Physical Intelligence Unveils MEM Architecture for Long-Horizon Robotic Memory

ycombinator · x · 2026-08-08

Physical Intelligence (π) presented MEM (Multi-Scale Embodied Memory) at the Y Combinator Paper Club, targeting the memory bottleneck of Vision-Language-Action (VLA) models in long-horizon tasks.

Core Problem: While current robotic foundation models master individual skills like washing pans, they struggle with multi-stage tasks like cleaning a kitchen due to a lack of coherent memory.
Memory Challenge: Retaining full observation history causes context explosion, but discarding it leads to "causal confusion," exacerbating spurious correlations in imitation learning.
Solution: The MEM architecture introduces multi-scale long and short-term memory for VLA models, enabling robots to track task progress, remember out-of-view objects, and recall past trial-and-error outcomes.

Original post →

Physical Intelligence Unveils MEM Architecture for Long-Horizon Robotic Memory

More from Embodied