Machine learning is a very important factor that helps robots to collect items consisting of very small parts more quickly and efficiently. For example, smartphones and other electronics. The robot that plays the popular game Jenga has advantages over previous systems. He has the ability to self-development – the robot masters several ways to complete the task and selects the best option. Mastering occurs not only through visual perception, but also through tactile sensations, physical interactions.
Research takes place in the basement of the building 3 MIT. Here the robot ponders its subsequent moves, the goal of which is to choose the best block to extract without destroying the Jenga tower.
The MIT engineers who created the player robot equipped it with a grip, a cuff to measure strength, and an external camera used by the robot to visualize and feel the individual blocks and the tower as a whole.
With a careful push of the new part, the computer processes the received tactile and visual information. Comparing and measuring the data obtained, and the best option is chosen from those made by the robot earlier. Also, the calculations take into account the results of each step – including the computer analyzing whether the block was successfully removed after the move made, and also takes into account the specific configuration and strength of the block movement. In these few seconds, the robot understands whether it is worth continuing this move or it is better to move to another block so that the whole structure does not fall.
The first results of the Jenga robot published the famous journal Science Robotics. Associate Professor of Mechanical Engineering at the Massachusetts Institute of Technology named after Walter Henry Gale – Alberto Rodriguez, he was noted that the robot shows amazing results precisely because of the presence of visual and tactile interaction.
“Jenga has major differences from the more common machine learning games – chess or go are inferior to this game, because here you need to master physical skills: pushing, probing, placing, stretching and aligning figures.”
According to a specialist, it is very difficult to simulate, so how manipulations require interactive perception and manipulation.
“The robot is constantly learning new things in real time. It interacts with the real Jenga tower. The key task is to draw lessons from relatively little the number of experiments, using common sense in relation to objects and physics “.
Alberto is sure that the tactile learning system developed by researchers can be used in applications not only in the proposed game, but also outside it. It is especially useful when careful physical interaction is required. For example, for the separation of recycled materials and garbage and the collection of consumer goods.
Push and pull
In the version of the game, which got the robot fifty-four blocks, which are folded into 18 layers of three blocks each. In this case, the blocks in each layer are oriented perpendicular to the blocks below. The robot has one task – to carefully remove one of the blocks, move it to the very top of the tower so that when building a new level the whole structure does not fall.
For programming the robot, traditional machine learning schemes were not suitable. For this, it would be necessary to take into account everything that can happen when the tower, blocks and the robot interact. This computational task is quite expensive and complicated, since it requires the construction of tens of thousands of variants of attempts to extract blocks.
The creators were looking for another method that would be more effective and resembled human knowledge – how would a person approach this task – playing Jenga.
For this developers:
- adapted the standard robotic manipulator ABB IRB 120;
- set the tower Jenga within reach of the robot, and started learning.
During the training, the robot first selected a random block and a place on the block to be pressed. Then he made a little effort, trying to push the block out of the tower.
For each attempt at pulling a block, the computer recorded the corresponding visual and force measurements and noted whether each attempt was successful.
Instead of performing tens of thousands of such attempts (which would have required the reconstruction of the tower almost as many times), the robot trained only 300 with attempts at similar measurements and results formed on clusters representing a certain block behavior. For example, one cluster stores data on attempts for the block that the robot was difficult to move, compared to a block that was easier to move, or that overturned the tower when moving. For each data cluster, the robot developed a simple model for predicting behavior based on its current visual and tactile measurements.
One scientist says that this method using clusters significantly increases the efficiency with which the robot can learn to play Jenga. He is inspired by the natural way people group similar behaviors.
Researchers tested their method by comparing with other modern machine learning algorithms. For this was used simulator MuJoCo. The lessons learned in the simulator informed the scientists how the robot is being trained in real time.
“We gave these algorithms the same data that our system received. We saw how both systems learned to play the same game on the same level. ”
“Compared to the unique approach we developed, standard algorithms had to be investigated several orders of magnitude more towers in order to study the game.”
Also, scientists organized an amazing experiment with the participation of real people – the volunteers also tried to get blocks from the tower without dropping it. And in comparison, the results of the robot and man practically did not differ in performance.
But one difference is still not fully developed by developers. In addition to physical interactions in Jenga, a strategy is needed that finds the right block, which is not only well pulled out, but also prevents the enemy from making the next move without destroying the tower.
The team of scientists does not set the task to make the champion of the game out of their robot – they have other more important tasks. The skills acquired by the robot will be applied in other more important areas.
“There are many tasks that a person performs manually when the feeling that this is done correctly comes in the language of strength and tactile signals,” says the project’s head, Rodriguez. “For such complex tasks, our developed approach will help us.”
The study with the robot and Jenga supported the National Science Foundation within the National Robotic Initiative.
Editor of IMD News