Proposes a framework that restores the Markov property for RL fine-tuning of VLAs despite asynchronous inference delays, enabling effective RL training on large vision-language-action models. This directly unblocks post-training RL for frontier robot policies operating under real-world latency constraints.