Key-Gram: Extensible World Knowledge for Embodied Manipulation
Researchers have developed Key-Gram, a new framework designed to improve embodied control systems by separating linguistic knowledge from visual reasoning. This approach uses a conditional-memory module to store and retrieve instruction-derived knowledge, allowing the main model backbone to focus on visual processing and action inference. Key-Gram has demonstrated significant performance gains across various robotic manipulation tasks, including RoboTwin2.0 and real-world dual-arm scenarios, by enhancing compositional grounding and transfer learning. AI
IMPACT Externalizing linguistic memory in embodied AI could lead to more adaptable and efficient robotic systems capable of complex instruction following.