Interesting! Question: when the robo-dog toppled over because of low battery/high heat, did you have to put new batteries to get it functioningn again? Or is its weight light enough to be picked up and carried in cases like that?
30 pounds, even with a carrying strap, could prove a considerable barrier for children, the elderly, smaller adults, and people with certain disabilities. The risk of having to call for help or abandon a $4,000 device on the sidewalk when it malfunctions seems to me to be too high to be appealing.
Thanks for this! You mentioned that Unitree has been shipping the quadrupeds for a few years. Do you know if it's mostly to the research/enthusiast market or if they have commercial applications too?
I don't know of detailed data but very roughly my sense is that academic research is the biggest market, with consumer enthusiasts and commercial entertainment (like using a robot in a performance or as a novelty at a shopping mall) as runner-up applications. People talk about using quadrupeds for applications like factory inspections but I'm not sure how common that is in practice. But again that's just based on the anecdotes I've seen.
The core engineering bottleneck that lab demonstrations routinely gloss over is organic resilience and self-healing under extreme chaotic wear.
The Impact Problem: High-torque planetary gearboxes, cycloidal drives, and harmonic reducers hate sudden shock loads. Catching an errant hoof, taking a tumble down a rocky draw, or absorbing mud, grit, and torrential rain will shear gear teeth or seize bearings in short order.
Self-Repair vs. Supply Chains: Biological tissue can absorb dynamic micro-trauma, remodel bone density, regenerate muscle tears, and maintain its own lubrication continuous-duty for decades. A broken robotic limb requires a clean workspace, custom replacement actuators, specialized tooling, and downtime.
Environmental Sealing: The moment fine dust, river water, freezing slush, or abrasive grime meets an actuator seal or exposed rotary joint, friction spikes, internal temperatures climb, and the platform faults.
Until machines are grown out of self-repairing synthetic musculature rather than machined aluminum, copper coils, and silicon boards, the high-friction, unscripted punishment of the open range remains strictly biological territory.
Decades ago, when I started in technology, computers took up full rooms and required heavy duty air conditioning and water cooling. Today, our cellphones have more memory than that computer had and in fact, have far more memory than the whole computer system had attached in the form of external disk drives, which at that point cost many of millions of $$.
Ai is helping advances move much faster. Humanoid development is advancing on an exponential curve. Humanoid robots will likely be on battlefields in 5 years or less.
Maybe I'm too cynical, but it really feels like aside from GPUs from a single company the US has completely lost the ability to compete in global markets, and instead of trying to find a solution we're just relying on protectionism to kick the can down the road. I really struggle to believe that it's just Chinese subsidies, because even when we shower companies with money (Intel) it doesn't seem to lead to any improvements. Maybe it's a mix of factors (quarter to quarter thinking, US AI companies being run by cultists, B2B focus leading to no consumer feedback, etc.). I'm kind of just ranting since this is only tangentially related to your article.
I agree it's not just subsidies. China has become the center of the global manufacturing economy, and that gives their companies a leg up in electronics in much the same way Silicon Valley gives the US a leg up in high tech.
Do you think consumer-facing tech like this is meaningfully related to increasing overall awareness/salience of the 'electric tech stack'? Currently I can't figure out what if any mechanics are likely to actually get the American public/American industry activated around this core technology, beyond something really bad happening like e.g. a large scale, high-visibility drone assault against an ally.
It doesn't hurt! I think US policymakers are paying attention, bootstrapping an industrial supply chain is just really hard. Biden tried to boost US chip manufacturing with the CHIPS Act and it didn't work that great.
Oh for sure - one of the reason I love all the CHIPS debriefs from e.g. Mike et al is that it really was kind of a vanguard effort to build subsequent industrial policy efforts atop, and I have mad respect for the Biden administration for attempting it. Agree that it's not just like "oh if you will it into being, it shall come" or etc etc.
Are these robots “open platforms” in the sense that the owner can write software to make it do new things? Or is the “crude app” the only way to control it?
Good question! When I was deciding which robot to get, I saw there was an "Air" model and a "Pro" model. I bought the more expensive "Pro" model for ~$4,000 hoping that I'd be able to program it. I didn't realize there's actually a third model called the Go2 EDU that costs $15,000. And only this third model allows third-party software out of the box. Even if I'd known this my wife would not have let me spend $15,000 on a robot for an article lol.
I looked into jailbreaking the Pro. It looked non-trivial, and I worried I might damage the robot and thereby hurt its resale value.
