Have you ever watched a soccer match where the players were comprised entirely of robots?

Peter Puchwein has. In fact, he co-designed those soccer-playing robots while at university in the early 2000s. (Fun aside: take a look at the premier robotic soccer playing tournament, RoboCup!).

Without the capabilities of today’s high-resolution cameras, he and his co-designer relied on ultrasonic sensors to navigate. A great idea, until the robots’ 360-degree ability led to a signal interference nightmare between ultrasonic sensors on multiple robots that rendered them incapable of looking anything like Pelé.

But Puchwein, the longtime VP of Innovation at KNAPP, one of the world’s market leaders in logistics and automation, didn’t take the moment as a failure. Instead, he went back to his algorithms and wrote code to filter out the noise.

This story, as told to Covariant’s Chief Scientist, Pieter Abbeel, in an interview on The Robot Brains podcast, captures not only Puchwein’s innovative spirit but also the heart of how he and his team drive innovation at KNAPP – and in turn, how that innovation has fundamentally changed the world of fulfillment and logistics.

We’re revisiting that interview below with key takeaways on what it really means to innovate, why now is the right time to do it, and why AI Robotics is central to doing so.

What is innovation, really?

What is innovation, really?

Puchwein’s role at KNAPP is to advocate for investment in innovation and R&D and then to make it happen.

Says Puchwein, “Innovation means doing something new and having the courage to actually do it.”

This is the job, mission, and goal of the innovation team at KNAPP. But the company doesn’t view innovation as siloed. Anyone, whether they’re an engineer or a salesperson, can submit an idea on the company intranet or join an intensive workshop to develop those ideas. Then the innovation team evaluates which ideas are worth pursuing and operationalizes them.

This cross-team collaboration extends to how KNAPP views its relationship with both customers and solutions providers, such as Covariant: as long-term partners. But often, the best ideas come from being on the warehouse floor with customers to deeply understand their pain points. Then it’s time to go back to the drawing board and customize solutions – or think of new ones entirely.

Modern warehouse automation

What innovation looks like in an automated modern warehouse

Software optimizes operations

During the Industrial Revolution, technological innovation was centered around repetitive mechanical processes such as welding, painting, and some assembly, all of which occurred in stages. One set of tasks led to another set of tasks which led to another set of tasks predictably down the assembly line.

“The big difference in warehousing,” says Puchwein, “is that there are so many factors that you don’t know…that influence the system at any time and you have to react to [them]. Not everything is predictable. For instance, if you have bad weather outside, a lot of people buy clothing for rain or umbrellas. This is something you don’t know maybe two days before.”

The same goes for quickly re-prioritizing your operations. Puchwein gives the example of a customer who unexpectedly runs a sale on white t-shirts and suddenly goes from selling 800 t-shirts per day to 100K. This necessitates a rapid re-prioritization of operations and reorganizing of the storage of physical goods so that the right inventory is easily accessible.

Warehouse operations can’t stop or pause just because parameters have changed when, in fact, they’re changing all the time.

In order to navigate in this kind of environment, it’s essential for operators to optimize in real-time. This is a place where AI can and does play a very big role.

Says host Pieter Abbeel: “[This is] a reinforcement learning problem where effectively you’re constantly making decisions. The state of the warehouse changes [and] the outside world has stochastic input as far as the model is concerned. You need to make new decisions all the time with possible long-term consequences. If you store something in the wrong place now you’re stuck with that being harder to retrieve if you didn’t realize that was going to be popular.”

“With reinforcement learning, it’s really possible with the right information and rewards to train models that exactly take care of these kinds of parameters where you can optimize and you can train it,” adds Peter Puchwein.

Picking is the holy grail of intralogistics automation

In today’s innovative warehouses, much has been automated around the “feet” of the operations. Autonomous mobile robots (AMRs), conveyors, and other similar technologies handle getting items to stations where individual SKUs can be picked and packed. But the piece-picking has long proven difficult to automate. With millions of SKUs to process, each of which requires subtly different approaches for grasping and moving, many traditional attempts at robotic automation have failed, creating operational bottlenecks.

KNAPP, as Puchwein relates, tried several different picking technologies over the years, but they often struggled to grasp a product because of technical challenges, such as understanding ghost reflections from shrink wrap. There were many nice marketing videos, but in reality, nobody had a solution that worked in real-world environments to pick the infinite variety of SKUs that come through a warehouse.

Now, the advent of robotics powered by modern AI, such as the Covariant Brain, is changing all of that by making piece-picking automation possible at scale in real-world settings.

Innovation across the value chain

While optimization of warehouse operations is key, it needs to be considered within the context of the full value chain. Everything we’ve discussed so far shaves seconds or even minutes off of processing times within the warehouse, which has big implications across millions of SKUs and orders processed. However, those warehouse optimizations are neutralized if there is no optimization in the last mile of the delivery, as just one example.

As Puchwein puts it, “If you order something and you’re not home to get it, it doesn’t matter how much we optimized the high automation warehouse, we lose one day because nobody is at home. Those seconds don’t matter anymore [if we have to redeliver it].”

Solving piece-picking

How AI Robotics solve piece-picking challenges

In order to be successful, piece-picking robots need to understand what to do with items they’ve never seen before. To do that, they need to see, think, and act, using AI technology like the Covariant Brain, which is trained on millions of picks and continues to grow its capabilities through fleet learning. As Pieter Abbeel explains in the episode, technology like this is “trained on data. It’s not exactly what it will see in the future, but it’s sufficiently enough related that it will still make good predictions and decisions in these future situations.”

It’s not predicting what will happen in the next twenty years, it’s understanding the world for machines. That’s, for me, the power of AI.
Peter Puchwein

“If we think about a glass of water standing in front of us,” says Puchwein, “we don’t go with both fingers, with the trigger finger, and with the thumb five millimeters from the left and to the right, and then I close the fingers, that’s not the reality…The human goes with one finger to the glass, touches it a little bit, and understands where the glass is positioned and then closes the hand and takes it.”

That’s what robots need to do while piece-picking. AI software enables robots to do just that, testing the world around them, learning from their experiences, and passing their learnings onto the rest of the robots in the network. In stark contrast to what happens with hardcoding, which requires knowing exactly what each robot will face every time, the AI is trained before deployment on a customer’s items, but that’s just the base of knowledge from which the robot makes accurate predictions so it can try, learn, and grow – just like a human would.

The role of AI

The role AI plays in intralogistics and beyond

Friends and family often ask Peter Puchwein what the point of AI is.

“People always think AI is about predicting the future,” says Puchwein. “AI is not capable of predicting the next Grand Slam champion. It is capable of understanding the game without knowing the rules and making a machine able to play against a human…That’s AI for me. It’s not predicting what will happen in the next twenty years. It’s understanding the world for machines.”

Watch the full episode below, or listen on Apple, Spotify, Google, Amazon, or Acast.

Watch: The Robot Brains Podcast