Watch a humanoid robot perform a standing backflip and you are watching the most impressive solved problem in modern robotics. Watch one try to pick up a wet sponge, turn a key in a lock, or fold a crumpled shirt, and you are watching the actual frontier. This inversion, that spectacular dynamic locomotion is now routine while ordinary household manipulation remains unsolved, is the single most important fact about the humanoid industry heading into 2027, the year the first serious consumer volumes are supposed to ship.
A recent industry survey summed up by Benzinga put it plainly: basic humanoid locomotion is effectively a solved problem, while complex dexterity remains the hard part. The question the field is now debating is whether manipulation is approaching its own ChatGPT moment. The evidence says: not yet, and the reason is worth understanding in detail, because the entire 2027 production calendar depends on it.
The Strange Economics of a Backflip
A backflip looks harder than laundry. For a robot, it is much easier. A flip is a tightly defined movement: the robot knows exactly what it is trying to do, the environment can be controlled, and the sequence can be rehearsed thousands of times in simulation before transfer. Interesting Engineering’s recent deep dive on robotic hands captures the core asymmetry: household work is full of uncertainty. Objects bend, slip, deform, break, and appear in slightly different places every time.
The factory has always been the easy case for manipulation. An arm assembling the same component thousands of times gets a fixed trajectory, predictable lighting, known object dimensions, and specialized tooling. The home is the opposite case. A shirt is crumpled differently every time. A glass may be partly filled. A drawer is slightly open. A plastic bag has no fixed geometry at all. Researchers at Ohio State University have described physical contact with soft, fragile, or irregular objects as one of the fundamental unsolved problems of household robotics, because contact is genuinely difficult to model and control.
This is why the demos have evolved the way they have. Unitree’s H1 backflips, its G1 runs through athletic routines, Chinese firms stage kung fu and dance performances. Meanwhile the same companies are only now beginning to publish results on picking up litter and arranging flowers. The performance layer of the industry matured years ahead of the utility layer.
What the Best Hand Research Actually Achieves
The current state of the art in dexterous manipulation is best measured not by demos but by controlled studies, and the most cited one this year comes from Zhejiang University. Published in Science Robotics and covered in detail by Interesting Engineering, the work combined visual and tactile information with reinforcement learning and online imitation learning. The result: an 85% success rate across five complex tasks involving 25 objects, including precision work like sharpening a pencil and twisting open a bottle cap.
Two things about that result deserve attention. First, 85% is genuinely strong by research standards; five years ago, multi-object dexterous transfer at that success rate was not on the table. Second, and this is the part that matters for deployment planning: 85% is nowhere near good enough for a consumer product. A home robot that fails one task in seven is returned within a month. A warehouse robot that drops one part in seven is a liability. Industrial reliability conventionally means multiple nines, which implies the remaining error surface of manipulation must shrink by an order of magnitude or more before unsupervised home economics close.
The Zhejiang result also validates a specific technical thesis: tactile sensing is the lever. Vision alone cannot solve manipulation, because a camera can tell a robot it is holding an egg, but it cannot tell the robot how close it is to crushing it. Force and slip must be felt, not inferred.
The Data Gap Nobody Can Buy Their Way Out Of
Underneath the physics problem sits a data problem, and it is more structural than the physics one. Modern robots are trained rather than programmed, but useful manipulation data is qualitatively harder to collect than images. A useful demonstration needs synchronized joint positions, camera streams, force, torque, and tactile readings. As IEEE’s robotics society recently highlighted, tactile datasets remain tiny compared with visual datasets, and fine manipulation stays hard partly because of it.
The contrast with language models is stark. The internet handed LLM researchers trillions of words for free. There is no internet-scale dataset of humans manipulating thousands of objects while recording everything their hands feel. Every player is forced to manufacture its own substitute, and the strategies reveal what each company believes the missing ingredient is:
- 1X is betting on video. CEO Bernt Børnich, whose 2027 shipment target of 50,000 NEO robots was reported this month, has said the company expects web video to provide roughly 99% of its training data, supplemented by simulation, synthetic data, egocentric video, wearable sensors, teleoperation, and autonomous robot experience. Video is abundant and cheap. It is also silent on contact forces.
