{"id":23519,"date":"2026-08-25T02:38:51","date_gmt":"2026-08-25T02:38:51","guid":{"rendered":"https:\/\/scannn.com\/approaching-robotics-hardware-takeoff-by-chris-paxton\/"},"modified":"2026-08-25T02:38:51","modified_gmt":"2026-08-25T02:38:51","slug":"approaching-robotics-hardware-takeoff-by-chris-paxton","status":"publish","type":"post","link":"https:\/\/scannn.com\/lv\/approaching-robotics-hardware-takeoff-by-chris-paxton\/","title":{"rendered":"Approaching Robotics Hardware Takeoff - by Chris Paxton"},"content":{"rendered":"\n
\n
\n
<\/p>\n
<\/picture><\/div>\n

<\/a>

Image from ChatGPT<\/figcaption><\/figure>\n<\/div>\n

The <\/span>World Humanoid Games are underway in Beijing<\/a>, and the world is watching. Robots are sprinting, long jumping, boxing, and playing soccer. The results are, at times, comical. Sights of one robot after another slamming into a wall at the end of a race, or soccer-playing Booster robots tripping over one another into a gigantic heap, make this all seem like a bit of a joke sometimes.<\/span><\/p>\n

But feeling this way would be a mistake. Instead, I would argue that these games show a robust, mature ecosystem which can both build things and, crucially, isn\u2019t afraid to break them. And that\u2019s key to real progress.<\/p>\n

It\u2019s staggering how much different things are from last year. The robots are faster, more intelligent, and more capable. They\u2019re also a lot more flammable, with multiple instances of fast-moving sprinters exploding or bursting into flames due to the tremendous heat being generated.<\/p>\n

A very rapid pace of robotic hardware and software innovation is enabling a robust ecosystem, in which many different companies and laboratories can build high-quality products, and even develop one-off robots like the <\/span>Honor Lightning<\/a>, a humanoid robot designed purely for sprinting and which recently beat (human) Usain Bolt\u2019s 100 meter record.<\/span><\/p>\n

Unitree <\/span>sold over 5,500 humanoid robots in 2025<\/a>, mostly its G1 humanoids. And after an explosive IPO, it seems unlikely that this will change any time soon. There\u2019s tremendous demand right now for capable, usable, and deployable robots \u2014 and Unitree\u2019s humanoids are something of a gold standard for the industry.<\/span><\/p>\n

Their secret, in larger part, has been shipping and iterating. Unitree is ruthless about punishing and testing their robots. In 2024, they were hauling the first G1 around to every academic conference, watching the robot slowly accumulate scratches and nicks, even though it could at the time barely walk. At the same time, they would bring out their quadruped robots and kick them down the stairs, showing exactly how durable and reliable they were when the robot could be piloted right back up.<\/p>\n

They\u2019ve been iterating quickly on their humanoids, going from what used to be a relatively useless G1 to an incredibly popular and well-oiled product, in part powered by rapid hardware cycles, aggressive vertical integration, and intense optimization.<\/p>\n

And yet by most accounts they\u2019re not even the biggest producer of humanoid robots. That would be Agibot:<\/p>\n

\n
<\/p>\n
\"Image\"<\/picture><\/div>\n

<\/a><\/figure>\n<\/div>\n

Agibot is far less seen outside China, though robots like the Agibot X2 (their answer to the Unitree G1). US companies like Faraday Future have been <\/span>selling some suspiciously similar<\/a> robots to those Agibot offers. Agibot also sells a number of industrial semi-humanoids like the A2-W:<\/span><\/p>\n

These robots are designed for industrial use cases as they\u2019re somewhat easier to deploy in different industrial use cases. Robots like the <\/span>Agibot G2 support autonomous charging <\/a>and have been <\/span>shown in real industrial use cases<\/a> already. <\/span><\/p>\n

Companies like Galbot on the industrial side, <\/span>with its heavy-duty S1 robots<\/a>, and <\/span>Booster with its developer-focused K1 and T2<\/a>, also provide great robots. The <\/span>Galaxea R1<\/a> has established itself as a very standard manipulation research robot, starring in the <\/span>outstanding Stanford Behavior-1k benchmark<\/a>. In short, there\u2019s a massive number of companies, producing a truly amazing number of robots.<\/span><\/p>\n

And lest we think the companies mass producing mobile robots are ALL Chinese, American humanoid company Figure, clearly at this point the manufacturing leader in the United States, also built its 1000th robot in July of this year:<\/p>\n

And multiple other humanoid companies are gearing up for increased production as well. But no one seems close to reaching the scale of hardware depth and quality as the ecosystem apparent in China right now.<\/p>\n

Humanoid robot designs are both showing diversity, but also converging on a couple form factors around certain roles. We\u2019ve seen a number of robots pick up a wheeled elevator base pattern seen on <\/span>Sunday<\/a> and <\/span>Weave <\/a>robots, a number of full-sized humanoids like <\/span>Figure<\/a>, <\/span>1X<\/a>, Tesla, and the Unitree H2, and a range of smaller developer-focused robots. The biggest robot producers (like Unitree and Agibot) all have a range of different platforms now, filling different roles \u2014 though the biggest sellers are still development-focused, child-sized robots like the Unitree G1.<\/span><\/p>\n

