Chinese humanoid robots

It runs real ROS, but the hobby servos and Pi compute mean it is a teaching tool with a hard ceiling, not a path to research.

It is a delightful classroom toy with a chatbot bolted on, but the hobby servos and Python-only SDK are a dead end for real robotics research.

The $1,400 price only exists inside China; getting one to Europe means a gray-market import at four times the sticker and no warranty.

This is a 37 cm classroom teaching robot with no hands and about an hour of battery: it will delight students learning to code and do computer vision, but it will not fetch, carry, or do any real-world work.

The Air price does not buy the R1 hardware or EDU development access. Both consumer editions are closed to custom code; choose EDU for programming.

The backflips are real and the price is right, but this is an agility showpiece for classrooms and demos, not a manipulation or autonomy platform.

This is the rare cheap humanoid with no SDK paywall: even the entry Geek tier is fully programmable, so you are not buying a locked demo box.

This is the R1 you can actually code, but the one-hour battery and configuration-dependent hardware mean extra equipment depends on the work you plan to do.

The most interesting price in the catalog and the thinnest paperwork behind it: orders open in China and, since 29 Sep 2026, in the US from $18,000, but no product page, no spec sheet, no checkout and no independent delivery evidence yet, so treat the 89,900 RMB figure as a domestic sticker and get the US terms in writing.

You are buying a purpose-built cleaning appliance, not a programmable humanoid: it has two-finger grippers rather than hands, no public SDK, no Western certification or support, and the $13,600 figure is the China fleet price, not what a Western reseller (near $44,000) will charge you.

You are buying a lifelike face and a good conversation, not a helper: it will talk with you and remember you, but it will not walk your dog or wash a dish, the software is Chinese-first, and outside mainland China there is no official channel, warranty, or support.

Now in mass delivery in China, but it is brand-new hardware with little independent validation and no channel outside China, so treat the aggressive price as an early-adopter bet, not a proven product.

Buy it for the open stack and the datasets behind it, not for hands or autonomy: out of the box it is a locomotion demo you develop on.

It walks beautifully out of the box, but the base G1 is a sealed demo unit: no hands, no code, no way to ever unlock it, so if you plan to develop, only the EDU counts.

Forget the famous $12k headline: that promo expired, EngineAI now lists it at about RMB 188,000 (~$26k), and even then you get open motion code and no hands, not a working worker.

The human-like walk is genuinely impressive, but with quote-only ordering and no set sticker, you are buying an R&D chassis, not a product with a fixed price.

The precision is real and the vendor is established, but the base price does not lock in the dexterous hands the demos show, and unstructured autonomy still has to be programmed.

The Pro's Thor compute is genuinely top-of-class, but the arms ship with a bare configurable end effector: budget roughly $11,000 for grippers or $17,000 for the dexterous-hand pair on top of the base.

The K2 is a real, mass-produced industrial humanoid with genuinely strong dexterous hands and a full-shift battery, but the ~¥248,000 headline is the China Basic price: what you actually pay overseas is a quote after margin, shipping, and duties, and what you receive is a capable demonstrator that your own engineers must teach before it does a single one of your tasks unattended.

Confirm the distributor is really authorized before you wire five figures abroad, and remember the hands and spare battery are extra on top of the sticker.

There is no product to unbox: this is a free design plus a parts list, so the real cost is the workshop, the sourcing, and the months of assembly it takes to get one walking.

This is a performer and greeter, not a worker: 3 kg per arm and a fist-shaped standard hand mean it dances and interacts rather than manipulates, and outside China and South Korea the only stock we can find is a Dutch reseller at EUR 75,999 excl. VAT, far above the China price (LimX's own rollout is a 2027 target).

It undercuts the H1 dramatically on paper, but there is no western storefront or support, so getting one to Europe or the US is a China-import project you run yourself.

You are buying a polished showroom host, not a general-purpose or developer robot: it walks slowly, lifts 1.5 kg, runs about 2 hours per charge, and has no public SDK. Its impressive moments (opening a car door, serving a drink) are a narrow demonstrated set, with VR teleoperation there for tasks beyond it.

This really is running in real Chinese factories with a fully open control stack, at a price no Western full-size humanoid comes close to. But modular means you are buying a configuration, not a robot: legs, wheeled base, and each hand type are separate choices, the five-finger hands are an upgrade over the base gripper, and outside China you are importing by quote with no local support. Leju publishes no datasheet either, so even the specs move between sources (payload is quoted at 14 kg on one database and 20 kg in launch coverage), which is one more reason to get your exact build written into the quote.

It is a fully open developer platform with no SDK paywall, but the sticker is bare: budget several thousand more for the grippers or hands, because the arms ship empty.

This is the G1 you actually program, but the sticker is just the entrance fee: add a GPU workstation, hands, and a person who knows ROS2, or it sits in the crate.

You are buying a capable ROS 2 mobile-manipulation platform, not an autonomous worker: the grippers are standard (hands cost thousands more), and it does nothing useful until your engineers teleoperate it and train the skills yourself.

It is genuinely good at reception and guiding out of the box, but heavy manipulation is an integration project, and the impressive-looking Ultra hides its price.

This is not the bipedal Adam and it does not do chores: it is a data-collection station on wheels that a human drives with the included mocap suit, so you are buying it to generate training data, and it stays a demo until your lab trains a policy on that data.

The $29,900 headline is a bare China price on a locked demo unit: add import, duty, VAT, hands, and the EDU upgrade before it does anything but dance on a remote.

