VLA-Pro
VLA-Pro: Cross-Task Procedural Memory Transfer for Vision-Language-Action Models
Shengyu Si, Yuanzhuo Lu, Ruimeng Yang, Ziyi Ye, Zuxuan Wu, Yu-Gang Jiang · 2026-05-28
- Management
- T1 Static Memory
- Memory is built once and reused without change during deployment.
- Representation
- Hybrid
- Every structured procedural-state index points to a parameterized LoRA memory that is activated after retrieval.
How experience becomes memory
Each path is one memory component, traced from what is stored to how it is found again.
structured action, object, orientation, and interaction-point sequences
- Represented asSpatial, temporal & cross-modal structured memoryCross-represented explicit
- Stored inVector-indexed collection
- Retrieved bySimilarity search
- Source evidenceSource not retained
Each task adapter is indexed by a sequence of structured procedural states; action-aware field similarity retrieves the top-k relevant source tasks.Full paper: Memory · PDF p. 4
task-specific LoRA procedural memories
- Represented asExperience-conditioned parametric memoryInternalized latent & parametric
- Stored inModular adapter bank
- Retrieved byLatent / parametric activation
- Source evidenceSource not retained
Offline task-specific LoRA adapters store executable experience; retrieved adapters are similarity-weighted, fused, and temporarily loaded for the current action chunk.Full paper: Memory · PDF p. 4 · Algorithm 1
Related systems
- DRAET1 · Embodied & 3D
- BrainNavT2 · Embodied & 3D
- HIMMT2 · Embodied & 3D
- MemoryExplorer / LMEE-BenchT2 · Embodied & 3D
- Affordance RAGT1 · Embodied & 3D
- C-NavT1 · Embodied & 3D