\n
<\/p>\n
\"\"<\/picture><\/div>\n

<\/a>

TRON2 is a modular robot architecture, built by LimX Dynamics. <\/span>Source: X<\/a><\/figcaption><\/figure>\n<\/div>\n

Hardware and software innovation is everywhere. LimX, for example, has been developing its modular TRON series robots, and Fourier has built its <\/span>more care-centric GR-3 humanoids<\/a>, with soft and friendly bodies.<\/span><\/p>\n

This empowers learning research as well. Massive open datasets like the <\/span>Galaxea Open World Dataset<\/a> and <\/span>Agibot World<\/a> are invaluable for researchers, and couldn\u2019t exist without abundant hardware. It also enables research and development of <\/span>robot foundation models like Lingbot-VA-2<\/a>.<\/span><\/p>\n

It\u2019s a robust ecosystem that will pay dividends in the future, because one of the biggest barriers in robotics is just having robots, and having robots that can do useful things! No longer is this an issue. We have a lot of robots, for every possible niche, and the real barrier now is just making them smart enough. And as we saw with LLMs, good data is key to solving that problem.<\/p>\n

Think back to events like the DARPA Grand Challenges of the 2000s in the United States. DARPA asked universities and researchers to build a self-driving car, as American troops were dying to roadside bombs in Afghanistan and Iraq.<\/p>\n

The researchers gathered in the Mojave desert to drive along a 240-km route. Not a single vehicle completed the route. Next year, five vehicles out of 23 completed a new, 212-km route. In 2007, at the Urban Grand Challenge, only six out of 11 teams completed a much shorter 89-km route <\/span>through George Airforce Base<\/a>. <\/span><\/p>\n

Despite the atrocious failure rate, and the poor performance by human standards even of the winners, the finalists would go on to form the core of the self-driving car boom, founding companies like Waymo and Zoox which have autonomous vehicles on the roads today.<\/p>\n

And this is where we come back to the World Humanoid Games. Competitions like this might be filled with failures \u2014 <\/span>watch this amazing compilation of falls from the DARPA robotics challenge<\/a> \u2014 but they spread information, ideas, and help enable real progress, just the same.<\/span><\/p>\n

We\u2019re at a strange point in time for robotics. Robotic hardware \u2014 good hardware \u2014 is available in a way that has never been possible before. It\u2019s more robust \u2014 massively so \u2014 and robots are <\/span>increasingly both affordable and capable<\/a>. In this environment, it\u2019s thrilling to see stuff like the Humanoid Robot Games, which will I think inspire and drive forward a new generation of engineers and researchers who will build robots that are actually useful.<\/span><\/p>\n

One thing that strikes me, is that the most important thing when building robots seems to be <\/span>just building robots<\/em>. That\u2019s been a huge part of Unitree\u2019s success: they build a lot of robots. If a robot breaks, they figure out why, and they make it stop breaking in that way. <\/span>They build actuators<\/a> because those are the bottleneck in the price of a robot. They sell robots and make money doing it \u2014 <\/span>278.21 million yuan of profit in 2025 alone<\/a> \u2014 and their explosive IPO saw their company value rise to $50 billion. They\u2019re <\/span>the<\/em> company building and selling general-purpose robots at scale, and the recipe works for them. But, importantly, it did not start that way: it started with a line of robot dogs and a Unitree G1 that could barely put one foot in front of the other.<\/span><\/p>\n

So, even if right now a lot of these robots fall down and burst into flames, a massive wave of real, working robots seems like it\u2019s just on the horizon.<\/p>\n

Share<\/span><\/a><\/p>\n

Leave a comment<\/span><\/a><\/p>\n

<\/p>\n
\n

Unitree’s Impossible Trajectory Is Still Overlooked<\/p>\n<\/div>\n

We are witnessing the birth of another Chinese hardware giant. Three years ago, Unitree was a quadruped company. By last year, they parlayed quadruped dominance into creating and leading the humanoid market. This year, their G1 humanoids are finally entering into viable deployments, and three new designs are on the way, including their most-direct\u2026<\/p>\n

Read more<\/span><\/p>\n

3 months ago \u00b7 312 likes \u00b7 2 comments \u00b7 Reyk Knuhtsen, Niko Ciminelli, Jacob Rintamaki, Robert Ghilduta, Joe Ryu, Jeremie Eliahou Ontiveros, and Dylan Patel<\/p>\n

<\/a><\/div>\n<\/div>\n

Source link <\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

Image from ChatGPT The World Humanoid Games are underway in Beijing, and the world is watching. Robots are sprinting, long jumping, boxing, and playing soccer. The results are, at times, comical. Sights of one robot after another slamming into a wall at the end of a race, or soccer-playing Booster robots tripping over one another […]<\/p>\n","protected":false},"author":16,"featured_media":23520,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[143],"tags":[],"class_list":["post-23519","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/posts\/23519","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/users\/16"}],"replies":[{"embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/comments?post=23519"}],"version-history":[{"count":0,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/posts\/23519\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/media\/23520"}],"wp:attachment":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/media?parent=23519"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/categories?post=23519"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/tags?post=23519"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}