---
license: mit
task_categories:
- robotics
tags:
- LeRobot
- teleoperation
- dual-arm
configs:
- config_name: default
  data_files: data/*/*.parquet
---

# nero_bag_box — dual-arm bag/box manipulation (LeRobot v2.0)

Teleoperated demonstrations on the AgileX NERO dual-arm rig (two 7-DoF
S-R-S arms), collected with a Pico 4 WebXR teleop rig and cut into
single-task clips. Ready for π0.5 / openpi-style fine-tuning.

## Tasks

- **Put the bag into the box**
- **Pour the bag out of the box**

## Features

| key | shape | notes |
|---|---|---|
| `observation.state` | float32[16] | left joint1–7 (rad), left gripper (0–1, measured), right joint1–7, right gripper |
| `action` | float32[16] | the **commanded** teleop targets for the same tick (IK joint targets + gripper command), same layout |
| `observation.images.base` | 640×480 h264 @ 15 fps | scene camera |
| `observation.images.wrist_left` | 640×480 h264 @ 15 fps | left wrist |
| `observation.images.wrist_right` | 640×480 h264 @ 15 fps | right wrist |
| `observation.images.depth_wrist_left` | 640×480 HEVC gray12le @ 15 fps | aligned left-wrist metric depth |
| `observation.images.depth_wrist_right` | 640×480 HEVC gray12le @ 15 fps | aligned right-wrist metric depth |

15 fps; every frame is one fresh camera observation with the joint state
sampled at the same tick. Actions are the next measured robot state. Wrist
depth is stored as LeRobot depth-video features next to the corresponding RGB
videos with identical episode and frame indices. Raw uint16 depth is converted
to metres using each episode's RealSense scale, logarithmically quantized to
12 bits over 0–3 m, and encoded losslessly as HEVC `gray12le`; its quantizer
parameters are recorded in `meta/info.json` for dequantization on load. A raw
zero remains zero and represents invalid/no-return depth. The fixed base camera
runs a stereo-IR profile and therefore has no depth stream.

Collected & converted 2026-08-08.