Linguistic Trajectory Encoding for Efficient Long-Horizon Spatial Memory in Embodied Agents

To address the issue of the lack of memory for querying dynamic object state transitions in long-term observations ranging from hours to days for embodied intelligent agents, researchers proposed the Linguistic Trajectory Encoding (LTE) scheme. This scheme uses a hybrid representation of natural language descriptions, sparse spatial anchors, and visual anchors to adaptively compress the motion history of dynamic objects. To evaluate its long-time delay capabilities, the researchers constructed the Spatial Memory Benchmark (SMB), which included semantic trajectory retrieval and long-range object retrieval tasks. In the SMB tests, the LTE-based system achieved a semantic trajectory retrieval success rate of 45.3% and a long-range object retrieval rate of 48.7%, which were superior to those of structured memory and visual language models (with best priors of 31.9% and 34.4%, respectively).