Summary
VLA-Pro is a plug-and-play framework that adapts task-relevant manipulation patterns at inference time using a procedural memory bank. The current procedural state (summarizing ongoing execution stage) queries the bank, which stores task-specific LoRA adapters as parameterized procedural experience; retrieved adapters are fused into the base VLA to reuse relevant patterns while avoiding interference from unrelated task behaviors.
Key Contributions
- Procedural state extraction module for querying a task-specific LoRA adapter memory bank
- Each bank entry stores a task-specific LoRA adapter encoding a distinct manipulation pattern
- Runtime adapter fusion enables cross-task knowledge transfer without full retraining
- Plug-and-play design compatible with existing VLA backbones
Significance
VLA-Pro frames cross-task generalization as a memory retrieval problem, achieving dynamic adaptation at inference time—an approach that scales naturally as the skill library grows without modifying the base model.