Key-Gram: Extensible World Knowledge for Embodied Manipulation

In three linesKey-Gram is a conditional-memory framework separating linguistic knowledge from visual reasoning for embodied control. It decomposes instructions into task-specific key-grams, retrieves linguistic priors via O(1) hashed lookup, and injects them into hidden layers. Achieves 29.5% gains on RoboTwin2.0, 35.8% on LIBERO-Plus, 15.4% on real-world tasks.

Summary generated by Claude — human-verified