I hear it's incredibly easy to jailbreak, but when you do you realize it's full of spyware that's really hard to remove. I remember seeing a youtube video about it...
* Their "AI enabled" versions are subject to attacks that can bypass safety guardrails. Leveraging RoboPAIR, researchers bypassed the safety guardrails built into the LLM. The AI governor failed to differentiate between safe, regular operating requests and malicious commands cloaked in adversarial prompt engineering. UniPwn & UniBLEed (CVE-2026-76639 / CVE-2026-76640)
* They steal designs and replicate them (Unitree's $1,600 Go2 model reportedly matches, almost to the millimeter, a robot called Mini Cheetah developed under the US Army's DEVCOM Army Research Laboratory program)- hardly a recipe for building trust among consumers
The take-aways: before we start wagging our collective tails (robotic or otherwise) about "cheap" robots, we should understand what we are really getting...
- Trustworthiness matters
- Provenance matters. Where did the design come from?
- Transparency matters. What does it really do?
- REAL safety testing and protocols matter.
I'm not bashing CN robots or that "robots suck."
Rather, it is important to assess the risk of chasing the latest shiny bauble when:
- there is a lack of understanding of the threat and potential risks
- a lack of oversight regarding what is being sold to people and how it might be (mis)used
- and no real privacy, safety and security regulations regarding consumer robotics.
Before someone brings a system capable of causing real harm into their household, there should be a way of ensuring its safety and the vendor should earn the trust of the consumer.
Interesting! Question: when the robo-dog toppled over because of low battery/high heat, did you have to put new batteries to get it functioningn again? Or is its weight light enough to be picked up and carried in cases like that?
It weighs about 30 pounds and has a handy strap on the back. So I picked it up and carried it home.
Thank you! I missed that detail in your article.
I'm not sure I mentioned it.
30 pounds, even with a carrying strap, could prove a considerable barrier for children, the elderly, smaller adults, and people with certain disabilities. The risk of having to call for help or abandon a $4,000 device on the sidewalk when it malfunctions seems to me to be too high to be appealing.
Thanks for this! You mentioned that Unitree has been shipping the quadrupeds for a few years. Do you know if it's mostly to the research/enthusiast market or if they have commercial applications too?
I don't know of detailed data but very roughly my sense is that academic research is the biggest market, with consumer enthusiasts and commercial entertainment (like using a robot in a performance or as a novelty at a shopping mall) as runner-up applications. People talk about using quadrupeds for applications like factory inspections but I'm not sure how common that is in practice. But again that's just based on the anecdotes I've seen.
Just seeing this article an hour or so after hearing that ICE is buying a bunch of robot dogs. It was reported on Democracy Now radio show.
Why robots cain't ever be cowboys.
The core engineering bottleneck that lab demonstrations routinely gloss over is organic resilience and self-healing under extreme chaotic wear.
The Impact Problem: High-torque planetary gearboxes, cycloidal drives, and harmonic reducers hate sudden shock loads. Catching an errant hoof, taking a tumble down a rocky draw, or absorbing mud, grit, and torrential rain will shear gear teeth or seize bearings in short order.
Self-Repair vs. Supply Chains: Biological tissue can absorb dynamic micro-trauma, remodel bone density, regenerate muscle tears, and maintain its own lubrication continuous-duty for decades. A broken robotic limb requires a clean workspace, custom replacement actuators, specialized tooling, and downtime.
Environmental Sealing: The moment fine dust, river water, freezing slush, or abrasive grime meets an actuator seal or exposed rotary joint, friction spikes, internal temperatures climb, and the platform faults.
Until machines are grown out of self-repairing synthetic musculature rather than machined aluminum, copper coils, and silicon boards, the high-friction, unscripted punishment of the open range remains strictly biological territory.
Decades ago, when I started in technology, computers took up full rooms and required heavy duty air conditioning and water cooling. Today, our cellphones have more memory than that computer had and in fact, have far more memory than the whole computer system had attached in the form of external disk drives, which at that point cost many of millions of $$.
Ai is helping advances move much faster. Humanoid development is advancing on an exponential curve. Humanoid robots will likely be on battlefields in 5 years or less.
A possible use case: the dog can walk the human.
This would give the human some exercise
Or put a seat on it and it could replace annoying scooters as transportation.
