Robotics technology has evolved to a stage where pioneering scientists are holding deep discussions regarding the foundational structure of smart entities—focusing intently on the anatomy of robots. This entails a closer examination of how the cognitive "brains" of these non-human agents are configured and how they interface with their mechanical forms.

Just as humans possess an innate, highly integrated mind-body connection, sophisticated autonomous machines designed to navigate human environments require a corresponding architectural blueprint.

That foundational perspective drives AMD Chief Technology Officer and Executive Vice President Mark Papermaster, who helped launch the inaugural Physical AI Summit hosted by Imagination in Action on October 30.

During the opening sessions, Papermaster delivered a detailed presentation on connecting robot bodies and cognitive frameworks in unprecedented ways, outlining what lies ahead for the industry.

Reflex Systems for Robots

"Alongside the brain, you need reflexes," Papermaster explained. "You need that muscle memory, and a nervous system that reacts in the moment, without waiting to be told, without those instructions having to come to it. So, as we bring AI into the physical world, we’re really relearning that lesson: keep everything we’ve thought about the brain and, for that matter, classical computing, but build reflexes into the system."

He elaborated further on the mechanics of these reflexes:

"The reflex," he noted, "I like to think about it as sort of the next challenge, as we think about computing and AI, where it’s heading, because as AI moves into the physical world, it has to sense, it has to decide, and act in real time."

Robots in Manufacturing and Beyond

"We’re already at work with physical AI," Papermaster stated, "to try to bring physical AI to reality, to practical applications."

These concepts are already finding practical deployment on factory floors.

"Robots are bringing touch and vision into the production line," Papermaster explained. "I heard the stat earlier about how many new Amazon robots that they’ve brought on … they’re moving very, very quickly. And what does this mean? It means they’re able to catch errors before those errors can leave the assembly line. They’re working side by side with people, so that trust has been established in those environments, and it’s boosting productivity while supporting safety."

He emphasized that this operational capability involves vigilant defect detection.

"Defects kill you," Papermaster continued. "They kill your ability to yield products, and defects mean you could get wrong answers if you don’t have perfect test coverage. So there’s really no margin for error."

Pushing humanoids into novel operational domains introduces unique hurdles, yet the trajectory remains remarkably promising.

"We at AMD are very aspirational in what we are doing in driving computing," he added, "because what we believe is that AI amplifies human potential. And from the kind of deployments of robotics that we described today, to tomorrow’s possibilities, it’s just amazing to think about what we can do to be an accelerator, and what humanity can do."

Latency, Power, and Safety

The conversation shifted toward the core triad of attributes required to optimize physical robotics: power, safety, and low latency for rapid reflexes.

"We’ve got vision-language-action models, whole-body controllers, world action models, and in the cloud, we optimize for throughput," he remarked regarding latency constraints. "All of those are very demanding from a compute standpoint, and a compute throughput standpoint. But in a robot, the key is also latency. You can’t miss that time-domain requirement. The robot has to respond within a real-time window, and the brain is planning, and it can be in the tens of milliseconds, but those reflexes underneath really have to respond in milliseconds, and the joints, microseconds. So it’s orders of magnitude of additional challenge."

Addressing data center power dynamics, he highlighted:

"We think about power every day. Power is performance. If I can get more power efficiency, I get more performance."

This dynamic translates directly into physical robotics.

"Most of the robot’s power goes to the motors. And so, as we’re building increasingly capable models, the robot can really only shed so much heat and carry so much battery," he explained. "So power optimization is actually a key constraint."

Safety represents the final pillar of the equation.

"These machines are going into all sorts of environments, right next to people," Papermaster noted. "In the cloud, we talk about security. We need that equally in robotic applications. But we also need safety to be, you know, just an absolute number one design factor. And it means really distributed safety systems that are fail-operational. They keep operating safely when something goes wrong, rather than just coming to a full stop. And capturing this opportunity requires fundamentally different computing architecture, where workload placement and compute selection are really central to the very system design."

Holistic Design and Conclusions

Papermaster later emphasized the necessity of closing control loops in advanced robotics.

"You start thinking about it from the top down, from the brain," he stated. "That’s where the planning and the bulk of the AI processing happens, and the demands keep growing, because robots are adding more cameras, more sensors, and that the brain has to be able to handle more and more (input) … more of that processing in those tens of milliseconds that that I mentioned. And then, the decision is only useful if you can act on it."

The dialogue also touched on holistic robotic architecture, open ecosystems, and cloud-based development before opening the floor to questions from MIT CSAIL Lab Director Daniela Rus.

Rus questioned the panel on the critical triad of safety, power, and low latency.

"I look at the work that we’re doing with Ramin and Liquid AI," Papermaster responded, referencing collaborations with Ramin Hasani on liquid models. "I’ve personally been engaged in I don’t know how many brainstorming sessions between AMD lead engineers and Ramin’s key engineers at Liquid AI. And the optimization must go from the model, and its tuning, down to our engine implementation, and our networking, and the enablement that we do physically. And I have to tell you, it’s a huge challenge. … in the world of physical AI, you have to think about that distributed computing. You can’t achieve the latency and that reflex response you need without full-stack optimization. It’s a big challenge, and it’s one that we’re going to all need to collaborate on very closely."

When pressed further regarding energy demands and the requirements for next-generation machines, Papermaster concluded:

"Power is performance: that’s true in the data center. It’s even more so in robotics. We really have to drive (with) power, with everything else that we do, including safety … we’re going to need yet more holistic design optimization across our open AI stack."