China Hot Tech @ChinaHotTech
← All posts
Robotics · Open Models · Embodied AI

AgiBot Opened LIFE, a Between-Task Memory Stack — The Code Is Apache, the RoboBeats Weights Are Not

Shanghai’s AgiBot (智元) has published LIFE — Lifelong Intelligence from Experience — under the paper title Learning to Live: A Step Toward Lifelong Embodied Intelligence. The technical report is on arXiv as 2611.00180 (dated October 2026); RoboBeats inference code and 0.8B / 2B / 4B weights, plus the SSS-Bench evaluation pack, were posted on November 3, and Chinese coverage such as IT之家 landed on November 11. The pitch is not a new humanoid SKU. It is a stack for what a robot does between tasks: keep watching, write only some frames into memory, and turn a solved problem into a reusable tool. You can clone the repo today. The catch for a product team is in the licenses: the code is Apache 2.0; the RoboBeats adapters on Hugging Face, and the SSS-Bench data, are CC BY-NC 4.0. Asterisk paper and vendor benches.

· ~8 min read · China Hot Tech

Official AgiBot LIFE project-site still: a bipedal humanoid at an office pantry counter, reaching toward a snack bag beside a beverage fridge branded 魔盒 and 智元

Official still from the LIFE project site’s demo cover (assets/s2-demos/s2-scene-cover.webp at agibotgeneral.github.io/life-site) — company-published footage of a humanoid in a real office pantry, not an image we generated.

Most robot stacks are organized around a single instruction: perceive, plan, act, then reset. The physical world does not reset. Elevator doors open, a visitor arrives, a cup changes hands, all while the policy is between goals. LIFE, from AgiBot’s HI / 远征 research team, is an attempt to treat that in-between time as the product. The report and the project site describe three cooperating pieces around a multimodal agent: RoboBeats (a light streaming gate), Plastic Memory (a flat store of selected images plus JSON records), and Explore–Distill–Mount (EDM), reused from the team’s earlier ReSight work, which turns a working procedure into a validated tool.

RoboBeats is the part you can actually download as a model. It is a family of LoRA adapters on Qwen3.5-0.8B / 2B / 4B, hosted as superbigsaw/RoboBeats-2B and the matching 0.8B and 4B cards. The gate reads a 1 fps (1 Hz) causal video stream and, instead of writing a caption for every frame, reads true/false scores for natural-language conditions — “has the elevator door opened?”, “has a new event just occurred?”. Multiple conditions share the visual tokens and cached history and keep their own prediction state. The 2B card’s example selector opens an event when the score is above 0.5 and closes it below 0.3. Only frames around those events are meant to be written into memory. The site’s official office demos overlay exactly this: a first-person fisheye next to a third-person shot, with the monitored condition and the triggered action on screen.

Official RoboBeats architecture diagram: a shared-visual multi-condition causal gate on a 2B VLM, per-condition event readouts, condition tracking, and an event-discovery path into memory question answering

Official RoboBeats architecture figure from the LIFE project site (assets/s2-research/robobeats.webp), corresponding to the paper’s Figure 4. Company diagram, not artwork we drew.

Plastic Memory is deliberately boring, which is the point. Selected frames are kept as files, with a JSON record for time, optional pose, and extracted fields for people, objects, and events. Because the original image is retained, a later question can harvest a detail — a room number on a door, whose cup that was — that nobody extracted when the frame was first saved. EDM sits on top when the current toolbox is not enough: an encoding agent explores a procedure, distills out scene-specific hard-coding, and mounts the result so the execution agent can call it later. The paper is explicit that this is skill and tool construction, not a requirement that the base model’s weights be updated after every task.

Can you use this from outside China?

The repo, yes — the weights, only if CC BY-NC works for you. AgibotTech/LIFE is Apache 2.0. The three RoboBeats adapters on Hugging Face are labeled cc-by-nc-4.0, and the README says SSS-Bench data and rubrics are also CC BY-NC 4.0; source videos stay under their original providers’ licenses and are not in the same dump. You will also need the Qwen3.5 base checkpoints the LoRAs sit on, and a CUDA GPU: the README’s default torch==2.13.0 wheel is built for CUDA 13 and asks for NVIDIA driver ≥ 580; a CUDA 12.9 install path is documented for 12.x drivers. We did not run the stack. This is research code plus a non-commercial gate, not a robot you can add to a cart, and not a license we can interpret as permission to ship a commercial product on top of the adapters.

What the paper actually measures

The useful numbers are protocol-bound. Treat them as the authors’ results, not as something we reproduced.

Official LIFE figure of six real-robot memory question-answering cases, including two cups and their owners, a cushion handover, a jacket left on a chair, an early-morning visitor, the first identified arrival, and the end of a lunch break, each with observed frames and the evidence used to answer

Official six-case figure from the LIFE project site (assets/s2-research/robot-cases.webp, paper Figure 13): offline memory built from a day of A3 Ultra office video, then used to answer later questions. Company figure, not our screenshot of a run.

Honest caveats

For an overseas engineer, the clean split is this. If you want a small, causal gate that watches 1 fps video and decides when a frame is worth remembering, the code and the three LoRAs are public, the serving command is one line, and the research question — what should a robot do with the hours when nobody is issuing a task — is a real one. If you want to productize that gate, start with the CC BY-NC label, not with the Apache badge on the repo. And if a social post tells you AgiBot has “solved lifelong learning,” the paper’s own 24-hour section is the corrective: a careful offline case study, not a robot that has been living in a room for a month.

Sources