With the lowest unemployment in nearly 50 years, the need for automated bin picking is urgent. In the U.S., where 38 percent of the manufacturing labor force moves parts between bins and manufacturing machines, 500,000 jobs remain unfilled. The automation industry is trying to meet that need: every automation trade show adds a few new companies that claim to have finally solved the puzzle of bin picking.
Automatica 2018 in Munich, for example, had no fewer than 14 bin-picking demonstrations. But even at large manufacturers, few bin-picking stations can be found. And at small and mid- sized enterprises (SMEs), the number is close to zero.
Why is the adoption rate of automated bin picking systems so low, when the need is so great and so many vendors claim to offer a solution?
Automated bin picking a complex problem
The simple answer is that for the most part, automated bin picking is only a partially solved problem. Picking randomly positioned parts from a bin and placing them precisely into a machine is a simple task for a human but a daunting task for robots. Robots must be able to grasp parts in an infinite number of orientations and reach deep into the corners of the bin all while avoiding collisions with the bin, other parts, or the work cell itself.
Currently, bin picking is able to be fully automated only with a huge system integration project that requires multiple advanced technologies to work together. These include:
- A 3D model of the part, the bin, the robot end effector, the placement target, and any environmental obstacles
- A model of one or more ways to pick up the part with the end effector and deposit it at the placement target
- A 3D sensor to map the bin
- Image analysis software to locate each part and potential obstacles in the bin
- Path planning software to find a collision-free route from the part’s pick-up point to the placement target
- Robot control software to maneuver the robot, end effector, and part along the route
There are commercial bin picking systems that include some of these components and address a subset of bin picking challenges. Usually these systems combine a 3D sensor with image analysis software that runs on a separate computer. A robotics expert is expected to integrate the sensor, computer, software, and robot controller, and then write a program to retrieve the location of each part and figure out how to get it to the placement target. (See Figure 1.)
Creating a general path-planning algorithm starting with infinite variation in part orientation is a near-impossible task. At best, weeks to months of experimentation and tuning yields a specialized algorithm with unreliable performance.
Because bin picking is so complex, specialists have attempted to apply deep-learning techniques. So far, the results have been disappointing, at least for industrial use. Artificial intelligence (AI) is effective for applications such as image classification and voice recognition, where lower accuracy is acceptable. But AI simply doesn’t have the reliability and accuracy that manufacturers need to replace human operators for bin picking.
With the large integration and programming effort required, it’s no surprise that most real-world bin picking deployments are found at large, sophisticated manufacturers such as automotive OEMs. But most bins and manufacturing machines are found at SMEs, with 69% of the industrial labor force worldwide. SMEs have the greatest labor shortages but the least capital and expertise required to create a bin-picking system.
Path planning key to ease of use and reliability
A truly universal bin picking solution needs to be usable by non-experts, configurable in a few hours, provide sophisticated path planning that works with little or no tuning, and be cost-effective for SMEs. This solution will truly democratize bin picking. (See Figure 2.)
The underlying software in a completely pre-integrated system manages all the complexity. All the system components (sensor, imaging and path planning software and co-processor) are pre-configured and plug in to the robot controller with no setup required, reducing installation time to hours.
All programming is integrated with the robot’s teach pendant, enabling bin picking actions to be freely mixed with standard robot commands using the same programming interface. Training time is minimal.
Setup and programming are guided by a series of wizards that mimic the training a human operator would have to go through, e.g., how to pick up a part, where to put it down, obstacles to avoid, etc. (See Figure 3.)
Planning a unique, collision-free path for each part in the bin to the placement target is by far the trickiest challenge for bin picking, and is often left to the ingenuity of an automation expert. Path planning is the main determinant of system reliability, and if not done well will result in collisions, parts left in the bin or dropped, and missed targets.
Finally, a bin-picking solution for SMEs
So, can we declare that the bin picking problem has been solved once and for all, and manual machine tending is a thing of the past? Not quite, since there will always be applications that are just too complex for automation and will require human operators for the foreseeable future.
However, advances in cost-effective cobots, 3D imaging, and intelligent motion control software have aligned to enable a new generation of bin picking solutions that can handle many more of today’s machine-tending applications, and enable manufacturers to shift their limited resources to higher-value tasks. More importantly, bin picking has become accessible and cost effective for SMEs.
About the Author
Eric Truebenbach is director of corporate development at Teradyne. Eric is currently acting as head of business development for robot applications at Energid, which was recently acquired by Teradyne. He specializes in identifying market opportunities created by combining new technology, applications, and best-in-class companies. Prior to his current position, Eric has served as leader of an internal startup, engineering manager, project manager, architect, and electronics engineer. He holds many patents in automation and instrumentation design, and has authored papers and been a speaker at technical conferences.
Filed Under: AI • machine learning, The Robot Report