---
license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - RoboTwin
  - RMBench
  - Mem-0
  - manipulation
  - dual-arm
size_categories:
  - n<1K
---

# Mem-0 `m1_mix` — RMBench / RoboTwin 2.0 (LeRobot dataset)

The **`m1_mix`** training dataset for the Mem-0 execution module: the five RMBench
**M1** tasks merged into a single [LeRobot](https://github.com/huggingface/lerobot)
`v2.1` dataset with **globally unique episode indices**. This is the exact data used to
train the checkpoint released at
[`qiuly/Mem-0-m1mix-RMBench`](https://huggingface.co/qiuly/Mem-0-m1mix-RMBench).

## Summary

| | |
|---|---|
| Format | LeRobot `v2.1` |
| Episodes | **250** (50 per task × 5 tasks) |
| Frames | 92,520 |
| FPS | 30 |
| Tasks | 5 (see below) |
| Robot | dual-arm (2× 7-DoF + 2 grippers) |
| Size | ~234 MB |

### Tasks (`meta/tasks.jsonl`)

`observe_and_pickup`, `put_back_block`, `rearrange_blocks`, `swap_blocks`, `swap_T`.

Episodes are laid out contiguously by task and indexed **0–249** with no collisions —
this global indexing is required by the Mem-0 MemoryBank, which groups frames by
`episode_id`; merging per-task datasets without re-indexing would alias episodes across
tasks.

## Features (`meta/info.json` is authoritative)

| key | dtype | shape | notes |
|---|---|---|---|
| `observation.image.head_camera` | video | 240×320×3 | av1-encoded, one mp4 per episode |
| `observation.state` | float32 | (16,) | left j0–j6, right j0–j6, left_gripper, right_gripper |
| `action` | float32 | (16,) | same 16-D dual-arm joint+gripper layout |

The per-episode parquet files additionally carry the Mem-0 training columns
(`subtask`, `subtask_end`, `episode_id`, timestamps/indices). See `meta/info.json` for
the full schema and `meta/episodes_stats.jsonl` for per-episode statistics.

## Layout

```
m1_mix/
├── data/chunk-000/episode_{000000..000249}.parquet     # states, actions, subtask labels
├── videos/chunk-000/observation.image.head_camera/
│   └── episode_{000000..000249}.mp4                     # head-camera RGB (av1)
└── meta/
    ├── info.json              # schema, counts, fps, paths
    ├── episodes.jsonl         # per-episode task + length
    ├── episodes_stats.jsonl   # per-episode feature stats
    └── tasks.jsonl            # task_index -> task name
```

## Usage

```python
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("qiuly/Mem-0-m1mix-dataset-RMBench")
print(ds.meta.info["total_episodes"], "episodes")
```

## Normalization

The released model uses **min-max → [-1, 1]** normalization over the 14 arm dimensions;
the exact stats (`norm_stats.json`) ship with the model repo
[`qiuly/Mem-0-m1mix-RMBench`](https://huggingface.co/qiuly/Mem-0-m1mix-RMBench), not in
this dataset. `meta/episodes_stats.jsonl` here gives raw per-episode statistics.

## License & attribution

Released under **Apache-2.0**. Generated with the **RMBench / RoboTwin 2.0** simulator;
refer to the upstream RMBench repository for simulator terms and task definitions.