Deep Learning Boosts Robotic Picking Flexibility

Gripping and manipulating items of diverse shapes and sizes have long been one of the biggest challenges facing industrial robotics. The difficulty is perhaps best summed up by the Polanyi Paradox, which states that we "know more than we can tell." In essence, while it may be easy to teach machines to exhibit a high level of performance on tasks that require abstract reasoning such as running computations, it is substantially harder to grant them the sensory-motor skills of even a small child in all but the most standardized and predictable environments.

However, with the need for flexible picking to accommodate reduced changeover time for more varied product runs on the rise, industry is pursuing new solutions to the problem.

A new collaboration project between Festo and the Karlsruhe Institute of Technology (KIT) called FLAIRPOP (Federated Learning for Robot Picking), seeks to use training data from multiple stations, plants, or even companies to more effectively drive deep learning algorithms with the goal of helping picking robots to become more adaptable. The approach is similar to that used by cloud-driven machine learning algorithms, which leverage larger amounts of data than individual end-users have access to.

• [Data Management - Data Access]
  

  The process works by allowing autonomous robots at multiple different picking stations to grip and transfer items of various shapes and sizes. This data is then aggregated and shared, allowing other robots to more effectively manipulate objects they have not yet encountered.

• [Efficiency Improvement - Productivity]
  

  The FLAIROP research project is developing new ways for robots to learn from each other without sharing sensitive data and company secrets. This brings two major benefits: 1) protecting customers' data and 2) gain speed because the robots can take over many tasks more quickly. In this way, the collaborative robots can support production workers with repetitive, heavy, and tiring tasks.