ShubhamK32/so101_declutter_v1

---
license: apache-2.0
task_categories:
- robotics
tags:
- lerobot
- so101
- manipulation
- teleoperation
- visuomotor
- decluttering
- real-robot
- pick-and-place
- object-retrieval
- language-conditioned
pretty_name: SO-101 Targeted Pick-and-Place — Blue Duster Retrieval v1
size_categories:
- n<1K
---

# SO-101 Targeted Pick-and-Place — Blue Duster Retrieval v1

A multi-view visuomotor teleoperation dataset collected on a 6-DoF SO-101 robotic arm for space decluttering tasks. Recorded using [LeRobot](https://github.com/huggingface/lerobot).

## Camera View Samples

| Top View | Wrist View |
|----------|------------|
|  |  |

## Task Description

The robot learns to identify and retrieve a specific target object — a blue duster — from a cluttered tabletop containing multiple distractor objects, and place it into a bin.

Spatial distractors are deliberately injected across episodes to prevent the policy from learning visual shortcuts (e.g. positional bias or background cues), forcing it to learn genuine object-level recognition and targeted grasping.

## Setup

- **Robot:** SO-101 6-DoF leader-follower teleoperation
- **Cameras:** Dual-view — fixed top-view + wrist-mounted egocentric
- **Frequency:** 60Hz joint telemetry with synchronized RGB streams
- **Pipeline:** Asynchronous data collection with kinematic filtering to remove demonstration jitter

## Camera Views

- `observation.images.topview` — Fixed overhead camera. Performs better on unoccluded pick-place tasks.
- `observation.images.wristview` — Egocentric wrist-mounted camera. Better handles overlapping objects and cluttered scenes.

## Dataset Structure
```
data/ # Joint states and actions (Parquet)
videos/
observation.images.topview/ # Fixed overhead RGB
observation.images.wristview/ # Wrist-mounted egocentric RGB
meta/ # Episode metadata, task descriptions, stats
```

## Trained Models

| Policy | Checkpoint | Steps |
|--------|-----------|-------|
| ACT (Action Chunking Transformer) | [ShubhamK32/act_so101_declutter](https://huggingface.co/ShubhamK32/act_so101_declutter) | 100,000 |
| SmolVLA (Vision-Language-Action) | [ShubhamK32/smolvla_so101_declutter](https://huggingface.co/ShubhamK32/smolvla_so101_declutter) | 20,000 |

## Citation
```bibtex
@dataset{kanitkar2025so101declutter,
author = {Shubham Kanitkar},
title = {SO-101 Space Decluttering Dataset v1},
year = {2025},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/ShubhamK32/so101_declutter_v1}
}
```