C-LEARN – the robot’s journey from apprentice to master
MIT is currently undertaking research to improve robot learning, making it universally accessible. The C-LEARN project could well revolutionise this sector by enabling robots to adapt to different situations and transfer their knowledge to other machines.
Today, robot learning uses two methods:
- Teaching from demonstration. By reproducing what they see, robots can perform a task quickly.
- Teaching through programming. This method requires an expert programmer.
Both methods have their drawbacks. The first does not allow the robot to adapt to constraints or unplanned changes. The second requires coding, and by proxy, a qualified programmer. MIT wanted to overcome these problems by experimenting with C-LEARN.
This is a bi-phase method for teaching the robot. The first phase consists in transferring a knowledge base to the robot, made up of the objects it will later manipulate. Thanks to this knowledge base, it can know how a particular object must be used.
In the second phase, an operator (who may also be a novice), uses a simple 3D interface to show the robot how the task must be carried out, as with teaching from demonstration.
The robot will then combine these two knowledge sets in order to identify the best method for performing the task. It is then able to better understand how the task must be executed and how to adapt to constraints. Where an obstacle might arise, it will know how to divert its course of action as it understands that a screw, for instance, must be put in a specific place in a specific way. It does not learn one unique way to perform an action.
Other than saving time and being more efficient, this information can be transferred to other robots so that they can undertake the same task even if they are not the same type of robot. MIT performed a test with a very compact, two-armed robot named Optimus. The transferred knowledge was then passed on to another robot, Atlas, a humanoid weighing in at nearly 400 lbs.
Thanks to C-LEARN, even non-coders can teach a robot, which will then, in turn, go on to transfer this knowledge.
The following video shows a demo of C-LEARN in action:
What are the impacts and consequences?
At present the tests are still underway, and a robot trained using C-LEARN is still at its best when assisted by a human. Today, a robot can successfully complete 87.5% of its tasks when unsupervised. When an operator is present to correct errors caused by measurement defects in the sensors, the success rate will reach 100%.
Ultimately, C-LEARN is getting closer and closer to human learning. We learn by observing an action and by linking it back to what we know already.
Such technology can radically change the way in which a robot functions. It will go from being a model that knows how to perform a task, to a model that understands how to perform a task. The direct benefit from this is greater flexibility when compared to teaching by demonstration, and greater speed when compared to teaching with programming.
One concrete application is the use of emergency robots. In a complex environment, they are able to adapt to the situation in order to perform their tasks.
Another benefit is that the wider the knowledge base, the more capable the robot is in performing its tasks. Machine Learning and Big Data have an important role to play here in the transfer of information in large-scale projects.
For simple tasks, the C-LEARN method requires little data and little knowledge from the operator who will be teaching the robot the task at hand. Once perfected, the MIT method could very quickly become an international standard in the world of robotics. Thanks to C-LEARN, anyone can transfer a skill-set to a robot, after which it can go on to transfer to others.