Computer Vision Improving Worker Safety by Eliminating Heavy Lifting
For years, efforts to deploy robotic automation to take over dangerous and harmful tasks for workers have over-promised and under-delivered.
Industries seeking to protect and retain their employees implemented numerous robotic systems designed to tackle heavy and repetitive lifting, but each time they fell short or failed at these tasks, significant human intervention remained a requirement, putting workers right back into harm’s way.
What’s been holding them back? Substandard item detection, identification and prioritization; faulty information about item packaging, center of gravity, item obstruction; and more. These perception deficiencies result in frequently failed picks, but new advances in computer vision address those issues, increasing successful picks and worker safety.
Injuries mounting
While the Occupational Safety and Health Administration (OSHA) and the European Agency for Safety and Health at Work (EU-OSHA) don’t limit maximum lift weights for workers, the National Institute for Occupational Safety & Health (NIOSH) does have a recommended weight limit formula and suggests a maximum of 51 pounds/23kg. It’s well established that lifting heavy objects at work comes with a number of health risks.
In 2020 (the most recent year of available data), nearly 480,000 occupational injuries treated in U.S. hospital emergency departments were the result of overexertion and bodily reaction, according to the Center For Disease Control and Prevention’s (CDC) Work-Related Injury Statistics Query System
As injured workers feel the pain, so does the productivity of the organizations relying on them. The U.S. Bureau of Labor Statistics reports more than half (255,490) of those cases resulted in days away from work with injuries including sprains/strains/tears, fractures, bruises/contusions, carpal tunnel syndrome, tendonitis, and soreness/pain.
Extrapolate those rates globally and the number of workers injured on the job from heavy lifting and repetitive motion soars along with medical and insurance costs. This creates a snowball effect that limits the availability of workers due to injury, as well as challenges related to staffing and retaining employees for these jobs.
Improving worker safety
Workers' safety and well-being are among the primary issues automation can address, but not all automation is created equal. Many older installations struggle with 20-25% error rates. If we conservatively estimate a pick rate of 500 parcels per hour, that’s more than 100 parcels per hour that require human intervention. If those parcels exceed the suggested maximum lift weight, the workers handling those parcels are at extreme risk of injury. In dry goods and other industries where variably stacked ingredients and materials require intake, many jobs require heavy lifting for every pick.
Fortunately, advanced computer vision can drive these error rates down and is already being deployed by leading brands across various industries. DHL, for example, worked with robotics integrator AWL to develop a robot for picking and placing parcels from a randomly mixed pallet onto the conveyor belt of a sorting installation. The solution combines computer vision with gripping technology capable of picking various packages up to 31.5 kg at a rate of 800 parcels per hour, automating work previously handled by workers.
AVT Europe NV has deployed depalletizing robots for different end users that currently pick and place heavy bags exceeding 25kg from variable pallet stacks at a rate of more than 400 bags per hour and are capable of reaching 1000 bags per hour.
While robots and grippers have long been capable of handling heavy objects, their muted success rates have been driven by lackluster vision software providing limited information. As robots become able to collect more and better data faster, their capabilities and reliability grow, and new applications become possible and more profitable.
Advanced computer vision makes the difference
A new breed of advanced computer vision has made significant improvements in successful pick rates of even the most difficult objects. What separates these new vision solutions from their predecessors:
●Training methods: multiple methods exist for teaching computer vision to see, strategize and provide direction, but supervised machine learning has separated from the pack by controlling how, what and when the system learns. This removes the need to unlearn faulty information that other teaching techniques allow into the systems, reduces ramp-up time and provides ongoing flexibility.
● Object identification capabilities: a lot can be learned from a photograph in just one tenth of a second, including segmentation, shape detection, material detection (cardboard, polybag, etc.), size, transparency, edges in poor lighting, slip sheets, damages and center of gravity.
●Prioritization: combining 2D and 3D data clarifies which objects are on top or overlapping and the flattest, largest, and sturdiest surfaces for picking.
●Brand agnostic robot instruction: communicating to the specs of each robot’s coded capabilities provides precise pick positions, angle of approach, and the right force and grasp pose for gripping and movement.
With high-quality algorithms, advanced computer vision will increasingly adapt to unique situations, improving the performance of a wide range of robotic cells in even the most complex scenarios. It detects deformations, damage, wrinkles and labels and will adjust accordingly to optimize the gripper’s capabilities, tuning the picking and placing process on the go and remaining accurate and productive in always-changing conditions.
The average error rate of new robotic installations with computer vision lands around 6%, a significant improvement from the 20-25% of older installations, but the installations utilizing the most advanced computer vision reduce those error rates down to less than 1%. In the 500 parcels per hour scenario, that reduces the more than 100 parcels per hour that require human intervention to less than five parcels per hour. Over an eight-hour shift, that’s 40 parcels requiring human intervention, 60 fewer than older installations require in a single hour.
More automation on the horizon
Just as advanced computer vision unlocked the picking capabilities of heavy objects, it’s also unlocking the potential for automation in other scenarios that are hazardous to workers. Trailer and container unloading, for example, is especially dangerous, with extreme heat and cold temperatures, poor visibility, uncomfortable and unstable surfaces, and the need to deal with objects that have shifted during transport.
To date, automation of trailer unloading has only been a pipedream. The same poor visibility and object disarray that has contributed to worker safety problems has also hampered the automation of this task. The unsophisticated solutions of the past struggled to pick and place objects that had no pattern, but advanced computer vision solves these problems with the same three-step approach used for picking heavy objects: item identification, prioritization and precise picking instructions.
Fewer manual interventions for better health and safety
As more industries, companies and governments around the world consider the importance of worker safety, more employers will turn to automation to tackle the jobs that cause the greatest harm to workers. While a 75% success rate in automating these tasks is a good deal better than no automation at all, it still leaves workers in a position to sustain an injury and negatively impact their long-term health. As more automated solutions utilize advanced computer vision, they greatly reduce exposure to these harmful tasks.
In the case of heavy lifting, they can nearly alleviate the need for human intervention altogether by reducing error rates to 1% or less. Requiring manual intervention 40 times over an eight-hour period is a much safer proposition than 100 interventions per hour, and employers are seeing the difference it can make.
Ken Fleming is the CEO of Fizyr which offers advanced computer vision for robots, providing the smartest, fastest, and most effective brain available to maximize robotic capabilities.
