Radar / AI robotics / LIFE
LIFE: AgiBot's framework that gives robots a heartbeat, memory, and reusable skills
AgiBot (Zhiyuan) open-sourced LIFE, a lifelong embodied-intelligence framework pairing the tiny RoboBeats decision model (0.8B/2B/4B) with plastic memory and an explore-distill-mount skill-reuse loop; task success rose from 37.5% to 65.8% on FindingDory.
Why it matters
First open framework that treats a robot's life between tasks as first-class: continuous perception, selective memory, and turning solved problems into mounted tools — plus a live camera demo that runs on one RTX 5090.
What you could build with it
A robotics team building a warehouse or household robot could run RoboBeats as an always-on event watcher on the robot's camera feed, so the robot notices and remembers changes between jobs and reuses previously distilled skills instead of re-solving the same task from scratch.
Does it hold up?
Early but runnable: code, weights and benchmark shipped together on launch day with a real-time demo, but evidence so far is the team's own FindingDory numbers and a 24-hour single-robot test — no independent deployments yet.
Built with LIFE
- 智元发布终身具身智能框架 LIFE:给机器人接上'心跳'与记忆pedaily.cn · Chinese report on the LIFE launch: RoboBeats, Plastic Memory and EDM, with paper, project page, code and SSS-Bench all open.
- Zhiyuan has released the lifelong embodied intelligence framework LIFE, with the task success rate increasing from 37.5% to 65.8%substack · English summary of the launch, the FindingDory 597-query result, and the open RoboBeats weights and SSS-Bench resources.
- ZhiYuan Unveils LIFE Framework for Lifelong Embodied Intelligence and RoboBeats Resourcesrobottoday.com · Industry briefing on LIFE and the RoboBeats model family as a step toward robots that learn continuously.
- AgibotTech/LIFEgithub · 11 ★ · Official code: RoboBeats inference with a live camera demo, SSS-Bench benchmark and evaluator, Apache 2.0.
Learn more
First spotted on github: source.
More AI robotics
RuView: see people through walls with WiFi, no camerasOpen-source Rust system that turns ordinary WiFi signals into real-time pose estimation, vital-sign monitoring, and…robotics · JEV 0.76Runway Praxis-1: open-weight robot model that learned to act from videoRunway released Praxis-1, a 27B-parameter world-action model, with open weights, inference and fine-tuning code under…robotics · JEV 0.66NVIDIA GR00T 2: a world-action robot model that learns from human videoNVIDIA's open frontier world-action model predicts future video and robot actions together, hitting 71% one-shot…robotics · JEV 0.66RoboParty RP1: a full-stack open-source humanoid robot for embodied AI researchRoboParty's 1.3-metre humanoid, launched 16 October and back in coverage on 2 November, comes in 24-DoF standard and…robotics · JEV 0.62Gemini Robotics 2.5: DeepMind's most capable robot model, tested on a live Mercedes lineGoogle DeepMind introduced Gemini Robotics 2.5, its most capable vision-language-action model, reporting 30-100%…robotics · JEV 0.61T-Rex: a robot manipulation AI that reacts to fingertip touchPanasonic, UC Berkeley, NVIDIA and Stanford built T-Rex (Tactile-Reactive Dexterous Manipulation), a VLA model that…robotics · JEV 0.61
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.