NextFin News — Chinese robotics company AgiBot has released LIFE (Learning to Live: A Step Toward Lifelong Embodied Intelligence), a framework that integrates continuous perception, long-term memory and capability reuse so robots can keep gathering environmental information between tasks and draw on historical records for later decisions.
The system comprises three main components: the lightweight decision model RoboBeats for condition tracking and event detection, the Plastic Memory module that stores selected frames and structured records, and the Explore-Distill-Mount mechanism that converts problem-solving methods into reusable tools and skills. In FindingDory’s 597-query tests, reusing existing tools and skills raised task success rates from 37.5% to 65.8%. Additional benchmarks showed RoboBeats-4B lifting StreamingBench active-output accuracy from 21.2% to 64.0% and SSS-Bench question-answering accuracy from 40.9% to 60.5%.
AgiBot has open-sourced the related paper, project page, GitHub code and RoboBeats model weights in 0.8B, 2B and 4B sizes, along with evaluation resources, enabling developers to study continuous robot perception and experience reuse.
Explore more exclusive insights at nextfin.ai.
