Embodied Memory Enables Robots to Perform Long Tasks

StanfordAILab · x · 2026-07-13

This post introduces a project for embodied robots titled MEM: Multi-Scale Embodied Memory for Visual Language Action Models, which will be presented at an RSS 2026 workshop.

The core idea is to add memory capabilities to PI policies, allowing robots to handle longer-horizon real-world tasks. Mentioned demos include prepping recipe ingredients, making grilled cheese, and cleaning the kitchen.

The speaker notes they will attend remotely but will provide more details in a subsequent thread.

Original post →

More from Embodied

Embodied channel →