6 Questions With … Anu Saha

The artificial intelligence revolution is here and it is bringing along with it robotics. ChatGPT and other large language models are transforming supply chains.

One of the company’s at the forefront of that revolution is Covariant with its Robotics Foundation Models (RFM).

 Anu Saha, head of product marketing at Covariant, joined Supply Chain Management Review to discuss more of what  ChatGPT and AI robotics have in common and more broadly, AI’s impact on fulfillment in this 6 Questions With … feature.

SCMR: What do supply chain executives need to know about the recent AI advancements like ChatGPT, Gemini, and Sora?

SAHA: Applications like ChatGPT have revolutionized AI for language. In the past two years, we have seen generative AI applications perform seemingly any digital task. The key behind these capabilities is foundation models, which will one day be the basis of all AI-powered software.

These networks are trained on massive amounts of data without specific use cases. Because a single model is trained on a wider set of tasks, performance on each particular task improves. For example, the same AI model enables ChatGPT to write your cover letter, reply to customer complaints, code video games, and translate dozens of languages.

SCMR: What do these AI advancements have to do with logistics?

SAHA: Traditionally, robots have been explicitly programmed to perform tasks in a constrained, predictable environment. We’ve seen autonomous robots at scale in industries like automotive manufacturing where tasks are repetitive and unchanging.

However, a warehouse or fulfillment center is a highly dynamic environment with constantly changing items, complex workflows, and varying tasks. To unlock the next level of efficiency and reliability in our fulfillment networks, we need robots that can adapt to dynamic environments without sacrificing speed and accuracy.

Robots powered by foundation models are not trained or programmed for a single, constrained task, so they can adapt to unforeseen changes and maintain more reliable performance. These models equip robots with the capability to comprehend their surroundings and to adapt their actions according to changing circumstances.

SCMR: Why now? Why is it necessary to automate logistics with Robotics Foundation Models?

SAHA: Automation is critical to meeting the performance requirements of today’s logistics operations. We see this mandate to automate driven primarily by immediate (e.g. labor availability) and long-term needs (e.g. operational resiliency).

Even in the world’s most automated facilities, the picking, packing, loading, and unloading activities within fulfillment centers remain largely manual—representing up to 60% of warehouse labor. That’s because automating item handling requires an adaptability that traditional robotics can’t offer.

These item-handling tasks are dangerous and repetitive activities with high turnover rates. Our customers report that about 50% of new warehouse workers quit within the first 60 days. In peak season, that tenure can be reduced to less than two weeks. This new generation of robots is coming at a time when we are facing an urgent labor problem that could cripple e-commerce growth and the global economy if not solved.

SCMR: Can you share some real-world use cases?

SAHA: In warehouses and fulfillment centers worldwide, robots powered by foundation models are helping organizations adapt to a variety of operational changes without human intervention, downtime, or throughput disruptions.

At the core, Robotics Foundation Models enable robots to discover new ways to handle never-before-seen items while enabling non-technical operators to also instruct robots on handling new, challenging situations in just minutes (alternatively, extensive reprogramming can take weeks or months of an engineer's time).

Here are some examples:

• AI-powered robots can depalletize mixed-SKU pallets with heavy cases and shrink-wrapped items. To avoid costly interventions that disrupt throughput, the robot must also be able to identify distractor items (empty trays or slip sheets) and discard them appropriately. Since each pallet is packed differently and the distractor items are not predictable or consistent, Robotics Foundation Model’s generalized understanding helps robots adapt on the fly.
• Robots powered by a Robotics Foundation Model can detect damaged products with torn packaging or broken items. Upon detection, those items can be disposed of in an exception bin. In these situations, robots understand what a cereal box looks like and what a damaged or broken cereal box looks like.

SCMR: Explain how Robotic Foundation Models differ from previous generations of AI-powered robotic automation.

SAHA: The key to foundation models is the training data. AI models like ChatGPT have benefited from large amounts of data that are widely available on the internet. However, until recently, this scale of data has not been available to train a Robotics Foundation Model because robots need to be trained on data from real-world environments. Internet data or lab settings are not enough.

• Volume: Large amounts of data must be captured from hundreds of connected robots operating in warehouses and fulfillment centers worldwide.
• Variety: To create a truly generalized model, data must capture different use cases (e.g., robotic induction, sorting, order picking) and a wide variety of items and SKUs (e.g., apparel, health & beauty items, parcels, produce, groceries, electronics, and so on).
• Multimodal: Instead of just one type of data, like text, robots need much richer physical data, such as sensor information, videos of robots performing warehouse tasks, images of SKUs and the robot environment, actions the robot intended to perform, etc.

This scale of high-fidelity data powers robots to autonomously adapt to unforeseen changes in operations, which unlocks operational efficiency that was not possible with previous generations of robotics.

SCMR: Any parting advice for our readers? 

SAHA: While new, the growth trajectory of Robotics Foundation Models is accelerating rapidly. Robotic applications, particularly within tasks that require precise object manipulation like picking, packing, loading, and unloading, are already being applied in production environments by some of the world’s most innovative logistics and retail leaders.

I predict we’ll see an exponential number of commercially viable applications deployed at scale in the next few years, which means the time to explore how your organization can apply foundation models is now.

SCMR: Thank you.

SC
MR