top of page
How Aloha Works
STEP 1
Data Collection
It all starts with data. Begin your project by either collecting training data yourself or downloading pre-existing training data sets. These data sets will provide the machine learning models with “experience” to train and inform the ML model for performing tasks.
OR
Collect
Repeatedly perform individual tasks in a variety of ways. An example is trying to pick up a colored cube placed in multiple locations and from a variety of angles.
Download
To save time, you can use publically available training data provided by the community and Trossen Robotics.
Hugging Face
We have partnered with Hugging Face as Aloha's community data-sharing platform.
Data Structure
Multi-View Video Streams
Joint Speed (rad/sec)
Joint Position (rad)
Per Time Step (50Hz)
STEP 2
Train & Evaluate
Once you have a significant training data set, it’s time to begin the cyclical process of training and evaluating. You begin with the pre-loaded ACT++ Machine Learning Model, adjusting parameters and evaluating the success rate of your data inputs to task outputs.
ML Model Training
Training can be done on a variety of hardware platforms including the Aloha pre-loaded laptop, high-performance edge computing nodes, or even cloud computing.
Pre-loaded Laptop
Edge Computing
Cloud Computing
ACT++ Tip
Adjust parameters to yield better results when evaluating.
Examples:
-
VAE KL weight
-
Feedforward
layer dimension -
Hidden layer
dimension -
learning rate
-
batch size
ML Loss Curve
Physically on Aloha Kit
Virtually (Digital Twin)*
ML Model Evaluating
Training can be done on a variety of hardware platforms including the Aloha pre-loaded laptop, high-performance edge computing nodes, or even cloud computing.
* Virtual evaluating with a digital twin is only available on Aloha Stationary as of 2024.
STEP 3
Build Your Own Model
After mastering ACT++ and achieving repeatable high-success rate tasks, it's time to build upon your work and make your own machine-learning model.
bottom of page