Jenga whips: AI robot knocks Jenga blocks out of tower 100 percent of the time

Scientists at the University of California Berkeley (UC Berkeley) have trained a robot so that it can independently knock individual Jenga blocks out of a tower using a whip without the tower becoming unstable and collapsing. The researchers combined reinforcement learning methods with corrections by a human when training the robot's artificial intelligence.

Jenga whipping is a popular sport that is also played competitively. Individual Jenga blocks are knocked out of a stacked tower with a whip. The tower must not collapse after being hit. To do this, the athletes must have a good eye, excellent reflexes and precise hand coordination. The success rate varies significantly depending on experience and ability.

The researchers at UC Berkeley wanted to have this done by a robot. The goal: the robot should be able to complete the task with a success rate of 100 percent. In the technical report "Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning", the scientists describe the AI method they used to train the robot.

According to the report, the researchers first used reinforcement learning to train the robot AI. A camera documented the successful and unsuccessful attempts. The successful attempts were recorded in an AI database, while the unsuccessful ones were discarded. The scientists deliberately avoided training in a virtual simulation. Instead, they concentrated exclusively on training in the real world. The reason: modeling the whip strokes proved to be very complex. It is therefore less efficient to train the robot in a simulation alone.

Human correction factor

However, training using reinforcement learning was not enough for the scientists. In order to achieve success more quickly, they incorporated a human correction factor. To do this, they used an input device that can precisely control the movement of the robot manually. This correction information was also included in the database if it proved successful. Initially, a human had to repeatedly intervene and correct the robot. From around 30 percent of the trials, the robot was gradually given less attention.

At the end of the training, the scientists tested the robot. It managed to knock individual blocks out of a Jenga tower with a whip 100 percent of the time without the tower collapsing. Jianlan Luo, one of the researchers involved in the project, believes that the robot has an advantage over a human Jenga whip – even if it is an experienced player. This is because, compared to a human, the robot has no muscles that could tire over time. A robot therefore performs every stroke with the same precision, as a video shows.

The scientists tested their training method on other tasks for a robot from –, including assembling a shelf, loading a circuit board and turning a fried egg in a pan. The researchers compared their combined training method of reinforcement learning and human intervention with a common method of behavioral cloning. In both methods, the robot was trained with the same amount of demonstration data. As a result, however, the combined training method made the robot faster and more precise.

(olb)