---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- robotics
- isaac-sim
- assembly
- manipulation
configs:
- config_name: default
  data_files: data/*/*.parquet
---

This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).

<a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path=lukasskellijs%2Fassembly_pick_insert">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-dark-xl.svg"/>
</a>

# assembly_pick_insert

Contact-rich **pick-and-insert** demonstrations (grasp a free-standing 8 mm peg, transport it,
and insert it straight down into a tight socket) generated by a privileged scripted expert in
NVIDIA Isaac Lab. The dataset is **model-agnostic**: keys follow the LIBERO/robosuite VLA
convention so pretrained policies (pi0.5/openpi, Qwen robot-manip, Nano/DROID/Cosmos, XVLA)
can fine-tune with a trivial adapter.

- **Robot:** Franka Panda + Robotiq 2F-85 (DROID standardized platform), operational-space
  (OSC) delta-EE control with locked gains for this demo set.
- **Control / record rate:** **15 Hz** (sim 120 Hz, decimation 8). Actions are EE deltas
  between consecutive recorded frames, so the action rate equals the frame rate.
- **Instruction:** `"pick up the peg and insert it into the hole"`

## Observation / action / state layout

| key | dtype | shape | meaning / units |
|-----|-------|-------|-----------------|
| `observation.images.front` | video (H.264, CRF 20) | (224, 224, 3) | third-person/agentview RGB, uint8 |
| `observation.images.wrist` | video (H.264, CRF 20) | (224, 224, 3) | eye-in-hand RGB, uint8 |
| `observation.state` | float32 | (21,) | joint_position (7, rad) + gripper_position (1, 0..1) + EE pos (3, m) + EE quat (4, wxyz) + EE wrench (6: force N, torque N·m) |
| `action` | float32 | (7,) | EE delta pos (3, m) + delta rotation (3, axis-angle rad) + gripper target (1, normalized) |

Frames are in the robot base frame. The first 8 dims of `observation.state` are the DROID
proprio contract (`joint_position` + `gripper_position`); EE pose + wrench follow as extras.
Wrench models can slice the last 6; DROID-native fine-tuning uses `observation.state[:8]`.