The buzz around artificial intelligence in robotics often focuses on flashy demos and theoretical capabilities. But what about machines working in warehouses, factories, and beyond? That's precisely what robotic experts Melonee Wise (Agility Robotics) and SK Gupta (GrayMatter Robotics) tackled in their recent conversation with Brian Heater from A3.
Their insights cut through the hype to reveal what's happening at the intersection of AI and robotics today. Watch the on-demand discussion and read some of the takeaways below.
TAKEAWAY 1: RELIABILITY IS KEY
There’s no doubt artificial intelligence is making robotics smarter, but you still need to focus on implementations that can survive the demanding world of industry.
Gray Matter Robotics has successfully deployed systems for surface finishing applications — sanding, polishing, and treating surfaces of various materials — where AI plays multiple roles in the technology stack.
"AI is needed to basically build a model of the part... to plan the trajectory of the robots, as well as to ensure safety and control. We also use AI for health monitoring and making sure that tools get monitored and changed at appropriate times," Gupta explains.
At Agility Robotics, Wise shares how they "use trained policies in quite a few different places... If the robot starts to lose its balance, it actually has a trained RL policy for basically restoring its balance." Why? Because when customers frequently demand uptime exceeding 99%, robots need to work consistently, providing confidence and results.
TAKEAWAY 2: DEMOS ARE ONLY A PART OF THE STORY
Know those amazing videos of robots performing complex tasks? Turns out there's usually more to the story.
"In the demos that I've seen, some of the technology looks promising, but you don't see the before and after. You don't see how many times it can do it reliably. You see a lot of staging," Wise observes. Just like with anything online, think about how many demos benefit from perfect lighting, carefully positioned objects, and cherry-picked success stories.
Gupta adds that there’s often a mismatch between what’s being researched and what’s needed in the field: "There is significant progress being made. But the question, ultimately, is that when we see the demo, it's very difficult right now to assess reliability." He stresses that even if a system achieves high reliability in testing, "that's definitely not gonna be anywhere close to being deployed in manufacturing," where failure tolerances are measured in parts per million.
TAKEAWAY 3: SIMULATION CAN'T REPLACE REAL-WORLD TESTING (YET)
While simulation environments have come a long way, both experts reveal where they still fall short. For training AI systems, simulations provide valuable data but often fail to capture the messiness of physical interactions.
"The challenge that we run into in simulation, especially high fidelity physics simulations, is that they have a lot of challenges. They run relatively slowly, and they're a little bit noisy," Wise explains. Sometimes, simulations hit weird physics glitches that completely undermine their training value.
Gupta breaks it down nicely: "If you are positioning a robot from point A to point B in free space, simulation is excellent for that kind of thing. But the moment the sanding tool makes contact with the workpiece… there isn't any simulation which can currently simulate [that effectively]."
This gap forces companies to spend months testing in the real world. As Wise puts it, "Until you get a month or two of data, it's really hard to draw a trend line through that data," — which means validating robotic systems takes time, patience, and a lot of troubleshooting.
TAKEAWAY 4: ROBOTS CAN'T "WING IT" WHEN HANDLING OBJECTS
One of the most eye-opening parts of the conversation is how robots handle objects. What may seem simple to us is, in fact, incredibly nuanced in industrial settings.
"When people are talking about these things, they're doing it from the naive perspective that you can just grab things," Wise explains. For example, Agility works with a bearing manufacturer where there are constraints on where, when, and how to touch parts of the object. “All of those things add up for how you look at and solve the problem,” Wise continues.
But that is just one example of many. Today’s manufacturers have to worry about contamination, electrostatic discharge, not damaging their products, and so on. It's much more involved compared to a simplified demo environment.
TAKEAWAY 5: SAFETY AND AI HAVEN'T FULLY MADE FRIENDS YET
When asked about AI's role in safety systems, Wise doesn’t sugarcoat it: "Right now? Pretty little."
The core issue is that "safety has a very high determinism and introspectability requirement. So if you put input A in, you expect output B," Wise explains. Current AI systems don't always give that level of predictability, especially when faced with unusual situations.
As robots start moving beyond protected work cells, new challenges emerge. "As we start using AI for walking around freely in the real world, navigating, manipulating things in the real world, the first big thing that we're going to have to accomplish is safe human detection," Wise notes.
Gupta also points out that safety isn't just about protecting humans; it's also about product integrity. "If there's a potential that you're going to feed an outer distribution data into [an] AI model, then it's going to produce a result that nobody knows what's going to come out of there," he says. That unpredictability could damage expensive parts or create dangerous situations.
Both experts think the near future will likely involve hybrid approaches — combining traditional safety systems with AI enhancements rather than going all-in on learned methods.
TAKEAWAY 6: HARDWARE DESIGN CAN MAKE OR BREAK YOUR AI
The conversation highlighted hardware design, something that often gets overshadowed by all the algorithm talk.
"The easier it is to model your hardware in simulation, and the more replicable the hardware is in simulation, the easier it is to make that SIM to real transfer," Wise comments.
Gupta adds how different types of complexity impact AI differently. Just like with most things in life, it’s about fine-tuning and adding layers to handle those complexities and improve.
Find Melonee Wise and SK Gupta — along with hundreds of other industry experts and pioneers — at the Automate Show and Conference. From thought-provoking conversations and eye-opening insights to seeing solutions in action, Automate is the place to experience the future of automation.
« View All Blogs