Elastic Queries Reinforcement Learning: Self-Aware Policy Execution for VLA Models

The article presents Elastic Queries Reinforcement Learning (EQRL), a novel framework for enhancing the execution of vision-language-action (VLA) models in robotic manipulation. EQRL introduces a lightweight latent-schedule adaptor that dynamically adjusts the computation resources based on the difficulty of the task, allowing for variable chunk execution without fine-tuning the VLA model. This approach significantly reduces inference costs while maintaining or improving task success rates, making it valuable for practitioners seeking to optimize resource allocation in robotic control tasks.