- 1X is also betting on humans in the loop. NEO’s Expert Mode schedules a trained remote operator to teleoperate the robot through tasks it cannot yet do itself, completing the chore for the customer while generating demonstration data as a byproduct. It is an honest architecture: the product works from day one, autonomy grows asymptotically. It also means early “autonomy” metrics deserve careful reading.
- Figure is betting on recorded human experience at scale. Its Index dataset and Helix 2.5 results, covered here yesterday, took zero-shot whole-body household success from 9% to 56% via pretraining alone, the cleanest published evidence that broad human-experience transfer works.
Each strategy is a different answer to the same hole in the data supply chain, and none of them yet produces the tactile channel the Zhejiang results say matters most. That is the gap to watch in 2027.
Hands Become a Product Category
While the AI labs fight over data, the component market is moving fast, and the center of gravity is in China. Data from the Qianzhan Industry Research Institute, cited in a Gasgoo interview with Zhaowei executive Chen Yidong, puts China’s dexterous hand market at 960 million yuan in 2026, a 170% year-on-year increase. That is a small market growing at an absurd rate, the classic signature of a component category being born.
The product launches are coming thick and fast. Prima 1, a new dexterous hand aimed at industrial manipulation, entered tracking this month with a 7/10 capability score from Humanoid.Guide’s assessment framework. Unitree’s G1 can be fitted with a force-controlled three-fingered dexterous hand with optional tactile sensing. Academic labs keep pushing the design envelope in stranger directions: a recent paper demonstrated a commercial anthropomorphic hand learning self-supported locomotion and manipulation by using its fingers as legs, blurring the line between gripper and limb.
Chen Yidong’s design philosophy, that the first goal of a dexterous hand is to look good, sounds like marketing until you unpack it. Humanoid hands are increasingly consumer-facing products whose appearance drives trust and adoption, not just industrial end-effectors. The interview also reflects a supply chain truth: China’s hand makers are treating aesthetics, cost, and manufacturability as first-class engineering constraints, the same discipline that made it the world’s EV battery hub.
The Hand Meets the Assembly Line
The dexterity gap lands exactly when the industry is committing to volume. XPeng announced that its Iron humanoid walked off a production line that the company says operates at 80% autonomy, with mass production targeted for the end of 2026 and deliveries in 2027. He Xiaopeng now personally leads the robotics division, per Reuters reporting from June. XPeng’s press release calls Iron the world’s first advanced general-purpose humanoid, a claim Humanoid.Guide correctly flags as undefined and unverified, and notably absent from all coverage are output volumes, cycle times, and yield figures.
Here is the uncomfortable synthesis: automating the manufacture of humanoid bodies does little to solve the parts of those bodies that actually determine whether they are useful. An 80% automated line can stamp out torsos, legs, and locomotion-capable platforms all day. The yield risk concentrates in the hands: dozens of tiny actuators, tendon or linkage routings, and tactile sensor surfaces per unit, all subject to wear, all currently immature. In unit-economic terms, the hand is likely to behave like a consumable inside a capital good, the component that determines mean time between failures and therefore the true cost per task.
That reframes the 2027 volume race. 1X’s 50,000-unit target and XPeng’s mass production are, on one level, body-count competitions. But revenue per body will be set by tasks completed autonomously, and tasks completed autonomously will be throttled by the hand. A robot that walks flawlessly for 4,000 hours but averages a finger failure every 200 is not a household appliance; it is a lease with a service contract.
What to Watch
Three signals will tell you whether the hand problem is closing:
- Disclosed success rates at deployment, not demo time. Figure publishing 56% zero-shot with fixed checkpoints and blind grading is the standard now. Any company shipping home robots in 2027 that will not state interventions per task is telling you the number.
- Tactile data volumes. The first lab that shows manipulation gains from a tactile dataset measured in millions of contact-rich episodes, rather than thousands, has found the missing channel.
- Hand-specific reliability metrics. Mean time between finger failures, cycles per fingertip sensor, hand replacement cost as a percentage of robot BOM. When manufacturers start publishing these, hands have become a mature component industry. China’s 960 million yuan market says that transition has begun.
The locomotion race is over; several companies won it simultaneously, which is precisely why it stopped mattering. The dexterity race is just starting, and it is the one that decides whether the humanoid is the next automobile or the next Segway. The backflip was never the product. The folded towel is.