metadata
license: other
license_name: gemma
license_link: https://ai.google.dev/gemma/terms
pipeline_tag: robotics
tags:
- robotics
- vision-language-action
- robot-memory
- pi0.5
- robomme
- arxiv:2611.14488
Is Your Robot's Memory Just a Clock? — Checkpoints
Policies trained for the paper Is Your Robot's Memory Just a Clock? Temporal Shortcuts in Memory-Augmented VLA Policies.
Paper: arXiv:2611.14488
Code: github.com/bennidict23/memory-clock
All models are trained from the pi0.5 base checkpoint with the released RoboMME MME-VLA recipe (FrameSamp-Modul memory, 30k steps, global batch 64) and evaluated on RoboMME.
| Folder | Model | Used in |
|---|---|---|
clock/ |
FrameSamp-Modul memory with time stamps | Figs. 6, 8b |
noclock/ |
the same without time encoding (use_pos_emb: false) |
Fig. 6 |
clock-retrain/ |
independent Clock retraining, control for pause augmentation | Fig. 8a |
pauseaug/ |
Clock trained with pause augmentation | Fig. 8a |
pauseaug-weak/ |
weaker pause-augmentation variant | Pause-augmentation ablation |
Each folder holds 29999/{params,assets} (openpi / orbax format) and history_config.txt.
Usage
git clone https://github.com/bennidict23/memory-clock.git && cd memory-clock
bash install.sh && bash download.sh # fetches these checkpoints into checkpoints/<name>/29999
bash run_eval.sh --model clock --task PickXtimes --event pause --len 100
License
The weights are derived from pi0.5 / PaliGemma and are subject to the Gemma Terms of Use.
Citation
@misc{zhang2026robotmemoryclock,
title={Is Your Robot's Memory Just a Clock? Temporal Shortcuts in Memory-Augmented VLA Policies},
author={Chubin Zhang and Bo An},
year={2026},
eprint={2611.14488},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2611.14488}
}