Why You Should Care About Deep Reinforcement Learning
Category: Manufacturing Technology • Jul 1, 2021
The concept of deep reinforcement learning (DRL) may seem like science fiction, but in fact it’s a form of Artificial Intelligence (AI) that is going to advance manufacturing in the coming years.
OSARO, a San Francisco-based company focused on DRL for industrial automation, is building technology for the factories, warehouses, and logistics systems of the 21st century. DRL enables automated systems to manage more complex decision-making tasks than previously possible. This opens the door for robots to be used in more applications that were unthinkable even just a couple years ago. As the technology advances, so too will the viable operations for wider adoption of robotics in manufacturing.
Today’s traditional robot systems have highly structured environments that do the same thing over and over again, but if there is even a slight change to that environment, the system doesn’t work. DRL will change how manufacturers use robots in a few common applications.
Pick & Place
Many robots today are used for picking and placing parts and components. These systems involve a lot of hardware-based solutions such as shaker tables and conveyor belts to physically restrain the input parts and create a repeatable orientation that the robot can recognize.
For enhanced optics, some systems are installed with high-end sensors that can be very expensive but still have limitations on how much detail they provide. If these systems had DRL capabilities, they could replace shaker tables and conveyors with high-resolution vision systems and simpler, less expensive sensors.
In the case of robotic welding, people are required to manually place parts into fixtures so the robot can weld them in the exact same spot every time. If the parts are not in that exact position, or if the parts have changed, the robot will not recognize them or will deposit a weld bead improperly centered in the joint. A robot with DRL will be able to move parts of any position and size, and in some cases, it may remove the need for complex fixturing altogether.
For machining applications, robots that have established DRL will be able to load parts into fixturing and remove finished products regardless of the orientation, and with the ability to run many kinds of parts without changing a program.
Currently, DRL is best suited for tasks that don’t need extensive input data to train, such as picking up objects. Applications with higher fault tolerances are most logical candidates for DRL because if the robot drops something, it should be able to pick it up again and continue operating. Parts should not be breakable or expensive in case they are damaged or destroyed. This is most applicable for metal parts and food items, but may not be ideal for more delicate medical or electronic components.
Further advances in vision systems will make it possible for robots to detect items that have historically been a challenge, such as those that are shiny, reflective, and clear.
Eventually, DRL will be used for learning across robots. Once a robot has learned and established a proficient policy, it will be applied to other non-identical, uncalibrated robots on multiple production lines throughout the operation.
Deep reinforcement learning will soon have real-world benefits for manufacturing companies of all sizes. Research on the technology continues to grow, so it’s only a matter of time before we see this level of AI in robots across the industry.