Recently in the world of intralogistics, AI-powered robots (or AI Robotics) have been able to successfully automate picking applications at scale. From picking eaches and individual SKUs to parcels and cases, AI Robotics helps increase efficiency and throughput in the face of increasing order volumes and labor shortages.
Over the hundreds, if not thousands, of conversations we have had about AI…and specifically AI Robotics, there are a few questions that seem to come up repeatedly. While some people have a good understanding of AI, others have a more rudimentary understanding of what it is and why it is ideally suited to solve specific problems in many industries and sectors.
So it’s not uncommon for someone to ask: Can you explain AI to me like I’m an 8-year-old?
Here are some of the most common questions we hear about AI and AI Robotics as it relates to picking applications in warehouses and fulfillment centers.
1. Why do you need to use AI for automating warehouse robots?
If you walked into most automotive manufacturing facilities 20 years ago…or even 40 years ago… you would see robots in action assembling cars. Robots have been performing traditional automation for a while now—automation that works in a pre-programmed manner in a highly structured environment where a robot picks up known objects and places them in the exact same location repeatedly.
In a typical modern fulfillment center, however, a picking robot faces a much more unstructured and dynamic environment. Part of this is due to the sheer variety of items and scenes that a robot must interact with:
- Item variety: A fulfillment center might have tens or hundreds of thousands of SKUs, each with differing shapes, sizes, packaging, materials, etc. A single facility might be handling small cosmetics items, apparel wrapped in transparent polybags, and everything in between. And the SKU mix might change multiple times a year with new SKUs being introduced.
- Scene variety: These SKUs might be placed in a bin or tote in any configuration (tightly packed, chaotically placed, overlapping items, etc.). Lighting conditions, tote size, item orientation, etc. could be varied as well.
These can result in a virtually infinite number of variations that cannot be ‘hard-coded’ for traditional automation where a robot is explicitly pre-programmed to handle each specific item for every specific scene variation.
This is where AI comes in…specifically modern neural network-based AI using deep learning. Inspired by how humans learn and reason, modern AI is able to make inferences and decisions based on past experiences. It is uniquely suited to work in exactly such an unstructured environment where we cannot explicitly train on every single item and scene variation.
Check out the video below by our Chief Scientist and Co-founder, Pieter Abbeel, explaining modern AI in a very easy-to-understand manner.
1. Why do you need to use AI for automating warehouse robots?
If you walked into most automotive manufacturing facilities 20 years ago…or even 40 years ago… you would see robots in action assembling cars. Robots have been performing traditional automation for a while now—automation that works in a pre-programmed manner in a highly structured environment where a robot picks up known objects and places them in the exact same location repeatedly.
In a typical modern fulfillment center, however, a picking robot faces a much more unstructured and dynamic environment. Part of this is due to the sheer variety of items and scenes that a robot must interact with:
- Item variety: A fulfillment center might have tens or hundreds of thousands of SKUs, each with differing shapes, sizes, packaging, materials, etc. A single facility might be handling small cosmetics items, apparel wrapped in transparent polybags, and everything in between. And the SKU mix might change multiple times a year with new SKUs being introduced.
- Scene variety: These SKUs might be placed in a bin or tote in any configuration (tightly packed, chaotically placed, overlapping items, etc.). Lighting conditions, tote size, item orientation, etc. could be varied as well.
These can result in a virtually infinite number of variations that cannot be ‘hard-coded’ for traditional automation where a robot is explicitly pre-programmed to handle each specific item for every specific scene variation.
This is where AI comes in…specifically modern neural network-based AI using deep learning. Inspired by how humans learn and reason, modern AI is able to make inferences and decisions based on past experiences. It is uniquely suited to work in exactly such an unstructured environment where we cannot explicitly train on every single item and scene variation.
Check out the video below by our Chief Scientist and Co-founder, Pieter Abbeel, explaining modern AI in a very easy-to-understand manner.
2. Is AI as good as humans?
As humans, we are amazing at learning and adapting to our environment. The bar set for robots is pretty high in that sense. When it comes to repetitive lower-order item-picking tasks in a warehouse setting, such as sortation, induction, etc. AI-powered robots are able to perform at a human level in terms of speed and accuracy. This helps augment the efficiency of the human workforce, freeing up workers to focus on higher-order tasks that require strategic and critical thinking.
3. How is AI trained and how long does it take?
AI comes pre-trained based on millions of picks across hundreds of thousands of SKUs from warehouses around the world. Modern AI based on deep learning does not need to be explicitly trained on every single item type or scene — instead, it uses its vast existing knowledge to infer how to handle items it hasn’t seen before. This means that AI-powered robots are performant on Day One, and keep improving over time as it experiences new items and scenes.
To see AI Robotics in action, highlighting out-of-the-box intelligence that allows the robot to pick never-seen-before items, check out the video below of the Covariant Brain performing in and winning ABB’s picking challenge.
4. When new SKUs are added, how does AI know how to handle them?
Usually, there isn’t a need to do anything to specifically teach a robot how to pick up new SKUs — when an AI-powered robot experiences a new SKU or item it has never seen before, it will try to use its previous learnings to try to pick it up. It will also incorporate whether it failed or succeeded into its learnings to continue to adapt and improve.
5. What happens when the AI makes mistakes?
At a high level, like humans, AI can make mistakes. And like humans, it learns from its mistakes. To use a specific example, say a robot tries to pick up lipstick unsuccessfully. It might try to use the same grasp points to pick it up a couple of times…and maybe even try to perturb the scene and try another grasp point. All of this gets incorporated into the AI model so that when it experiences a similar item and scene in the future, it can leverage this learning.
5. What happens when the AI makes mistakes?
At a high level, like humans, AI can make mistakes. And like humans, it learns from its mistakes. To use a specific example, say a robot tries to pick up lipstick unsuccessfully. It might try to use the same grasp points to pick it up a couple of times…and maybe even try to perturb the scene and try another grasp point. All of this gets incorporated into the AI model so that when it experiences a similar item and scene in the future, it can leverage this learning.
6. How do I tell which AI Robotics solution is better than another?
It would be hard to differentiate AI solutions for warehouse automation purely based on claims of performance and capabilities from marketing materials. A rigorous performance-based assessment is really the best way to evaluate whether an AI solution can handle the real-world variations a robot might experience in your warehouse, with your specific items and SKUs. Such an assessment should test out-of-the-box performance, learning speed, and learning potential. Check out this blog post that talks more about conducting such performance-based assessments.
We could go miles deep on each of these questions about AI and AI Robotics. But at a high level, it is important to understand why modern AI is a key enabler for automating warehouse tasks that were previously mostly manual.
At Covariant, our researchers stay abreast of — and lead — the latest advancements in AI. Our engineers incorporate them into our robotic systems to continue pushing the boundaries of warehouse automation.