Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents
The episode discusses the paper "Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents." Hosts discuss how this memory-augmented agentic framework uses a division of labor between an LLM planner and a frozen Vision-Language-Action (VLA) model. Key features include using analytic primitives to handle uncertainty, task-specific memory for reusability, and isolating non-contact execution to improve reliability and debugging.