🌐 Project Page | 📄 arXiv | 🎵 RoboBeats | 🤗 Models | 📊 SSS-Bench | 📝 Citation

• [2026-11-10] ⚡ Real-time RoboBeats: per-frame inference with cache reuse, conditions of your own, and a local web demo for your camera or videos (Live Demo).
• [2026-11-03] 🎉 Released RoboBeats: inference code and model weights (0.8B, 2B, 4B). Released SSS-Bench: the benchmark, evaluator and reproduction scripts.
• [2026-10-30] 📄 The paper is on arXiv.

LIFE (Lifelong Intelligence from Experience) is a framework for robots that keep observing, remember selectively, and turn experience into new capabilities.

RoboBeats is a small, strictly causal VLM gate. It watches a 1 fps video stream and scores every frame. In event discovery mode, the score is the probability that a new event has just occurred. Only the frames around detected events are written into memory.

Each model is a LoRA adapter; the base model is downloaded automatically.

git clone https://github.com/AgibotTech/LIFE.git && cd LIFE
pip install -r requirements.txt

The default torch==2.13.0 wheel is built for CUDA 13 and needs NVIDIA driver ≥ 580. With a CUDA 12.x driver, switch to the CUDA 12.9 builds:

pip install --force-reinstall torch==2.13.0 --index-url https://download.pytorch.org/whl/cu129
pip install --force-reinstall --no-deps torchvision==0.28.0 torchaudio==2.11.0 --index-url https://download.pytorch.org/whl/cu129

python -m robobeats my_video.mp4 --gate-model superbigsaw/RoboBeats-2B --out probs.json

Run RoboBeats on your camera, your own video, or the bundled example, in real time, with conditions of your own:

pip install "fastapi>=0.110" "uvicorn[standard]>=0.29"
python -m robobeats.serve --gate-model superbigsaw/RoboBeats-2B
# open http://localhost:8000

Models, Python API, condition tracking, the real-time LiveGate and the readout: robobeats/README.md. Condition-tracking benchmarks (Charades-STA, StreamingBench, SPOT-Bench): benchmarks/README.md.

SSS-Bench (Sparse-event Streaming Suite) asks questions about videos up to 30 min long, under a strictly causal streaming protocol. The answers depend on details that are sparse in time. Videos are not redistributed; see sss_bench/README.md for download, data format, grading and reproduction.

Evaluate your model. Write predictions as {"id": ..., "answer": ...} JSONL, then:

python sss_bench/scripts/check_videos.py $SSS_DATA_ROOT
python -m sss_bench.evaluate --pred preds.jsonl --data-root $SSS_DATA_ROOT \
  --judge-base-url http://127.0.0.1:8000/v1 --judge-model Qwen3.8-27B --out results/my_model

Reproduce RoboBeats (gate → 27B memory writer → 27B reader → grading, 8 GPUs):

SSS_DATA_ROOT=/path/to/videos LLM_MODEL=/path/to/Qwen3.8-27B GATE=2B bash sss_bench/scripts/run_robobeats.sh

Code: Apache 2.0. SSS-Bench data and rubrics: CC BY-NC 4.0. Source videos stay under their original licenses.

@article{che2026learning,
  title         = {Learning to Live: A Step Toward Lifelong Embodied Intelligence},
  author        = {Che, Lirong and Wang, Peidong and Su, Yimeng and Wang, Jiong and Chen, Yang and Fu, Wenbo and Huang, Zhenyu and Xu, Tiancheng and Yang, Yuzhe and Xie, Dongli and Cao, Xu and Wang, Chuang and Liu, Daqi and Su, Jian},
  journal       = {arXiv preprint arXiv:2611.00180},
  eprint        = {2611.00180},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  year          = {2026},
  url           = {https://arxiv.org/abs/2611.00180}
}