license: apache-2.0
โ๏ธ ACT-SO100-Scoop
Action Chunking Transformer checkpoint that scoops coffee beans with a custom 3-D-printed scoop end-effector.
3-D-printed scoop designed for SO-100 and SO-101 robotic arms.
Tool STL is available for download in the SO-100 Tools repository.
Demo
Note: The model had trouble completing the final step of the task, pouring the coffee beans into the cup. This is likely due to some interference issues in the training data and limited size of the dataset.
Dataset
Name | Episodes | Frames / episode | Modalities |
---|---|---|---|
370-drawn-to-caffeine-coffee-scooper | 42 | ~450 | RGB 1080x1920, proprio 5-DoF, gripper state |
Training Details
See run details on wandb for more information: wandb run.
Hyper-parameter | Value |
---|---|
Chunk size | 100 |
Dim Feedforward | 3200 |
Dim Model | 512 |
Dropout | 0.1 |
Feedforward Activation | ReLU |
Decoder layers | 1 |
Encoder layers | 4 |
Attention heads | 8 |
VAE Encoder layers | 4 |
Batch size | 32 |
Optimizer | AdamW, lr = 1e-5 |
Citation
If you use this checkpoint in your work, please cite the following:
@misc{Rohling2025ACTSO100Scoop,
author = {Kevin Rohling},
title = {ACT Checkpoint for Coffee Scooping on SO-100},
year = {2025},
howpublished = {\url{https://huggingface.co/kevin510/ACT-SO100-Scoop}}
}
- Downloads last month
- 6
Inference Providers
NEW
This model isn't deployed by any Inference Provider.
๐
Ask for provider support