---
license: other
task_categories:
- robotics
tags:
- robot-memory
- vision-language-action
- robomemarena
- robomemo
pretty_name: RoboMemo annotations for RoboMemArena
size_categories:
- 1K<n<10K
---

# RoboMemo annotations for RoboMemArena

Writer memory and transition-Status annotations for all 2,600 training trajectories
of the [RoboMemArena](https://robomemarena.github.io/) release (26 tasks × 100
demonstrations). These are the annotations used to train the RoboMemo (PicMem Full F)
policy on RoboMemArena.

This repository holds annotations only. It contains no images, trajectories, or
features. Get the trajectories from the official RoboMemArena release
([GitHub](https://github.com/OpenHelix-Team/RoboMemArena),
[Hugging Face](https://huggingface.co/datasets/RoboMemArenaBenchmark/RoboMemArena)).
The annotations were produced from the ModelScope mirror at upstream revision
`d09f63b671f362985e1e0d038d2014c2d2519488`.

## Contents

| Path | Files | Description |
| --- | ---: | --- |
| `writer/episode_<id>.json` | 2,600 | Causal Writer updates every 16 frames: private memory bank, one historical front keyframe, one current subgoal |
| `writer/dataset-index.json` | 1 | Writer model, prompt version, system prompt, evidence layout, call and error counts |
| `dataset/status/targets/episode_<id>.json` | 2,600 | Transition-Status targets per 16-step execution window (`reviewed_coarse_interval_v2`) |
| `dataset/status/episode-store.json` | 1 | Status episode index and sample positions |
| `dataset/{train,validation,status}/meta/stats.json` | 3 | Split statistics |
| `dataset/dataset.json` | 1 | Episode list, state/action dimensions, history configuration |
| `dataset/assets/benchmark/norm_stats.json` | 1 | Normalization statistics used for training |

## Annotation summary

- Dataset version: `robomemarena-modelscope-full-official-prompts-v4-public-pouring`
- Writer: `gpt-6-astra`, prompt `picmem_grounded_supported_outcomes_v2_task_aware_demo`,
  Responses API, decision stride 16
- Writer calls: 176,261 execution calls, 0 errors, 0 refusals, `complete: true`
- Status: 2,600/2,600 episodes, 1,720,027 windows, 0 missing episodes
- Task instructions use the official public RoboMemArena prompts.

Absolute paths inside the JSON files (`/data/picmem/...`) are the original
build locations. They are kept for provenance and are not needed to use the data.

## License

Annotation license: TBD. The underlying trajectories remain under the RoboMemArena
release terms.