RBR50 Executive Q&A: Victor Anjos, Kindred.AI

Image: Kindred.AI

July 17, 2019      

Rising consumer demand for products through e-commerce has created incredible pressure on companies to improve their fulfillment operations, while at the same time they deal with a growing labor shortage for warehouse workers. Many have turned to robot companies to help them with this problem, including ones that are using robots to help pick and sort items through the use of artificial intelligence.

RBR50 2019 honoree Kindred.AI has developed systems that enable robots to interact with the physical world for the past five years. Its SORT system for piece-picking in the order fulfillment space includes the company’s AutoGrasp Robotics Intelligence Platform, which uses advanced AI algorithms in vision, grasping, and manipulation.

Robotics Business Review recently spoke with Victor Anjos, the company’s vice president of engineering, about recent developments driving the AI piece-picking space, how the company plans to make its robots even smarter, and how it reacts to manufacturers creating more and more products for robots to identify and manipulate.

Q: In the past few years, we’ve really seen piece-picking applications and companies emerge pretty dramatically. Is there a particular part of the technology that has enabled this growth – whether it’s computer vision, AI/machine learning, grasping/tools – or was it a “perfect storm” of factors all combining to create these achievements?

Kindred Victor AnjosAnjos: It’s due to a multitude of reasons. One is the downward pressure in the market coming from the big players in automation. To meet the consumer’s expectation for fast, free shipping, fulfillment centers to have to accelerate the process, without sacrificing completeness or accuracy. But hourly workforces are hard to find, staff and retain. The work is tedious and can lead to injury, so robots make more sense in those situations.

On the technology side, the advent of better computer vision systems, AI models, and compute capacity (through either GPUs or cloud infrastructure), combined with recent advances in machine learning have been paramount in driving the success of piece-picking systems, which is the area where Kindred shines the best.

Q: As you’ve developed the system for grabbing and sorting items in an e-commerce situation, what types of products have proven most difficult to master for the system? Is continual work needed to be able to grab those items, or are we at a point where a company will just use workers to grab those really hard items?

Anjos: Our customers expect our robots to be able to pick a majority of their items, and we are constantly refining our ability to do so. Our SORT robots come in several different configurations that are optimized for a variety of items like polybags, shoeboxes, small general merchandise items, and so on.

It’s important to understand that the challenge is broader than just picking up items with a variety of form factors. It’s about identifying items in a dynamic, fast-paced fulfillment environment. Then the item has to be grasped using human-like dexterity and placed in a specific area. All of those steps must be done very quickly, consistently, with as little wasted motion as possible. In some cases, we rely on a remote pilot to coach the robot on how to handle oddly-shaped or irregularly-sized items. In that case, the exception becomes a learning opportunity for SORT, our robot.

Q: On a similar note, do you think you’ll see companies from the manufacturing side start to develop packaging solutions to make their products more easily “grabbable” by robots? Wouldn’t that make your job easier?

Anjos: That would obviously make our job easier! But we don’t want to disrupt processes that may be working for a customer. We focus on building solutions that can meet our customer’s most challenging requirements.

Q: Another trend we’re noticing is that several robotics companies are teaming up with others to provide multiple-task applications. For example, a materials handling robot could deliver items to a central bin or sorting station, at which point the picking and sorting robot would then take over, and eventually a packing robot could then put the item into a box for delivery. Do you believe this would be the next stage for e-commerce fulfillment efficiency improvements?

Anjos: The amount of automation in fulfillment centers is growing exponentially. In the near term, there’s a need for associates to perform some of the tasks that robots can’t do now. But this need will diminish over time as we envision, design and create new robots optimized to perform specific tasks in order fulfillment.

Q: What developments are happening with making the system even smarter – have you found that the AI/ML algorithms, or your reinforcement learning approaches, have made the robots become better at their jobs? What’s next once the robot has “mastered” its tasks?

Anjos: Most everything we do is based on machine learning, and some of our best advances are in the area of reinforcement learning. With each iteration, we improve the performance and speed of our robots. When we believe we have an improved AI model we have a very scientific approach to confirm our hypothesis. We simulate, test and canary deploy the new software to a small subset of robots and gauge its performance compared to previous versions and the rest of the fleet.

Q: We’re also seeing a lot of robotics companies offer robots-as-a-service to help companies deploy these systems in their operations. Does this method put any additional pressure on Kindred to make sure that robots are maintained and serviced efficiently, since the hardware is still owned by Kindred?

Anjos: We see our customers as partners in this journey, so we want to ensure that we are aligned to their goals. In practice, this means that regardless of how our pricing model works, Kindred is committed to ensuring that SORT and upcoming robotic solutions are supported with upgrade paths, rapid issue resolution, and new features.

Q: How does your team keep up with the breadth of SKUs that manufacturers create; it doesn’t feel like consumer demand will decrease any time soon, so companies will likely see more and more product types come through the pipeline. How do you ensure that the system won’t get stumped?

Anjos: There are several key pieces to what we do; recognize that items are in the bin, choose a candidate to pick, identify it, and place it with the other items in the order. We have no issue handling new SKUs, it’s new form factors and packaging materials that must be solved for.

Q: What does it mean for Kindred to be a part of the RBR50?

Anjos: We consider it an honor to be selected. We are especially pleased because it’s based on more than attaining a certain degree of commercial success – the award also recognizes companies with great vision, ambition and customer focus, and that’s something that’s definitely part of our culture.