Self-Monitoring Robotic Learning has proven effective in training artificial agents on individual tasks. Yet, most existing enhanced learning approaches are still ineffective when teaching multipurpose robots capable of completing several tasks requiring different skills.
To overcome this challenge, a team of researchers from the University of California, Berkeley, has developed a new approach to enhance learning that can teach robots how to adapt their behavior based on the task required. This process they summarized in a research paper published in the arXiv archive. Presented at the Conference on Robot Learning robotic learning conference.
Robots can Re-Employ
This enhanced learning approach allows robots to automatically reach and practice actions over time and learn which actions can be used in a particular environment. So that robots can re-employ the knowledge they have acquired and apply them to complete new tasks assigned to them by human users.
Artificial agents: Smart, self-propelled entities that achieve specific goals in a particular environment based on sensors and engines. These entities may be complex or straightforward and may be able to learn or use knowledge to achieve their goals. The thermostat is an example of an artificial agent.
ArXiv Archive: An archive of draft electronic scientific papers written in physics, mathematics, astronomy, computer science, and statistics. Which you can access online. These archives are on the arXiv.org site.
“We know that data is the key to robotic control, and to get enough data to achieve this control, robots will have to collect data themselves. It is what we call self-supervised robot learning. The robot can collect consistent exploration data. Conclude if it has succeeded or failed to perform the task, and thus learn new skills.
The new approach is based on the enhanced learning framework associated with the same team’s goal-conditioned RL framework. Researchers used targeting technology in an underlying space to train the robot in skills such as pushing objects or opening doors. Directly through pixels, without the need for an external reward system or situation assessment.
“In our new work, we focus on generalization. How can we achieve self-monitoring robotic learning to learn a new skill and generalize visual diversity while practising this skill? We believe that generalizing in new situations will be key to achieving better robotic control.”
Instead of training the robot on several individual skills. Nir and his team designed the conditional goal-setting model to identify specific achievable goals commensurate with the robot’s condition. The algorithm developed by the researchers learns a particular pattern of representation. Separating things that a robot can control and those that it cannot control.
Self-monitoring robotic learning technology is based on the robot collecting data initially in the form of images. And actions through their interaction with the surrounding environment. And then training the robot compressed representation of this data. Converting ideas into limited-dimensional vectors that implicitly contain information such as object placements. So that, instead of teaching the robot directly what it should learn. This representation automatically absorbs concepts by compressing their objectives.
A Robot Can Achieve Different Goals
“Using acquired representation, a robot can achieve different goals and develop an enhanced learning policy.
Compressed representation is the basis for the practice phase and it uses to measure the similarity between two images to let the robot know if it has succeeded or failed. And is used to represent the goals so that the robot can achieve them so that when testing the robot can match the image of a target defined by a human being by implementing the policy it learns.”
The robot collects data from random interactions to be used to train representation, as data outside the enhanced learning policy
The researchers assessed the effectiveness of their approach using a series of experiments in which the robot processes objects it sees for the first time in an environment created using the MuJuCo simulation platform. And interestingly, their method of training has enabled the robot to acquire skills that can apply in new situations automatically.
More specifically, the robot was able to deal with various objects. Circulating the coping strategies it had previously acquired on new bodies that had not been matched in training.
First, we found that we could train a policy to push bodies in the real world, up to about 20. But the fact that acquired policy can drive more objects. This type of generalization is the main objective of deep learning methods. And we hope that it will be the beginning of more advanced images than generalization.
The team’s self-monitoring learning robotic learning model may enable robots to handle a wider range of tasks without having to train in a wide range of individual skills. At the same time, nir and his colleagues plan to test their approach in simulated environments while looking for ways to improve them.
“Inventors are now continuing to search on different axes, including solving tasks using more visual diversity. Solving a wide range of tasks simultaneously. And also, considering the possibility of using a task solution to speed up the resolution of the next job.