It is the H2 you can finally program, but doing real manipulation means paying separately for the Thor compute module and hands, and rigging a gantry for a 70 kg biped.

You are buying a development platform, not a worker. It walks and grips; anything task-shaped you build yourself, and the dexterous five-finger hands are a separate quote on top of a machine that already costs about 60,000 dollars.

It is a real, deployed picking-and-packing machine, but you are buying a wheeled two-armed mobile manipulator with a suction cup and a gripper (not a walking, five-fingered humanoid), and outside China you get it through a reseller at roughly $120K with import and support on you.

The Western $100k price is roughly triple the China B2B figure, and real care use is a pilot engagement with Fourier, not something the robot does on delivery.

The dancing, keytar-playing Adam you saw online is scripted and teleoperated: you are buying a superb, human-like moving platform and an SDK, not a robot that will autonomously do a job, and getting real work out of it takes a serious in-house RL and ROS2 team.

The H1 is a legacy machine (superseded by the H2, though Unitree has posted no formal end-of-life notice): it is a locomotion research chassis with no hands and no autonomy, and the newer H2 does most of it for a third of the money on the platform Unitree is actually developing, so buy the H2 unless you have a specific reason to want the aging H1 lineage.

You are buying a fast, capable research and pilot platform, not a turnkey worker: budget around $120,000, plan for an engineering team and battery-swap logistics, and know that runtime, safety certification, and ingress rating are still being validated.

You cannot buy this new any more: Unitree dropped it from the lineup and the US reseller is sold out. The prices you will still find quoted are historical, and the H2 does more of this job for a fraction of the money.

The flying-kick videos are real and genuinely impressive, but that is the point: the A3 is an athletic showcase and interaction unit, not a proven manipulation platform. The direct ~$45,000 price is closer to a display sticker than a fleet reality, and in Europe the honest access path is renting one at roughly GBP 1,999 a day.

Despite the 'humanoid' label, the G2 rolls on a four-wheel base; it does not walk on legs, so if you specifically wanted a bipedal robot this is not it (AgiBot's bipedal line is the A2 series). What you get is a very accurate, force-controlled dual-arm platform on wheels, sold to businesses by quote, with the dexterous OmniHand and task integration as separate work and cost.

This is a blueprint and a dev kit, not a product you unbox and use. There is no official price, no dexterous hands, a 0.5 kg per-arm payload and 2-hour battery, and it does nothing useful until your team trains and deploys the control policies. If you want a robot that works on arrival, this is not it; if you want to build and own a humanoid stack, it is the most open one there is.

It is a superb teleoperated dual-arm platform, but you are buying a research/deployment tool that learns tasks from your own VR demonstrations, not a robot that walks in and does chores by itself, and there is no public price or consumer purchase path.

The IP66 rating and -20C to 55C envelope are the real, sourced story here, and DEEP Robotics markets it as the first full-body sealed humanoid. Everything else (the price, the DOF, the hands, the SDK) is unpublished by the maker or comes from resellers, so this is an early-stage, quote-only industrial platform you buy on a purchase order and trust, not on a manufacturer spec sheet.

This is an announced, pre-market social-presence robot, not something you can buy or ship today. It walks and emotes convincingly but is not built to manipulate objects or do physical work, has no published SDK, and has no confirmed path outside China, and the ~$173K figure is a projected late-2026 price, not a live one.

You are buying a development platform, not a worker: it arrives programmable with real tactile hands and open tooling, but it does nothing useful until your team writes the software, there is no published price, and no consumer or retail channel exists.

This is a role-tailored appliance, not a general-purpose robot you can buy and hack. There is no public price, no SDK, and no off-the-shelf order button. KEENON leases it per job (aimed below half of local labor cost), some tasks only work when it is paired with KEENON's other robots, and as of its July 2025 WAIC debut it is still in pilot deployments (Shangri-La, Walmart) rather than open sale. CEO Li Tong is blunt: "At this stage, we don't expect to create an all-purpose robot."

This is a quote-only industrial pilot platform, not a product you can price or buy: there is no public price, no SDK, and no consumer path, and the folding-clothes and lamp-assembly demos are engineered showcases, not turnkey autonomy you get on delivery. If you want a robot with a checkout button, this is not it (and don't confuse it with the cheaper consumer CASBOT 02).

This is an open-source reference design, not a product you order and deploy: you can download every drawing and line of control code for free, but there is no price, no cart, and no support desk, and turning a Qinglong into anything useful is a full robotics engineering project, not an unboxing.

You are buying an early-stage pilot platform under a contract, not a finished worker: there is no list price, no public SDK, and what actually ships is whatever the sales agreement scopes, which is narrower than the launch video.

This is one of the few humanoids with a real track record of paid factory deployment rather than a stage demo, but you cannot just buy one: there is no list price, no cart, and no out-of-box autonomy. It is a quote-only enterprise deployment that only earns its keep after UBTech integrates and trains it on your line.

It is really working in factories, but there is no price and no purchase path for outsiders, so unless you are a large manufacturer becoming a partner, you cannot get one.

You are paying about $84,000 for a $29,900 H2 body with an NVIDIA Thor module and a reference-design badge, and the SharpaWave tactile hands that justify the whole thing add another $156,400. Read the DOF count carefully: the body is 31 DOF, not 75.

You cannot buy this. The "$89K" that gets quoted is Xiaomi's own build-cost estimate for a single prototype, not a price, and there has never been an SDK, a hands option, or a way to order one.

The gait demos went viral, but you cannot buy one and will not be able to for years: mass production is a 2027 story, and any price you see is invented.