ECHO memory boosts robot task success by 12.8%

A team of researchers has introduced ECHO (Experience Consolidation and Hierarchical Organization), a novel memory framework designed to overcome the limitations of flat storage in Vision-Language-Action (VLA) models for long-horizon robot manipulation tasks. Traditional memory-augmented architectures rely on linear or unstructured storage, which hampers efficient retrieval and generalization to novel task compositions. ECHO mimics the hierarchical organization of human experience by mapping VLA hidden states into a continuous hierarchical space using a hyperbolic autoencoder. This approach leverages hyperbolic metrics and entailment constraints to structure experience vectors into a semantic memory tree, enabling top-down retrieval. A background consolidation mechanism continuously refines the tree through geometric interpolation and structural splitting, allowing for synthetic memory generation.

When integrated into the π0 foundation model, ECHO demonstrated substantial gains over the baseline. On the LIBERO-Long benchmark, ECHO achieved a 12.8% absolute improvement in execution success rate, and it significantly enhanced compositional generalization on cross-suite unseen long-horizon tasks. Preliminary real-world experiments further validated the practical effectiveness of the approach. These results highlight the potential of structured, hierarchical memory to advance robotic autonomy, enabling robots to perform complex, multi-step tasks with higher reliability and adaptability.