Gated Memory Policy

A research team from Columbia University has introduced Gated Memory Policy (GMP), a novel AI architecture designed to solve a fundamental problem in robotic manipulation: how to effectively use memory. Unlike traditional approaches that simply feed robots longer observation histories—which often causes performance drops due to distribution shift and overfitting—GMP intelligently learns both when to recall past information and what specific information to recall. This is achieved through a learned memory gate that activates historical context selectively, improving the robot's reactivity and robustness by preventing it from being bogged down by irrelevant past data.

To efficiently manage what to recall, GMP employs a lightweight cross-attention module that constructs compact, effective latent memory representations. The researchers further enhanced the system's resilience by injecting diffusion noise into historical actions during training and inference, making the policy less sensitive to noisy or imperfect memory inputs. Tested on their newly proposed non-Markovian benchmark, MemMimic, GMP outperformed standard long-history baseline methods by an impressive 30.1% in average success rate. Crucially, it also maintained competitive performance on standard Markovian tasks from the RoboMimic benchmark, proving its versatility.

The team has made all code, data, and real-world deployment instructions publicly available, providing a significant new tool for advancing robotic learning. This work addresses a key bottleneck in creating robots that can perform complex, multi-step tasks requiring an understanding of past interactions, moving beyond simple reactive behaviors.