README.md

---
license: other
pretty_name: Unitree G1 Dynamic Task (training data)
task_categories:
- robotics
tags:
- robotics
- lerobot
- unitree
- g1
- real-robot
- conveyor
configs:
- config_name: Dynamic_Cup
  data_files: "Dynamic_Cup/data/*/*.parquet"
- config_name: Speed_Cup_20
  data_files: "Speed_Cup_20/data/*/*.parquet"
- config_name: Speed_Cup_30
  data_files: "Speed_Cup_30/data/*/*.parquet"
- config_name: Speed_Cup_40
  data_files: "Speed_Cup_40/data/*/*.parquet"
- config_name: Speed_Cup_50
  data_files: "Speed_Cup_50/data/*/*.parquet"
extra_gated_prompt: "Private real-robot training data (Unitree G1, conveyor pick tasks). Access is granted manually by the owner; please do not redistribute."
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
  I agree not to redistribute this dataset: checkbox
---

# Unitree G1 Dynamic Task: training data

Real-robot demonstrations on a Unitree G1 picking cups from a moving conveyor. This is the exact
training data used for the per-task Long-WAM / Fast-WAM fine-tunes. There are 5 tasks,
250 episodes and 69750 frames in total, recorded at 30 Hz.

| task | instruction | episodes | frames | duration | action dim | cameras |
|---|---|---|---|---|---|---|
| `Dynamic_Cup` | pick up the blue cup and the green cup from the moving conveyor and stack the green cup into the blue cup. | 50 | 28643 | 15.9 min | 16 | `color_0`, `color_2`, `color_3` |
| `Speed_Cup_20` | pick up the cup. | 50 | 11502 | 6.4 min | 8 | `color_0`, `color_3` |
| `Speed_Cup_30` | pick up the cup. | 51 | 10344 | 5.7 min | 8 | `color_0`, `color_3` |
| `Speed_Cup_40` | pick up the cup. | 50 | 10462 | 5.8 min | 8 | `color_0`, `color_3` |
| `Speed_Cup_50` | pick up the cup. | 49 | 8799 | 4.9 min | 8 | `color_0`, `color_3` |

## Format

Each task is a standalone **LeRobot v2.1** dataset (`codebase_version: v2.1`, `robot_type: unitree_g1`):

```
<task>/meta/{info.json, episodes.jsonl, episodes_stats.jsonl, tasks.jsonl}
<task>/data/chunk-000/episode_NNNNNN.parquet              action, observation.state, timestamp, indices
<task>/videos/chunk-000/observation.images.<cam>/episode_NNNNNN.mp4   480x640, h264, 30 fps
<task>/fastwam_dataset_stats.json                         normalization stats used in training
<task>/conversion_manifest.json                           which source dims/cameras were kept
```

Cameras: `color_0` = head camera (left eye), `color_2` = left wrist camera, `color_3` = right wrist camera.

- **Dynamic_Cup**: 16-D action and state `[left_arm.qpos 0-6, right_arm.qpos 0-6, left_ee, right_ee]`, 3 cameras.
- **Speed_Cup_20/30/40/50**: only the right arm moves, so the data is right-arm only. It uses an 8-D action and state
  `[right_arm.qpos 0-6, right_ee]` and 2 cameras (head plus right wrist).

`action` is the absolute joint-position target and `observation.state` is the measured joint position. Both use the
units of the original Unitree export.

## Provenance

These files were converted from the original Unitree LeRobot v3.0 recordings, which have a 19-D action, a 51-D state
and 4 cameras. The conversion made these changes:

- dropped the 3 always-zero body dimensions
- dropped the second head camera (`color_1`)
- reduced Speed_Cup to the right arm only
- replaced the placeholder instruction "pick up cube." with the task instructions in the table above

Episode lengths and frame timing are unchanged.

## Loading

Each task folder is its own dataset. Download the task you need, then point LeRobot at that folder:

```bash
huggingface-cli download AaronHuangWei/Unitree-G1-Dynamic-Task --repo-type dataset --include "Dynamic_Cup/*" --local-dir g1_dynamic_task
```

```python
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset  # a lerobot version that reads v2.1
ds = LeRobotDataset("g1_dynamic_task/Dynamic_Cup", root="g1_dynamic_task/Dynamic_Cup")
```