While vision-language-action (VLA) models handle short manipulation primitives effectively, they struggle to maintain state coherence across extended sequences. This work introduces a neuro-symbolic framework coupling learned VLA control with explicit task graphs and multimodal procedural memory to track step transitions and context. The authors also inject demonstration-derived pseudo-gaze cues into fine-tuning and inference to guide spatial-temporal grounding. Tested on multi-step workspace clearing and surgical-instrument handling, the approach shows how symbolic boundaries can steady end-to-end policies across branching procedural workflows.
No heat snapshots are available in the last 24 hours.