Humanoid Robots Are Watching. Who's Watching Them?

Read enough robotics announcements and you develop a reflex. Strip out the torque specs, the degrees-of-freedom counts, the breathless language about "embodied intelligence" — and ask the question the press release never answers: what does this thing collect, and where does it go?

This week handed me three announcements worth running through that filter.

Jaka Pi: The Architecture That Actually Matters

Shanghai-based Jaka Robotics unveiled Pi, a compact humanoid standing about four feet tall and weighing roughly 92 pounds. The hardware story is legitimately interesting — 27 degrees of freedom, joint modules 15 to 27 percent smaller than previous generations, knee joints delivering up to 120 Nm of torque. Compact, capable, built for real-world deployment rather than conference theater.

But the design choice I keep coming back to is the dual-brain architecture. Jaka split Pi's cognitive system into what they call a cerebrum and a cerebellum. The cerebrum handles AI reasoning, vision perception, large language model processing, and environmental understanding. The cerebellum handles real-time motion control through an EtherCAT-based network with millisecond-level latency.

From a pure engineering standpoint, this separation is elegant — you need deterministic, low-latency control for movement, and you need a different kind of compute for understanding context and language. Keep them from stepping on each other and you get a more stable machine.

From a data standpoint, the cerebrum is the piece I'd want to audit. A robot that processes spoken instructions, runs vision through a large language model, and understands its environment in real time is generating a continuous stream of perceptual data. On Pi, that architecture runs on Intel's heterogeneous computing platform. The question Jaka hasn't answered publicly — and no one is making them answer yet — is what happens to the environmental and conversational data that the cerebrum processes. Does it stay onboard? Does it sync to infrastructure? Who can query it?

Pi is positioned as a research and development platform, so the immediate stakes are lower than a consumer or commercial deployment. But the architecture Pi normalizes today is the one that ships in higher-stakes contexts tomorrow.

Dino in the Wild: The Surveillance Question Vietnam Isn't Being Asked

VinBigData's humanoid Dino is the robot in this news cycle that should be getting more scrutiny than it is.

Dino debuted internationally at Computex Taipei 2026 and is being developed for — and this is the company's own language — "security and surveillance in urban areas, campuses, commercial spaces, and service complexes," alongside household assistance. That's a significant dual mandate, and the tension between those two use cases isn't just a design challenge. It's a data governance question dressed up as an engineering problem.

The most concrete deployment so far was at Vinpearl Safari Phu Quoc, where Dino operated as an autonomous robotic guide. It worked in outdoor conditions, used multilingual speech interaction, responded to visitor questions through natural language, and relied on real-time environmental awareness to navigate crowds. By the company's account, it handled the unpredictability of real public spaces reasonably well.

Now hold that picture. A robot equipped with environmental perception, autonomous navigation, multilingual speech processing, and situational awareness — operating in a commercial tourist venue, interacting with international visitors. Every multilingual exchange it handles is being processed somewhere. Every environmental scan is data. The faces, the voices, the movement patterns through a crowded outdoor space.

Vietnam does not have a data governance framework comparable to the EU's GDPR. The country passed a Personal Data Protection Decree in 2023, but implementation and enforcement are still developing, and the framework for how an autonomous humanoid's continuous perceptual feed should be handled — stored, retained, shared with whom, subject to what access controls — isn't established in the way that European regulators would require before a deployment like this.

That matters specifically for the international tourists at Vinpearl. A European visitor whose biometric and conversational data is processed by a Vietnamese-operated autonomous system is in a regulatory grey zone that neither their home country's protections nor local Vietnamese law currently resolves cleanly.

I want to be clear about what I'm raising here: this isn't an accusation. I don't know what VinBigData's data practices are. Neither does anyone outside the company — because the company hasn't said. And that absence of disclosure is precisely what should concern anyone thinking about where humanoid security robots are headed. A robot built for surveillance doesn't need bad intentions to create a surveillance problem. It just needs to collect data in a context where no one has established who owns it, who can access it, and how long it persists.

