With Autonomous Systems, Sberbank develops a unique AI robot control system

With Autonomous Systems, Sberbank develops a unique AI robot control system

Sberbank research labs developed a unique AI robot control system to manipulate heavy coin bags with an accuracy of 95% in real life conditions. Neal Analytics Autonomous Systems experts helped design and train the solution leveraging Microsoft Project Bonsai platform.

Challenges

Sberbank is one of the largest Russian and Eastern European banks. It needs to handle large quantities of coin bags. Carts containing up to forty coin bags need to be unloaded by human operators for counting and repackaging. Each bag weighs about 2kg (4.4 lbs.) and, for the ones at the bottom of the cart, can be hard to reach by hand. This caused unnecessary stress on operators’ back and arms as the process needed to be repeated multiple times each day.

The physical demands of this task also quickly tire human operators, making it difficult to scale operations during peak times.

Sberbank research lab decided to look for a solution that would leverage a robotic arm to perform the task of picking up each bag from a cart and to drop them on a table at operator level. However, existing simple robotic control systems would not be able to solve this problem as the coin bags could have unpredictable shapes and locations and traditional robotic control systems would not be able to dynamically adjust to these changing operating conditions.

Instead, Sberbank decided to work with the Microsoft AI team to find a workable and innovative solution.

Solutions

process overview for Sberbank AI robot control system

The customer decided to leverage Microsoft Project Bonsai, a platform used to design, train, and operate Autonomous Systems, to solve this challenge.

Autonomous System AI agents, aka “brains”, self-train using the concept of Deep Reinforcement Learning (DRL). This trial-and-error approach requires the availability of an accurate simulator as the AI cannot be trained on a real-life system. Therefore, the customer decided to develop a MuJoCo physics-based simulation to train the agent.

After analysis and considering the inputs of both Neal Analytics and Microsoft experts, as well as Sberbank’s operators and researchers, the team decided to build a solution that would use a single AI agent (i.e., Project Bonsai “brain”). Vision sensors would then be used to provide the necessary brain inputs to first train then operate the agent.

Using the simulator, each agent training – i.e., from start to working brain- required between 300,000 and 5 million DRL training cycles and would take up to 20 hours.

Once a training was complete, the customer would test the AI agent with different cart setups, ranging from 5 to 40 bags, to evaluate whether the robot arm was able to find, pick up, and finally drop each bag successfully on the table.

Results

After multiple tests using different training and design strategies, the AI-controlled robot was able to pick up bags 95% of the time on the first attempt. For the remaining 5%, a second try would take place if the robot missed the first time.

In addition, the robot was able to perform this task at a speed that was close to human speed, making it a viable solution for the next step in this project: field deployment.

As the robot’s speed was limited due to the requirements of the inverse kinetic calculations needed for the brain to instruct the controller, it also meant that further speed increase would be possible through optimization of these kinetic calculations.

Learn more: