ICRA-2026-RAMEN/airoa-public-filter-noeval

---
license: apache-2.0
task_categories:
- robotics
tags:
- lerobot
- lerobot-v3
- hsr
- manipulation
- mobile-manipulation
- imitation-learning
configs:
- config_name: default
data_files:
- split: train
path: data/**/*.parquet
---

# airoa-hsr-26k-filtered-v3.1

A quality-filtered subset of the **AIROA HSR** imitation-learning dataset,
converted to [LeRobot v3](https://github.com/huggingface/lerobot) format.

## Dataset summary

| Field | Value |
|---|---|
| Robot | Toyota Human Support Robot (HSR) |
| Total episodes | 3,609 |
| Total frames | 949,465 |
| FPS | 10 |
| Action dimensions | 11 |
| Cameras | hand, head |

## Action space

The `action` feature is an 11-dimensional composite vector built from
position-difference and relative-position signals:

| Index | Name | Source |
|---|---|---|
| 0 | `arm_lift_joint` | `action.state_diff[0]` |
| 1 | `arm_flex_joint` | `action.state_diff[1]` |
| 2 | `arm_roll_joint` | `action.state_diff[2]` |
| 3 | `wrist_flex_joint` | `action.state_diff[3]` |
| 4 | `wrist_roll_joint` | `action.state_diff[4]` |
| 5 | `hand_motor_joint` | `action.relative[5]` |
| 6 | `head_pan_joint` | `action.state_diff[6]` |
| 7 | `head_tilt_joint` | `action.state_diff[7]` |
| 8 | `base_x` | `action.relative[8]` |
| 9 | `base_y` | `action.relative[9]` |
| 10 | `base_t` | `action.relative[10]` |

## Observations

- `observation.image.hand` — wrist/hand camera (video, 10 fps)
- `observation.image.head` — head/overview camera (video, 10 fps)

## Quality filtering

Episodes were retained only if they passed **all** of the following checks:

**Servo motion filter** (ensures the robot actually moved):
- `std` of servo positions for joints 1, 3, 5 must each exceed `0.04`

**Video QC filter** (applied independently to hand and head cameras):
| Metric | Threshold |
|---|---|
| Blurry frame fraction | < 5 % |
| Dark frame fraction | < 3 % |
| Low-contrast frame fraction | < 1 % |
| Bright frame fraction | < 1 % |
| Frozen frame fraction | < 10 % |

## Format

This dataset follows the [LeRobot v3 dataset format](https://github.com/huggingface/lerobot).
Episode data is stored as chunked Parquet files under `data/`,
videos under `videos/`, and metadata under `meta/`.

## Filtering note (this dataset)

This dataset is `ICRA-2026-RAMEN/airoa-public-filter` with 18 episodes removed because they overlap
with the public-task evaluation datasets (`airoa-public-task-{pick/place}-{box/mug/coffee}-eval`).

Removed episode_indices (in original `ICRA-2026-RAMEN/airoa-public-filter` numbering): [181, 591, 646, 702, 771, 895, 994, 1026, 1316, 1418, 1644, 1880, 2050, 2164, 2925, 3078, 3545, 3577]

After filtering:
- total_episodes: 3609 → 3591
- total_frames: 949465 → 944488

Episode indices have been renumbered sequentially (0..3590).
Video files are copied unchanged; some frames in the original mp4 files are now orphan
(unreferenced by any episode) but this is benign for LeRobot v3 readers, which slice
videos via `from_timestamp` / `to_timestamp`.

## 32D layout note (this revision)

This revision overwrites the previous 11D layout with the 32D HSR sparse layout
required by π0.5 32D-action checkpoints (Run67/Run70 family). All state/action
columns are now `float32[32]` arranged as:

- state[0:5] = arm joints
- state[6] = hand_motor_joint (absolute gripper)
- state[11:13] = head (pan / tilt)
- state[5], [7:11], [13:32] = 0 (unused)
- action[0:5] = arm diffs
- action[6] = hand_motor_joint relative (gripper command)
- action[11:13] = head diffs
- action[13:16] = base x/y/theta relative
- action[5], [7:11], [16:32] = 0 (unused)

Episode-level filtering (18 eval-overlap episodes removed) is unchanged.