The Dual-Brain Architecture and the Global Humanoid Robot Race | 5th June 2026

**SPEAKER_1** (0:00)
Welcome to Colaberry AI Podcast, brought to you by Colaberry AI Research Labs and Carl Foundation.

**SPEAKER_2** (0:05)
Thanks for having me.

**SPEAKER_1** (0:06)
So imagine you've got this 90-pound machine. Right. And it's carrying a full tray of glass water pitchers.

**SPEAKER_2** (0:12)
Sounds like a disaster waiting to happen.

**SPEAKER_1** (0:14)
Oh, absolutely. And then it steps on a slick batch of floor. In about 150 milliseconds, that machine has to somehow calculate a completely new center of gravity.

**SPEAKER_2** (0:24)
Right.

**SPEAKER_1** (0:25)
It has to adjust 27 independent joints and apply exactly 120 newton meters of torque to its left knee just to catch itself.

**SPEAKER_2** (0:32)
All while still processing your verbal command to just be careful.

**SPEAKER_1** (0:35)
Exactly. And today's deep dive is really about the deeply technical methodologies making that kind of impossible math a reality. We are looking at a stack of sources today that cut straight through all those CGI heavy marketing videos.

**SPEAKER_2** (0:50)
Yeah, there are a lot of those lately.

**SPEAKER_1** (0:51)
So many. We want to look at the actual engineering. So for you listening, we are examining dual brain architectures, edge computing and extreme hardware integration coming out of three major hubs. We've got China, Vietnam and the Nvidia ecosystem.

**SPEAKER_2** (1:05)
And honestly, it's fascinating because the industry has moved entirely past those old philosophical debates about generalized AI.

**SPEAKER_1** (1:13)
Right. They are just grappling with hard physics now.

**SPEAKER_2** (1:15)
Exactly.
The core challenge in robotics has always been this massive latency bottleneck. It's an agonizing balancing act between high-level reasoning...

**SPEAKER_1** (1:26)
Like figuring out what an object is.

**SPEAKER_2** (1:27)
Right. Or parsing a complex verbal command, balancing that with low-level deterministic control, which is basically the immediate non-negotiable math required to just keep a bipedal machine from collapsing under its own weight.

**SPEAKER_1** (1:41)
So to understand how engineers are actually solving that latency bottleneck, we can look at Shanghai-based Jaca Robotics. Yeah.

**SPEAKER_2** (1:48)
Their new compact humanoid, Pi, is a really brilliant architectural baseline.

**SPEAKER_1** (1:52)
It really is. Because Jaca intentionally avoided building some massive imposing machine. Pi stands just four feet, so about 1.22 meters, and weighs 92 pounds, but roughly 42 kilograms.

**SPEAKER_2** (2:03)
Very compact.

**SPEAKER_1** (2:04)
Yeah. But within that tiny frame, they've managed to engineer 27 degrees of freedom, which for anyone fuzzy on the terminology, a degree of freedom, or DOF, is essentially an independent axis of motion.

**SPEAKER_2** (2:18)
Right. So like a simple hinge joint is one degree of freedom.

**SPEAKER_1** (2:21)
And a human shoulder would be three.

**SPEAKER_2** (2:22)
Exactly. So managing 27 of those simultaneously in a 92 pound frame requires some incredibly dense hardware.

**SPEAKER_1** (2:30)
Yeah. The sources detail that their new integrated joint modules are actually 15 to 27 percent smaller than previous iterations.

**SPEAKER_2** (2:38)
And shrinking those joint modules alters the fundamental physics of the robot's locomotion. Because when you reduce the mass at the end of a limb...

**SPEAKER_1** (2:46)
Like the wrists and the ankles.

**SPEAKER_2** (2:47)
Yeah, exactly. You drastically lower the moment of inertia. So it takes significantly less energy to swing a lighter leg forward.

**SPEAKER_1** (2:54)
Which makes the center of gravity much easier to control.

**SPEAKER_2** (2:56)
Spot on. But the crazy thing is, they achieved that size reduction without sacrificing raw output. The knee joints on Pi can still generate up to 120 Newton meters of torque.

**SPEAKER_1** (3:07)
Which is that critical threshold we talked about earlier.

**SPEAKER_2** (3:09)
Right. Because it provides the sudden explosive force necessary for dynamic recovery, like catching its balance on that slippery floor image.

**SPEAKER_1** (3:16)
And it still supports a 3 kilogram payload capacity per arm, which is impressive.
But OK, the physical hardware is great. The real breakthrough, though, is how they control those 27 axes without the whole system just crashing.

**SPEAKER_2** (3:31)
Right. What Jaca calls their fusion brain architecture.

**SPEAKER_1** (3:33)
Yeah. Built on Intel's heterogeneous computing platform. And they use these two biological metaphors, which I love, the cerebrum and the cerebellum.

**SPEAKER_2** (3:42)
Right. So heterogeneous computing is really just the practice of using discrete specialized processing cores for different tasks.

**SPEAKER_1** (3:50)
Instead of pushing everything through one central processor.

**SPEAKER_2** (3:52)
Exactly. So the cerebrum layer is dedicated entirely to high level intelligence. It runs on high performance GPUs or NPUs.

**SPEAKER_1** (3:59)
This is the layer processing the large language models, right?

**SPEAKER_2** (4:02)
Yeah. LLMs, vision algorithms, application logic. If you tell the robot, go grab the red wrench, the cerebrum translates that audio wave.

**SPEAKER_1** (4:11)
It identifies the wrench on the camera feed.

**SPEAKER_2** (4:14)
Right. And formulates a plan.