Virtual training and commissioning of industrial bin picking systems using synthetic sensor data and simulation

ABSTRACT

Defined handling of unsorted parts, known as bin picking, is a challenge in robotic automation. Available solution concepts for this problem are usually either costly or require considerable setup and tuning efforts. In this contribution, a setup for virtual commissioning of such automation systems is introduced. Using a physics-based simulation environment, a virtual stereo-camera simulation and robot controller integration, a full simulation of the bin picking cycle is possible. The setup is also used to generate realistic synthetic training data for learning-based computer vision routines. The functionality of the system is demonstrated for generating training data capable of enabling a real-life deployment of the pipeline. A simulation of both model-based and learning-based bin picking systems is also conducted. This simulation also involves the path planning and execution as well as the grasp itself, allowing for a full simulation of the bin picking cycle.

KEYWORDS:

• Bin picking
• synthetic training data
• convolutional neural networks
• robotics simulation
• virtual commissioning

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by the Bayerische Forschungsstiftung [FORobotics (AZ-1225-16)].