RoboticsRobot planningStory 06
Robots turn long memories into short plans
What changedMemory as Plans compresses completed language-and-image segments into compact plans, then gives a fixed-context executor the current world model, action plan and progress. The authors report 78% success on real-robot tasks while latency stays approximately constant as history grows. Those benchmark and hardware results remain author-reported and task-specific.
The useful part
Why it matters
Long-horizon robots need memory without making every next action slower and more expensive as the task history expands.
Worth doing
What to do next
Compare failure recovery and latency against a growing-context baseline on tasks longer than the paper’s demonstrations.
Keep in mind
Good to know
The state-of-the-art and success claims are author-reported and may not transfer beyond the evaluated tasks.
Evidence