AgiBot releases open-source LIFE framework for continuous robot perception

AgiBot has open-sourced the LIFE (“Learning to Live: A Step Toward Lifelong Embodied Intelligence”) framework, aimed at keeping robots perceptive and remembering between tasks. The paper was posted to arXiv on October 30 and the code released under Apache-2.0 on GitHub, with inference weights for RoboBeats models (0.8 B, 2 B, 4 B parameters) added on November 3 link.

LIFE combines a multimodal large model-driven agent with three components: RoboBeats, a lightweight decision model that reads true/false condition scores (“heartbeat”); Plastic Memory, which stores raw frames and JSON records for later reanalysis; and EDM (Explore-Distill-Mount), which converts successful problem-solving methods into reusable tools without weight updates.

In internal tests, RoboBeats-2B processes a new frame in ~75 ms on a single RTX 5090, and RoboBeats-4B lifts StreamingBench accuracy from 21.2 % to 64.0 %. On SSS-Bench (171 videos, 618 questions) text-memory QA rose from 40.9 % to 60.5 %, and EDM-mounted tools raised FindingDory success from 37.5 % to 65.8 %.

A 24-hour case study with an AgiBot A3 Ultra robot built a day-long memory, scoring 29.5/30 with an 80.6-second average response, compared with 28.5 and 554.7 seconds for a Codex-based baseline under identical conditions. The team notes multi-day deployment is next, and results have not yet been independently reproduced.