AI Learning in Humanoid Robots
We can divide learning of AI, which is called the Machine Learning (ML) into the following categories:
Supervised Learning
Unsupervised Learning
Reinforcement Learning
Hybrid / Newer Methods
A humanoid robot’s learning, although still falls under the main categories above, is quite different than an AI model that exists only as an algorithm because in the case of humanoid robot we have the data input that comes through physical sensors. In this case therefore, we have what is called the “physical AI”. The sensors we talked about in humanoid robots sensors post here, all come into play, to give robot a coherent understanding and awareness of itself and its surroundings.
Now let’s talk about each learning category….
Supervised Learning
In supervised learning, AI learns from human input, where humans show AI some data with given conditions / labels. For example teaching AI how to identify spam emails, by human identification of a great number of emails first, or teaching AI to diagnose a certain disease by doctors labeling healthy x-rays and x-rays with disease, for AI to learn and later identify them itself.
Supervised learning models:
Support Vector Machine (SVM)
Convolutional Neural Network (CNN) (CNN is most commonly in supervised learning but can also be used in unsupervised or reinforcement learning. It is one of the most commonly used types of deep learning models.)
In supervised learning, we solve either a regression or classification problem using SVM, CNN, or many other algorithms.
For a humanoid robot, supervised learning examples can be teaching hand gestures or objects by telling each gesture’s or object’s meaning.
Unsupervised Learning
In contrast to supervised learning where the models are trained by the data that were labeled by humans, in unsupervised learning, the models seek to find patterns, structures in unlabeled data and to be able to group them as necessary.
Common techniques include
- clustering, such as K-means or DBSCAN
- dimensionality reduction, such as PCA or t-SNE
- anomaly detection.
For example a humanoid robot can be set to learn certain tasks just by watching humans. It identifies patterns and relations in movements, postures, objects and much more. This is very similar in essence to how humans learn. Through this for example, the robot may learn how to manipulate objects to accomplish a certain task.
Reinforcement Learning
Reinforcement learning is learning by trial and error, where the AI is set to maximize rewards, which reinforces the actions that are towards a given goal.
Algorithms used include Deep Q-Learning or Proximal Policy Optimization (PPO).
A humanoid robot can learn how to walk by itself, through reinforcement learning. For example, every certain distance of progress and every move that is more energy efficient can be rewarded and after many trials, the robot will be able to learn walking efficiently.
Hybrid / Modern Learning Methods
In addition to three conventional categories listed above, there are methods that should be grouped differently, which are mainly newer approaches.
Self Supervised Learning:
In self supervised learning, the model creates its own labels from input data. For example by watching a lot of videos, a humanoid robot can get a sense of how objects interact and move in real world.
Imitation / Demonstration Learning:
Here a task is first shown to the robot by human, which the robot observes and learns. It can be counted as a shortcut to reinforcement learning.
Online / Continuous Learning:
This is learning and adapting on an ongoing basis, to the changing environment and input conditions.
A Final Note – AI > Machine Learning > Deep Learning
You probably heard the term “Deep Learning” a lot and might be wondering how it fits into the categories mentioned above. The categories above refer to how AI learns, which is called machine learning (ML), in other words ML is a subset of AI. During machine learning different methods can be used. Deep learning is a learning method based on complex and powerful models that can be employed in each group listed above. Therefore deep learning is a subset of machine learning, which is used as one of the tools in machine learning that can fit into any group listed and described above. Deep learning is a very useful method to recognize complex patterns within large amounts of data, and is used extensively in many useful application from autonomous driving to speech or face recognition. See our post on deep learning in Robot Magazine.
By: A. Tuter
——————————————————
Terms of use:
Copying or republishing of any type of content from this website is not allowed without written permission from us. We make dated records of our posts, keep originals of our images.
The content in this website may contain errors, inaccuracies or be incomplete. User assumes all liability as a result of usage of content in our site. Also see our Terms page for more details.