Maybe the “dog” can walk the dog, so the human owner of both can remain seated on the couch :)
Maybe I'm too cynical, but it really feels like aside from GPUs from a single company the US has completely lost the ability to compete in global markets, and instead of trying to find a solution we're just relying on protectionism to kick the can down the road. I really struggle to believe that it's just Chinese subsidies, because even when we shower companies with money (Intel) it doesn't seem to lead to any improvements. Maybe it's a mix of factors (quarter to quarter thinking, US AI companies being run by cultists, B2B focus leading to no consumer feedback, etc.). I'm kind of just ranting since this is only tangentially related to your article.
I agree it's not just subsidies. China has become the center of the global manufacturing economy, and that gives their companies a leg up in electronics in much the same way Silicon Valley gives the US a leg up in high tech.
Definitely true, and I probably should have clarified I didn't mean I thought you were arguing it was just subsidies.
No worries!
来自中国的问候！王兴兴确实是一个很有意思的人
They should put fur on it!
Do you think consumer-facing tech like this is meaningfully related to increasing overall awareness/salience of the 'electric tech stack'? Currently I can't figure out what if any mechanics are likely to actually get the American public/American industry activated around this core technology, beyond something really bad happening like e.g. a large scale, high-visibility drone assault against an ally.
It doesn't hurt! I think US policymakers are paying attention, bootstrapping an industrial supply chain is just really hard. Biden tried to boost US chip manufacturing with the CHIPS Act and it didn't work that great.
Oh for sure - one of the reason I love all the CHIPS debriefs from e.g. Mike et al is that it really was kind of a vanguard effort to build subsequent industrial policy efforts atop, and I have mad respect for the Biden administration for attempting it. Agree that it's not just like "oh if you will it into being, it shall come" or etc etc.
Are these robots “open platforms” in the sense that the owner can write software to make it do new things? Or is the “crude app” the only way to control it?
Good question! When I was deciding which robot to get, I saw there was an "Air" model and a "Pro" model. I bought the more expensive "Pro" model for ~$4,000 hoping that I'd be able to program it. I didn't realize there's actually a third model called the Go2 EDU that costs $15,000. And only this third model allows third-party software out of the box. Even if I'd known this my wife would not have let me spend $15,000 on a robot for an article lol.
I looked into jailbreaking the Pro. It looked non-trivial, and I worried I might damage the robot and thereby hurt its resale value.
I hear it's incredibly easy to jailbreak, but when you do you realize it's full of spyware that's really hard to remove. I remember seeing a youtube video about it...
Robo dogs have existed since the 90s. Nobody buys them. There is no market
Unitree is problematic on many fronts:
* They ship with undocumented features that can enable surveillance and information theft. Multiple technical reviews showed Unitree units systematically sending unencrypted operational telemetry back to remote servers in China without explicit user authorization. https://roboticsbiz.com/robot-cybersecurity-vulnerabilities-what-cve-2026-8153-and-unipwn-reveal-about-robot-safety/
* Their "AI enabled" versions are subject to attacks that can bypass safety guardrails. Leveraging RoboPAIR, researchers bypassed the safety guardrails built into the LLM. The AI governor failed to differentiate between safe, regular operating requests and malicious commands cloaked in adversarial prompt engineering. UniPwn & UniBLEed (CVE-2026-76639 / CVE-2026-76640)
(https://thehackernews.com/2026/08/two-unitree-g1-edu-humanoid-robot-flaws.html)
* They steal designs and replicate them (Unitree's $1,600 Go2 model reportedly matches, almost to the millimeter, a robot called Mini Cheetah developed under the US Army's DEVCOM Army Research Laboratory program)- hardly a recipe for building trust among consumers
(https://www.detroitnews.com/story/tech/2026/08/18/how-us-military-funding-propelled-china-robot-dogs/91349568007/)... among other sources
The take-aways: before we start wagging our collective tails (robotic or otherwise) about "cheap" robots, we should understand what we are really getting...
- Trustworthiness matters
- Provenance matters. Where did the design come from?
- Transparency matters. What does it really do?
- REAL safety testing and protocols matter.
I'm not bashing CN robots or that "robots suck."
Rather, it is important to assess the risk of chasing the latest shiny bauble when:
- there is a lack of understanding of the threat and potential risks
- a lack of oversight regarding what is being sold to people and how it might be (mis)used
- and no real privacy, safety and security regulations regarding consumer robotics.
Before someone brings a system capable of causing real harm into their household, there should be a way of ensuring its safety and the vendor should earn the trust of the consumer.