Gated Memory Policy (GMP)
Gated Memory Policy (GMP) is a visuomotor policy designed for robotic manipulation tasks that learns both when and what to recall from historical observation data. It addresses the challenges of distribution shift and overfitting often encountered when extending observation histories.
Description
GMP introduces several key mechanisms to handle varying memory requirements in robotic tasks:
- Memory Gate Mechanism: A learned gate that selectively activates history context only when necessary, improving robustness and reactivity.
- Latent Memory Representations: A lightweight cross-attention module that constructs efficient representations of historical information.
- Diffusion-based Robustness: The policy injects diffusion noise into historical actions during training and inference to mitigate sensitivity to noisy or inaccurate histories.
The model achieves significant performance improvements on the MemMimic benchmark for non-Markovian tasks while maintaining competitive performance on Markovian tasks in RoboMimic.
