Summary
Agentic-VLA is an agentic training framework that enables VLA models to efficiently adapt online to novel environments and tasks. It addresses two critical limitations of standard VLA training: poor generalization and low sample efficiency, by dynamically synthesizing rewards, guiding exploration with a critic, and maintaining an experience memory for warm-start adaptation.
Key Contributions
- Adaptive Reward Synthesis: dynamically generates and adjusts reward functions based on current VLA capabilities, decomposing complex tasks into learnable sub-goals for curriculum learning
- Language-Guided Exploration: a critic model provides structured, semantically grounded exploration guidance rather than random sampling
- Experience Memory: stores and retrieves task-relevant policy weights to warm-start adaptation on similar tasks, improving sample efficiency
Significance
Agentic-VLA bridges the gap between offline pre-training and online deployment, providing a principled agentic loop for continuous VLA improvement without requiring large additional demonstration datasets.