(Note: VinBigData's robotics division has been referred to in reporting as both Vin Big Dynamics and VinDynamics — I've used VinBigData here as the parent entity; the correct operating brand for the robotics unit is worth confirming before any regulatory or legal analysis.)

NVIDIA's "Open" Platform: Open to Whom, Exactly?

NVIDIA's entry into this week's news is structurally different. Rather than launching a finished robot, the company unveiled an open humanoid robot reference design built on its Isaac GR00T platform, centered on the Unitree H2 — a nearly six-foot, 150-pound human-scale platform with 31 degrees of freedom, paired with tactile five-finger hands that bring the total to 75 degrees of freedom. The onboard compute is NVIDIA's Jetson Thor system; the published AI performance figure from NVIDIA's own materials for this platform is over 2,000 FP4 teraflops, though the specific module designation and exact figures should be confirmed against NVIDIA's current spec sheets before treating them as settled.

The pitch is genuinely compelling for the research community. Instead of forcing every lab to assemble a robot stack from incompatible components, NVIDIA is offering a full workflow: Isaac Teleop for demonstration data capture, Isaac Sim and Isaac Lab for virtual training, Isaac ROS for deployment. Institutions including ETH Zurich, the Allen Institute for AI, UC San Diego's robotics lab, and the Stanford Robotics Center have committed to using it. (The video sources a quote attributed to Stanford's Steve Cousins — his exact current title and role should be independently confirmed before publication.)

When NVIDIA says this platform is "open," they mean researchers can access the hardware reference design, the software stack, and the foundation models. That's meaningfully open relative to closed commercial systems.

But "open" in robotics AI is not the same as "auditable." Here's the question I'd ask before deploying any system trained on NVIDIA's Isaac GR00T stack: what data is the robot collecting during operation, and what does it transmit back into training pipelines? When a robot trained on GR00T runs in a real environment — a factory floor, a research lab, eventually a hospital or a school — its cameras and sensors are continuously generating data about that environment and the people in it. The "open" in open platform typically describes the training tools and model access. It doesn't automatically mean transparent data flows during deployment.

This is worth watching carefully as GR00T moves from university robotics labs into institutional and eventually commercial settings. The research community tends to be thoughtful about these questions. The institutional buyers who follow them often aren't.

What to Ask When the Robot Shows Up

I don't cover robotics because robots are scary. I cover it because the security and privacy questions that matter most get asked after deployment, not before — and by then the architecture is set, the contracts are signed, and the data is already flowing.

So when a humanoid robot appears in your workplace lobby, your shopping center, your child's school, your city's transit hub — here's your working checklist:

What does it record? Video, audio, speech transcripts, biometric data, behavioral patterns? Ask specifically.

Where does that data go? Onboard processing only, or does it sync to cloud infrastructure? Whose cloud? Under what jurisdiction?

Who can access it? The operator? The manufacturer? Law enforcement? Third-party analytics vendors?

How long is it retained? "We delete it after X days" is a policy. Policies can change. What does the contract say?

What's the legal framework? If the robot is operated by a company headquartered in a different country, which data protection law actually applies to you?

The robotics industry is moving faster than the regulatory frameworks that govern it. That's not unique to robotics — it's the recurring story of every technology wave I've covered. But humanoid robots are distinct in one important way: they're not an app on your phone that you choose to install. They're infrastructure. They'll be in spaces you don't opt into, making decisions about who to watch and what to remember, in environments where the expectation of privacy is ambiguous at best.

Jaka Pi's dual-brain architecture is a genuinely clever piece of engineering. VinBigData's Dino demonstrated real capability in a messy outdoor deployment. NVIDIA's GR00T platform will probably accelerate academic robotics research by years. None of that is in question.

What is in question — and what nobody launching these platforms is rushing to answer — is who owns the continuous perceptual record of the world that these machines are building, one interaction at a time.

---

Rachel "Rach" Kovacs is Buzzrag's cybersecurity and privacy correspondent.