Computer Vision
Computer Vision in Robotics: Allowing Robots to See
Created on
October 22, 2022
Written by
Chinar Movsisyan

Throughout history, the world has seen several major technological and economic revolutions. More than 10’000 years ago, the Agricultural Revolution transformed society from a hunter-gatherer one to an agriculture and settlement based society. The Scientific Revolution of the 1600s saw the emergence of modern science and changed how we study and observe the world around us. Perhaps the biggest economic and technological revolution, the Industrial Revolution that began in the 1700s, completely transformed our economies and radically increased the standard of living of people.

Each of these transitionary periods brought with them radical disruption and new opportunities to improve society. It could be argued that the world is currently living through another one of these transitory periods. Advancements in robotics and artificial intelligence have made it possible to streamline manufacturing workflows and automate repetitive tasks that humans have traditionally done in factories. These technologies are already boosting economic productivity and changing the way companies think about their supply chain and manufacturing processes.

Where Robots Are Used Today


Robotics has long been a point of interest for the manufacturing industry. Manufacturing often involves repeatable processes which robotic systems can conduct efficiently and effectively. These tasks involve material handling and pick and place tasks (sorting, palletizing, stacking, etc.), which are handled using robotic arms. Incorporating robotics into a company’s manufacturing process increases productivity and saves on labor costs. It can be expected that this trend will only continue as robotics and AI technologies continue to advance.

Distribution & Supply Chain

As e-commerce continues to take up a greater and greater share of consumer spending, distribution and supply chain challenges need to be optimized in order to meet the demands of the market. Robots are increasingly being used in warehouses to boost overall productivity and create more efficient workflows. Several companies today build what are known as autonomous mobile robots. These robots are able to learn their surroundings and plot routes to move around the warehouse. They can be used for moving pallets and products from their picking locations to the shipping or staging area. According to Honeywell, one of the world’s largest developers of software for process manufacturing, picking times can be reduced by “nearly 50%” by incorporating mobile robots into the workflow.

Increasingly, there are also several companies making advancements towards autonomous systems for trucks. Self-driving trucks will eventually significantly disrupt the supply chain process, as they will be able to work without the constraints that human-operated trucks have. Companies such as Tesla, Waymo and Torc are all actively developing autonomous trucks.

Robotics for Recycling

Robots are also being utilized in the recycling industry to conduct waste sorting. The United States alone produces approximately 300 million tons of trash a year. The portion of this that is recycled needs to be sorted, as not all pieces that are sent for recycling are actually recyclable. Traditionally, this job has been done by humans, who often have to work on a conveyor belt sorting through trash for items that can be recycled. As you can imagine, this is a challenging job, and often recycling centers can’t recruit enough people necessary to do their job. Robots, with the help of computer vision, can fill the gaps to do this job. Computer vision models can be trained to detect and sort through the recyclable objects on a conveyor belt. The system will make a determination as to whether or not something is plastic, glass, paper—or something that is not recyclable—and sort them accordingly.

Computer Vision Challenges for Robotics

In order for robots to be useful in dynamic environments, they must have some understanding of the tasks they’re trying to accomplish. This is where machine learning and computer vision come into play. Tasks such as object detection and object tracking are done using computer vision algorithms. In the distribution and supply chain examples we discussed earlier, the mobile robots that operate on the warehouse floor use computer vision to detect and track objects that may be blocking their way, and adjust accordingly. As the robots are moving boxes onto a pallet, they must be able to detect whether the pallet is within reach, if there is room on it, and where to stack the object onto the pallet. All of this is done using object detection and tracking algorithms.

As with any other machine learning model, the key to training effective computer vision models that can operate in robust environments such as warehouses is having good data. There are several challenges that arise when acquiring data and training models for robotics.

In the example of robots that help sort through heaps of materials sent for recycling, the computer vision model must be able to identify cans as recyclable aluminum materials. While it may be easy to build a model that can identify soda cans on a shelf in a grocery store, the cans that end up on a conveyor belt in a recycling center can come in a variety of shapes. Some of them will be crushed to save space, others may be cut open or be significantly dented. Being able to correctly identify the object, regardless of its current shape, is an important aspect for the model to be effective in production. There is also the issue of occlusion, meaning that parts of the object that the model is trying to detect is not visible, usually because it is covered by something else. The position of the object can also differ from scenario to scenario, as can the position of the camera the robot is operating with. This can lead to some complexities for the computer vision algorithm to correctly detect the object.


Machine learning models’ performance decays over time. Fine tuning and improving machine learning models—including computer vision ones—start with identifying where your model is not performing well. Model performance monitoring tools help to detect the outliers on which your model is failing, allowing you to address the deficiencies and significantly reduce the feedback loop. This is especially true when working with systems that have to operate in dynamic environments such as warehouses and factories.